UNIVERSIDAD NACIONAL AUTÓNOMA DE MÉXICO
PROGRAMA DE MAESTRÍA Y DOCTORADO EN INGENIERÍA
INGENIERÍA ELÉCTRICA – TELECOMUNICACIONES
IMPLEMENTATION OF MACHINE LEARNING TECHNIQUES TO OPTIMIZE
THE GROUND SEGMENT FOR HIGH THROUGHPUT SATELLITE SYSTEMS IN
Q/V BAND
T E S I S
QUE PARA OPTAR POR EL GRADO DE:
DOCTOR EN INGENIERÍA
PRESENTA:
IVÁN ANDRÉS CORNEJO GAIBOR
TUTORES PRINCIPALES
DR. SALVADOR LANDEROS AYALA, FI-UNAM
DR. JOSÉ MARÍA MATÍAS MARURI, FI-UNAM
COMITÉ TUTOR
DR. VÍCTOR RANGEL LICEA, FI-UNAM
DR. RAMÓN MARTÍNEZ RODRÍGUEZ-OSORIO, GR-UPM
CIUDAD DE MÉXICO, NOVIEMBRE 2020
UNAM – Dirección General de Bibliotecas
Tesis Digitales
Restricciones de uso
DERECHOS RESERVADOS ©
PROHIBIDA SU REPRODUCCIÓN TOTAL O PARCIAL
Todo el material contenido en esta tesis esta protegido por la Ley Federal
del Derecho de Autor (LFDA) de los Estados Unidos Mexicanos (México).
El uso de imágenes, fragmentos de videos, y demás material que sea
objeto de protección de los derechos de autor, será exclusivamente para
fines educativos e informativos y deberá citar la fuente donde la obtuvo
mencionando el autor o autores. Cualquier uso distinto como el lucro,
reproducción, edición o modificación, será perseguido y sancionado por el
respectivo titular de los Derechos de Autor.
JURADO ASIGNADO:
Presidente: Dr. Víctor Rangel Licea
Secretario: Dr. Miguel Moctezuma Flores
1er. Vocal: Dr. José María Matías Maruri
2do. Vocal: Dr. Salvador Landeros Ayala
3er. Vocal: Dr. Ramón Martínez Rodriguez-Osorio
Lugar donde se realizó la tesis: FACULTAD DE INGENIERÍA, UNAM.
TUTOR DE TESIS:
DR. SALVADOR LANDEROS AYALA
--------------------------------------------------
FIRMA
To my parents, Iván and Lourdes, for their endless love and encouragement.
To my loving wife, María Fernanda, for her support and patience.
Thank you for making it possible.
Acknowledgements
My deepest thanks to my tutor and supervisor, Dr. Salvador Landeros-Ayala, for
sharing his knowledge and his support during my Ph.D study, whose guide and advice
were essentials in the development of this thesis. Also, he was always aware of me
and my family in Mexico. I am very grateful to him for allowing me the opportunity
to pursue my Ph.D at Universidad Nacional Autonoma de Mexico, UNAM. It has
been a privilege and an honor to work with him and learn from him.
I wish to thank my co-tutor and co-supervisor, Dr. Jose Maria Matias Maruri,
for his patience and his linking-up in my research work. During my Ph.D thesis, he
always supported me all-the-time. We had several meetings to discuss different topics
of my research, administrative issues at UNAM, and my residence in Mexico.
I wish also to express my gratitude to Dr. Ramon Martinez and his work-team at
Universidad Politecnica de Madrid, UPM, in Madrid, Spain, where I performed my
studies as a visiting student. We exchanged several criteria for my research and gave
me excellent feedback in order to improve some aspects of my thesis. I am eternally
grateful for all his comments and advice.
I would also like to thank Dr. Victor Rangel-Licea, he is a member of the
tutoring committee of this thesis and Head of the Telecommunications Engineering
Department at DIE, UNAM. Through his support and management, it was possible
to attend different events to enrich my knowledge in the telecommunications field.
I want to thank Dr. Miguel Moctezuma Flores for agreeing to be part of my Ph.D
committee member and review my research.
A special thanks to Mtra. Susana Kolb Cadwell. She is the coordinator at DGECI,
UNAM. DGECI has a great program in academic writing that I was part of this
program as a student, learned a lot from her as well as her work-group.
I am grateful to Consejo Nacional de Ciencia y Tecnologia, CONACYT, for
granting me a scholarship to study my Ph.D at Universidad Nacional Autonoma
de Mexico, UNAM, CVU/Scholarship number: 559998/298954, and the CONACYT
International Mobility Grant, 291250, to attend to Universidad Politecnica de Madrid,
UPM, in Madrid, Spain.
At last but not least, I must express my endless gratitude to my parents. They
have always supported my studies and have stayed with me in my life decisions. To
my grandmother, for her eternal blessings. I am grateful to my wife Fernanda for her
love and encouragement every time, in all good and bad moments. To my friends and
colleagues in Mexico, Spain, and Ecuador, thank you all.
ii
Abstract
Rain attenuation events are one of the most important drawbacks in satellite
communications, impairing directly on satellite link availability. For this reason,
it is necessary to foresee rain events in order to avoid an outage of the satellite
link. However, the lack of rainfall database hinders the development of a reliable
prediction method. In this thesis, we implement an alternative method to generate
rain attenuation time-series in the Q/V band for a specific geographic location based
on the recommendation of the International Telecommunication Union. With the
computed rain-attenuation data, we propose and develop a method to predict rain
attenuation events based on deep learning techniques without appealing to complex
mathematical models. To be specific, we implement the Long-short term memory
network, which is a Machine Learning technique based on supervised learning. Each
model is trained and validated by computational experiments, employing statistical
metrics to find the most accurate and reliable models. The outcomes of the prediction
model are employed and discussed to compare with other related methods and models.
In addition to the evaluation of the proposed method, we find that our method could
be notably able to improve mechanisms for switching either data traffic or satellite
links in order to increase the satellite-link availability.
For this purpose, two methods were proposed to optimize the ground network by
the implementation of the deep learning method as well as another method based
on the Markov chain. The smart-predictive method is based on the developed deep
learning model, where the predictions of the rain attenuation and the carrier-to-noise
and interference time series are able to detect, in advance, when one or more gateway
feeder uplinks are going to be affected by the rain, so that the smart method
can foresee the switching between the affected nominal gateway and an available
redundant gateway. This mechanism is applied to all gateways of the ground network
so that it is possible to find the number of nominal and redundant gateways needed
for outperforming the minimum required availability of 99.9%.
The second method, in essence, employs the same principle but using the Markov
chain to find a stationary distribution, which allows us to know the time spent of
each state according to its nominal gateway or redundant gateways in the process,
in addition to the switching probability between them. The obtained results are
discussed and compared with another well-known method to define the ground
network scheme, finding a ground network more efficient, lower complexity, and with
a lower number of gateways, that take part in the data managing of the satellite
system.
iii
Contents
Acknowledgements ii
Abstract iii
Mathematical Notation xi
Nomenclature and Units xiii
Abbreviations and Acronyms xix
1 Introduction 1
1.1 Traditional Broadband Satellite Systems . . . . . . . . . . . . . . . . 1
1.2 Multiple Spot Beam Satellite Systems . . . . . . . . . . . . . . . . . . 3
1.3 Evolution of High Throughput Satellite Systems . . . . . . . . . . . . 4
1.4 State-of-art in current HTS systems . . . . . . . . . . . . . . . . . . . 6
1.4.1 Feeder Link Design in Q/V band . . . . . . . . . . . . . . . . 6
1.4.2 Aggressive Frequency Reuse . . . . . . . . . . . . . . . . . . . 6
1.4.3 Ground Segment Architecture . . . . . . . . . . . . . . . . . . 7
1.5 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.6 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.7 Contributions of the Research . . . . . . . . . . . . . . . . . . . . . . 9
1.8 Thesis Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2 System Models for the Ground Segment 11
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Feeder Uplink Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Parameters of the Feeder Uplink . . . . . . . . . . . . . . . . . 13
2.3 Spatial Correlation Model . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 The Dynamic Rain Attenuation Model . . . . . . . . . . . . . . . . . 15
2.4.1 Procedure to Synthesize the Rain Attenuation Time-series . . 17
2.5 N + P Diversity Model on Ground Segment . . . . . . . . . . . . . . 18
3 Quantification of Multiple Spot Beams for HTS Systems in Ka and
Q/V bands 21
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.2 Connectivity in Latin America: A Review . . . . . . . . . . . . . . . 21
iv
CONTENTS v
3.3 Available Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.1 Ka-band Radio Regulations . . . . . . . . . . . . . . . . . . . 23
3.3.2 Q/V band Radio Regulations . . . . . . . . . . . . . . . . . . 24
3.4 Satellite Link Design in Ka and Q/V bands . . . . . . . . . . . . . . 26
3.5 Radio Interfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.6 Methods to Calculate the Performance of Satellite Antennas . . . . . 28
3.6.1 First Method: Antenna Size Evaluation . . . . . . . . . . . . . 29
3.6.2 Second Method: Calculations for Antenna Gain and EIRP . . 29
3.7 Analysis and Quantification of Multiple Spot Beams for HTS Systems
in Ka and Q/V bands: Numerical Results and Discussion . . . . . . . 30
3.7.1 Orbital Position Analysis . . . . . . . . . . . . . . . . . . . . . 31
3.7.2 Quantification of Multiple Spot Beams . . . . . . . . . . . . . 32
3.8 Contributions of the Research . . . . . . . . . . . . . . . . . . . . . . 37
4 Interference Evaluations in Frequency Reuse by Using Offset-
Parabolic-Reflector Antennas 38
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2 Geometric Design of an Offset-Parabolic-Reflector Antenna . . . . . . 38
4.3 Multiple Spot Beams Antenna for HTS Systems . . . . . . . . . . . . 41
4.3.1 Multiple Spot Beams: Design and Analysis . . . . . . . . . . . 42
4.3.2 Spot Beam Pattern: Model Analysis . . . . . . . . . . . . . . 43
4.4 Carrier-to-Interference: Evaluation Model . . . . . . . . . . . . . . . 44
4.5 Offset-Parabolic-Reflector Antennas: Numerical Results . . . . . . . . 45
4.5.1 Offset-Parabolic-Reflector Antenna: Geometric Parameters . . 45
4.5.2 Sizing of Spot Beams over the Coverage Area . . . . . . . . . 48
4.5.3 CIR Evaluation Results . . . . . . . . . . . . . . . . . . . . . 49
4.5.4 Link Margin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.6 Contributions of the Research . . . . . . . . . . . . . . . . . . . . . . 52
5 Method of Rain Attenuation Prediction Based On Long-Short Term
Memory Network 55
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.2 Implementation of the Proposed Deep Learning Method Based on
LSTM Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.2.1 Deep Learning Network: Model Architecture . . . . . . . . . . 56
5.2.2 Deep Learning Network: Model Description . . . . . . . . . . 63
5.3 Experimentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.3.1 Experiment 1: Training and validation subsets partitioning, 70/30 73
5.3.2 Experiment 2: Training and validation subsets partitioning, 90/10 74
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.5 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.6 Contributions of the Research . . . . . . . . . . . . . . . . . . . . . . 90
CONTENTS vi
6 Ground Segment Optimization by Using Smart Strategies of
Switching Between Gateway Stations 91
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
6.2 Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.2.1 Sizing of Nominal Gateways, NGWs . . . . . . . . . . . . . . . 92
6.2.2 Smart Method for Forecasting Rain Attenuation and CNIR at
each GW . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
6.2.3 Method to Decide the Best NGWs . . . . . . . . . . . . . . . . 93
6.2.4 A Strategy to Allocate the Best PGWs . . . . . . . . . . . . . 94
6.3 The 1 + P̄ Scheme Analyzed by a Markov Chain . . . . . . . . . . . . 95
6.3.1 Method to Calculate the Probability of Rain, P0 . . . . . . . . 96
6.3.2 The Switching Strategy, 1+P̄ , from the perspective of a Markov
Chain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.5 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.6 Contributions of the Research . . . . . . . . . . . . . . . . . . . . . . 115
7 Conclusions, Future Work, and Contributions 116
7.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.3 Contributions and Production . . . . . . . . . . . . . . . . . . . . . . 118
Appendix A List of Geographic Coordinates for Gateway (GW)
Locations 120
A.1 List of Geographic Coordinates for Gateway (GW) Locations, Chapter 3120
A.2 List of Geographic Coordinates for Gateway (GW) Locations, Chapter 5120
A.3 Distance Matrix Between Pairs of Locations, Chapter 6 . . . . . . . . 120
Bibliography 125
List of Figures
1.1 Architecture of a broadband satellite system. . . . . . . . . . . . . . . . 2
1.2 Architecture of a multiple spot beams satellite system with 4-color frequency
reuse scheme. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Capacity Evolution of HTS systems. . . . . . . . . . . . . . . . . . . . . 5
2.1 The correlation coefficient, ⇢a, between two sites separated by a distance of
D km. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.2 Block diagram of the rain attenuation time series synthesizer, ITU-R
P.1853. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Gateway diversity scheme with a redundant GW. . . . . . . . . . . . 20
3.1 Percentage of households with Internet access per country [40]. . . . . . . 22
3.2 Percentage of households with Internet access per geographical zone [40]. . 23
3.3 Es/N0 for each feeder uplink using the Monte Carlo Method. . . . . . . . 32
3.4 Normal distribution of the orbital-position observations. . . . . . . . . . . 33
3.5 Comparison of the obtained number of beams using DVB-S2X and DVB-S2
standards. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.6 The number of carriers in the feeder downlink. . . . . . . . . . . . . . . . 34
3.7 Total capacity comparison between different data traffic ratios. . . . . . . 35
3.8 The number of beams and carriers using both 3F1P and 3F2P schemes. 36
3.9 The number of beams and carriers using both 4F1P and 4F2P schemes. 37
4.1 The geometry for the offset-parabolic-reflector antenna. . . . . . . . . . . 40
4.2 The hexagonal-grid layouts of both 3-cell and 4-cell for frequency reuse,
illustrating the beam parameters. . . . . . . . . . . . . . . . . . . . . . 42
4.3 The interference geometry of the downlink. . . . . . . . . . . . . . . . . 46
4.4 Normalized pattern of the offset-parabolic-reflector antenna, 40 GHz. . . . 48
4.5 Spot beams of feeder downlinks over the Latin America region. . . . . . . 50
4.6 The 3-color frequency-reuse schemes over the Mexico region with 4 different
beamwidths. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.7 The 4-color frequency-reuse schemes over the Mexico region with 4 different
beamwidths. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.8 Forward link margin vs. co-channel CIR. . . . . . . . . . . . . . . . . . . 54
5.1 A single LSTM cell diagram [78]. . . . . . . . . . . . . . . . . . . . . . . 57
5.2 The architecture of an LSTM layer [77]. . . . . . . . . . . . . . . . . . . 59
vii
LIST OF FIGURES viii
5.3 The architecture of an LSTM multi-layer [80]. . . . . . . . . . . . . . . . 60
5.4 The architecture of the proposed LSTM layer. . . . . . . . . . . . . . . . 61
5.5 The architecture of the proposed deep learning network. . . . . . . . . . . 62
5.6 Sigmoid activation function. . . . . . . . . . . . . . . . . . . . . . . . . 64
5.7 Hyperbolic tangent activation function. . . . . . . . . . . . . . . . . . . 65
5.8 Rectifier-Linear-Unit activation function. . . . . . . . . . . . . . . . . . . 66
5.9 Scaled-Exponential-Linear-Unit activation function. . . . . . . . . . . . . 66
5.10 Block diagram of the proposed method based on the deep learning
network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.11 RMSE scores for training subsets. . . . . . . . . . . . . . . . . . . . . . 76
5.12 RMSE scores for validation subsets. . . . . . . . . . . . . . . . . . . . . 76
5.13 MAE scores for training subsets. . . . . . . . . . . . . . . . . . . . . . . 77
5.14 MAE scores for validation subsets. . . . . . . . . . . . . . . . . . . . . . 77
5.15 The training rain-attenuation subsets obtained from deep learning models. 78
5.15 The training rain-attenuation subsets obtained from deep learning models
(cont.). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.16 The validation rain-attenuation subsets obtained from deep learning models. 80
5.16 The validation rain-attenuation subsets obtained from deep learning models
(cont.). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.17 R2 for training subsets. . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.18 R2 for validation subsets. . . . . . . . . . . . . . . . . . . . . . . . . . . 82
5.19 Model performance results: measured vs. predicted rain-attenuation values
for validation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.19 Model performance results: measured vs. predicted rain-attenuation values
for validation (cont.). . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
5.20 The higher R-Squared, R2 = 0.9224 at Tuxtla, Gtz. . . . . . . . . . . . . 85
5.21 The lower R-Squared, R2 = 0.7265 at Cd. Juarez. . . . . . . . . . . . . . 85
5.22 Training and validation loss functions for each rain-attenuation time series. 86
5.22 Training and validation loss functions for each rain-attenuation time series
(cont.). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.23 Comparison between predictive models. . . . . . . . . . . . . . . . . . 89
6.1 Flowchart of the multiple 1 + P̄ switching strategy. . . . . . . . . . . . . 96
6.2 The 1 + P̄ strategy represented by a Markov Chain graph. . . . . . . . . . 99
6.3 Unavailability percentage per year, for all N̄ configuration schemes. . . . . 102
6.4 System availability vs. the feeder uplink capacity, on average, by ranging
the CNIR threshold, ⇠th. . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6.5 Performance of the N̄ + P̄ schemes, as a function of system unavailability
and the number of P̄ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.6 Performance of the N̄ + P̄ schemes, as a function of system outage
probability, Pout, and the number of P̄ . . . . . . . . . . . . . . . . . . . 113
6.7 N̄+P̄ configuration schemes obtained from the three aforementioned methods.114
List of Tables
2.1 Basic Propagation Parameters . . . . . . . . . . . . . . . . . . . . . . 13
3.1 Frequency Allocations for Satellite Downlink in Ka-band, ITU Region
R2 [4]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Frequency Allocations for Satellite Uplink in Ka-band, ITU Region R2
[4]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Frequency Allocations for Satellite Downlink in Q band, ITU Region
R2 [4]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Frequency Allocations for Satellite Uplink in V band, ITU Region R2
[4]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.5 Frequency Reuse and Polarization Schemes [18]. . . . . . . . . . . . . 27
3.6 Maximum Spectral Efficiencies . . . . . . . . . . . . . . . . . . . . . . 28
3.7 Feeder Uplink Parameters . . . . . . . . . . . . . . . . . . . . . . . . 32
3.8 Throughput Comparisons . . . . . . . . . . . . . . . . . . . . . . . . 35
4.1 Geometric Parameters of the Offset-Parabolic-Reflector Antenna . . . 47
4.2 Spherical Spread Losses and Edges illuminations with a Symmetric
Gaussian Radiation Pattern . . . . . . . . . . . . . . . . . . . . . . . 47
4.3 Spherical Spread Losses and Edges illuminations, f = 34.71 . . . . 47
4.4 Cross-polar isolation XPiso for both feeder and user downlinks. . . . . 48
4.5 The Number of Spot Beams (Nbmin
/Nb) for the Latin American Regions. 49
4.6 The CIR Evaluation for Feeder Downlink, 40 GHz. . . . . . . . . . . 49
4.7 The CIR Evaluation for the 3-color Frequency-reuse Scheme. . . . . . 51
4.8 The CIR Evaluation for the 4-color Frequency-reuse Scheme. . . . . . 53
4.9 The CNIR Values for the Forward Link. . . . . . . . . . . . . . . . . 53
5.1 Hyperparameters for Experiment 1 . . . . . . . . . . . . . . . . . . . 74
5.2 Hyperparameters for Experiment 2 . . . . . . . . . . . . . . . . . . . 75
6.1 Month Numbers and Number of Days . . . . . . . . . . . . . . . . . . 97
6.2 The Inbound and Total Capacity of the Ground Network Segment. . 101
6.3 Gateway Availabilities. . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.4 The Number of Outages of Each NGW for the 4 + P̄ Scheme. . . . . 104
6.5 The System Availability for the 4 + P̄ Scheme. . . . . . . . . . . . . . 104
6.6 The Number of Outages of Each NGW for the 8 + P̄ Scheme. . . . . 104
6.7 The System Availability for the 8 + P̄ Scheme. . . . . . . . . . . . . . 104
ix
LIST OF TABLES x
6.8 The Number of Outages of Each NGW for the 12 + P̄ Scheme. . . . . 105
6.9 The System Availability for the 12 + P̄ Scheme. . . . . . . . . . . . . 105
6.10 The Number of Outages of Each NGW for the 16 + P̄ Scheme. . . . . 106
6.11 The System Availability for the 16 + P̄ Scheme. . . . . . . . . . . . . 107
6.12 The Operation Probability of Gateways. . . . . . . . . . . . . . . . . 108
6.13 The Stationary Distribution of Each NGW for the 4 + P̄ Scheme. . . 109
6.14 The System Operation Probabilities for the 4 + P̄ Scheme. . . . . . . 109
6.15 The Stationary Distribution of Each NGW for the 8 + P̄ Scheme. . . 110
6.16 The System Operation Probabilities for the 8 + P̄ Scheme. . . . . . . 110
6.17 The Stationary Distribution of Each NGW for the 12 + P̄ Scheme. . . 111
6.18 The System Operation Probabilities for the 12 + P̄ Scheme. . . . . . 111
6.19 The Stationary Distribution of Each NGW for the 16 + P̄ Scheme. . . 112
6.20 The System Operation Probabilities for the 16 + P̄ Scheme. . . . . . 112
A.1 List of Geographic Coordinates for GW Locations, Chapter 3 . . . . . 121
A.2 List of Geographic Coordinates for GW Locations, Chapter 5 . . . . . 122
Mathematical Notation
Note: the vectors are denoted by bold lowercase letters whereas the matrices are
expressed by bold uppercase letters.
[N ] the set {1, 2, . . . , N}
x a dataset
^ logical conjunction, if statement
bxc the floor of a scalar x
_ logical disjunction
R the set of real numbers
U an interval [0, 1]
E{·} the mathematical expectation of a random variable
X a Matrix
x a vector
x y the Hadamard product (element-wise product of either vectors or matrices)
x(i) the i-th example (input) of a dataset
XT the transpose of matrix X
X:,i the column i of matrix X, (a column vector)
Xi,: the row i of matrix X, (a row vector)
Var{·} the variance of a set of data, 2
| x | absolute value of a scalar x
rXy the matrix derivatives of y with respect to X
rxy the gradient of y with respect to x
±U an interval [ 1, 1]
xi
MATHEMATICAL NOTATION xii
Pr{·} the probability of a random event
! prepositional logic, if ... then ...
{·} the standard deviation of a set of data
{0, 1} the set containing 0 and 1
f(·) the general form of the mathematical function
Jn(·) the Bessel function of the n kind
m{·} the mean of a set of data
Q complementary cumulative normal distribution
Q 1 inverse complementary cumulative normal distribution
s(t) continuous signal, where t is the independent variable
s[k] discrete signal, where k is the independent variable
x a scalar (real or integer)
xi the element i of a vector x, with indexing starting at 1
Xi,j the element i, j of matrix X
Nomenclature and Units
Ac the coverage area, [km2]
Ah the area of the hexagon cell, [km2]
Ai the rain attenuation in the i-th site
Aoffset the parameter that adjusts the time-series to match the probability rain, [dB]
BWT total-bandwidth available, [Hz]
Cpi the capacity of the satellite link, b/s
CpT total-capacity available, [b/s]
D distance, [km]
DL the download traffic, [b/s]
Dant the diameter of satellite antenna, [m]
E the signal’s average power
EIRPsat the effective isotropic radiated power from satellite antenna, [dBW]
Es/N0 the energy per symbol to noise power spectral density, [dB]
F the focal length of the parabolic-reflector-antenna, [m]
Freqs E the frequency on the space-Earth (s-E) path, [Hz]
G/T antenna gain-to-noise-temperature, [dB/K]
Gr(✓) the relative gain of the antenna, [dB]
GTX the transmitter-antenna gain, [dBi]
H the offset of the parabolic-reflector-antenna center, [m]
HPAsat the high power amplifier of the satellite in Ka-band, [W]
K the total number of samples, the K-dimensional Euclidean space
xiii
Nomenclature and Units xiv
LNA low noise amplifier, [W]
MTii the monthly mean total rainfall data, [mm]
N number of GWs on the Ground Segment
Nb the number of the beams
Np the potential for polarization reuse
Nii the number of days of each month
Nsb the number of the frequency sub-bands
PL the path loss from either s-E or E-s, [dB]
Pa conditional joint probability that the attenuations exceed a1 and a2, respectively
Pi the probability of occurrence in the i-th site
P rain
i the annual probability of rain (%) in the i-th site, a.k.a P0i
Pr joint probability that it is raining at both sites
P0ii the monthly probability of rain, [%]
Pout the total probability of a system outage
Psw the switching probability
R the range between HTS and GW, [km]
Ri solution of P rain
i for the i-th site
Rp the rainfall rate exceeded for the desired probability of exceedance, [mm/h]
R0.01 the annual rainfall rate data exceeded for 0.01% of an average year, [mm/hr]
Rsw the switching rate
Savail the total available spectrum, [Hz]
Seff the spectral efficiency, [b/s/Hz]
Ts the time interval between samples
Tii the monthly mean surfaces temperatures, [K]
Tsw the interval between switching instants
UL the upload traffic, [b/s]
Γk,i the k-th value of CNR matrix, Γ, according to the i-th feeder link of GW
Nomenclature and Units xv
↵ the learning rate
↵i phase component of the communication channel for the i-th GW feeder link
N̄ number of nominal gateways on the Ground Segment
P̄ number of redundant gateways on the Ground Segment
the parameter that describes the time dynamics, [s 1]
1 the linear decay factor
2 the quadratic decay factor
` the number of iteration
✏ the small constant in order to avoid division by zero
csi clear sky CNR for the i-th GW feeder link
Boltzmann’s constant, 1.380 649⇥ 10 23 [J/m]
the wavelength of the frequency band used in a specific satellite link, [m]
Ct the collection of all cell state at time step t
D the distance matrix between pair of locations
Ht the collection of all hidden state at time step t
P the state transition matrix
Wd the weights matrix of the dense layer
Wh the recurrent weights matrix, Whf ,Whi,Who
Wx the input weights matrix, Wxf ,Wxi,Wxo
ΘD the concatenation matrix of the dense layer (weights and biases)
ΘL the concatenation matrix of the LSTM layer (weights and biases)
bd the bias vector of the dense layer
b the bias vector of the LSTM layer, bf ,bi,bo
ct the stacked cell states at time step t
ht the stacked hidden states at time step t
D the mathematical function of a dense layer
F the deep learning model
Nomenclature and Units xvi
Ii( ) the function that describes the nature of the interference at the i-th beam, [dB]
LU the mathematical function of a LSTM layer
L the mathematical function of a single LSTM cell
Ni the noise variance at satellite interface for the i-th GW feeder link
S the mathematical function of a LSTM multi-layer
⌫ the system availability, where ⌫i is the feeder uplink availability of each NGW
! the angle that ranges from 0 (beam center) to ✓BW/2 (beam edge), [deg]
the angle between the point T and the center of the i-th beam, [deg]
B the bisector angle of parabolic-reflector-antenna, [deg]
C the f aims to the aperture center at the point C, [deg]
L the lower angle of parabolic-reflector-antenna, [deg]
P the angle from the lower edge of dish to feed pointing direction, [deg]
U the upper angle of parabolic-reflector-antenna, [deg]
f the angle of feed-antenna-pattern, [deg]
s the half of the angle subtended of parabolic-reflector-antenna, [deg]
⇢a the correlation factor for the conditional joint probability
⇢r the correlation factor for the joint probability
g the gate function activation based on the sigmoid function
✓0 the beam diameter at the triple beam crossover, [km]
✓c the closest spacing between beam centers reusing the same frequency, [km]
✓r the closest spacing between the reuse-beam edges, [km]
✓s the spacing between adjacent-beam centers, [km]
✓3dB the angle subtended by the half-power points of the main lobe, [deg]
✓BW Beamwidth in degrees, [deg]
⇠th the CNIR threshold
⇣ the value of the carrier-to-interference, [dB]
ai the attenuation threshold in the i-th site, [dB]
Nomenclature and Units xvii
ct the cell state at time step t
cu the cell candidate of the LSTM block
d0 GEO distance of 35786 [km] above Earth’s surface
fg the forget gate of the LSTM block
h the offset distance of the parabolic-reflector-antenna, [m]
hi[k] the kth value of the communication channel for the i-th GW feeder link
ht the hidden state at time step t
ig the input gate of the LSTM block
ii the month numbers
k index of each sample
m the number of dataset’s examples
n the integer number
n(t) the white gaussian noise process
og the output gate of the LSTM block
p the probability of the rain
pd the desired average annual probability of exceedance
rc the coverage radius of spot beam, [km]
ro roll-off parameter, [%]
tii the monthly mean surfaces temperatures, [ C]
v` the moving average
Ξk,i the k-th value of CNIR matrix, Ξ, according to the i-th feeder link of GW
t time in seconds, 1 [s]
b/s rate in bit per second, 1 [b/s]
dB decibels, 1 [dB]
Gb/s giga bit per second, 1⇥ 109 [b/s]
GHz giga Hertz, 1⇥ 109 [Hz]
Nomenclature and Units xviii
Hz frequency in hertz, 1 [Hz]
km kilometers, 1⇥ 103 [m]
km2 square kilometers, 1⇥ 106 [m2]
m distance in meters, 1 [m]
Mb/s mega bit per second, 1⇥ 106 [b/s]
MHz mega Hertz, 1⇥ 106 [Hz]
mm millimeters, 1⇥ 10 3 [m]
mW milli watts, 1⇥ 10 3 [W]
Tb/s tera bit per second, 1⇥ 1012 [b/s]
W power in watts, 1 [W]
Abbreviations and Acronyms
5G Fifth generation wireless technology for digital cellular networks
a.m.s.l. Height above mean sea level
ACI Adjacent Carrier Isolation factor
ACM Adaptive Coding and Modulation
ADAM Adaptive Moment Estimation Optimizer
ARIMA Autoregressive Integrated Moving Average
BRASILSAT Brazilian communications satellite
BSS Broadcast Satellite Service
CIR Carrier-to-interference ratio
CNIR Carrier-to-noise and interference ratio
CNR Carrier-to-noise ratio
DVB Digital Video Broadcasting
DVB-RCS2 DVB - Return Channel Satellite, Second Generation
DVB-S2X DVB - Second Generation, Extensions
E-s Earth-to-space satellite communication link
EHF Extremely high frequency
EI the edge illuminations which can be both upper and lower, [dB]
EIRP Effective Isotropic Radiated Power
EoC the beam End-of-Coverage
EUTELSAT European Telecommunications Satellite Organization
xix
Abbreviations and Acronyms xx
FMT Fade Mitigation Technique
FRF Frequency Reuse Factor
FS Fixed Service
FSL Free Space Loss
FSS Fixed Satellite Service
FT the feed edge tapers which can be both upper and lower, [dB]
GEO Geostationary Orbit
GPU Graphics Processor Units
GS Ground Segment
GW Gateway Station
HD High Definition Video
HDFSS High-Density application in the Fixed Satellite Services
HPA the High Power Amplifier
HTS High Throughput Satellite
i.i.d Independent and identically distributed random variables
IDU Indoor Unit
IoT Internet of Things
ITU International Telecommunication Union
LDPC Low-Density Parity-Check
LHCP Left Hand Circularly Polarized
LSTM Long Short-Term Memory
M-APSK M-ary Amplitud and Phase-Shift Keying modulation
M-PSK M-ary Phase-Shift Keying modulation
M-QAM M-ary Quadrature Amplitude Modulation
M2M Machine-to-machine communication
MAE Mean Absolute Error
Abbreviations and Acronyms xxi
MBA Multiple Beam Antenna
MF-TDMA Multiple-Frequency Time Division Multiple Access
MODCOD Modulation and Coding Scheme
modem Modulator and Demodulator device
MS Mobile Service
MSS Mobile Satellite Service
NCC Network Control Center
NGSO Non-geostationary-satellite orbit systems
NGW Nominal Gateway
ODU Outdoor Unit
PGW Redundant or Backup Gateway
QoS Quality of Service
R-Squared Coefficient of determination, R2
R2 ITU Region 2
ReLU Rectifier-Linear-Unit Activation Function
RF Radio Frequency
RHCP Right Hand Circularly Polarized
RMSE Root Mean Squared Error
RMSProp Root Mean Square Propagation
s-E space-to-Earth satellite communication link
SaaS Software as a Service
SatCom Satellite Communication
SELU Scaled-Exponential-Linear-Unit Activation Function
SFTSS Standard Frequency and Time Signal-Satellite
SGD Smart Gateway Diversity
SNN Self-Normalizing Neural Networks
Abbreviations and Acronyms xxii
SSL Spherical Spreading Loss, [dB]
TDMA Time Division Multiple Access
TPU Tensor Processor Units
UHTS Ultra High Throughput Satellite Systems
ULPC Uplink Power Control
UT User Terminal
VSAT Very Small Aperture Terminal
XPiso Cross-Polar Isolation factor
Chapter 1
Introduction
The thesis is focused on the performance analysis of the ground segment (GS)
architecture for a new generation of satellite communications (SatCom) systems.
Nowadays, the current requirements of throughput and broadband to users are
huge so that the traditional communication infrastructures need to expand their
resources and capabilities to cover the service demand. This main idea is also applied
to SatCom systems in order to increase their capacity and to become an efficient
solution to carry data traffic. Thus, new technological challenges emerge, providing
an important research frame for the network architecture on GS. This chapter presents
the state-of-the-art related to the problems arisen with the new satellite technology,
specifically in areas involved in the feeder link design, multiple beam satellite systems,
the frequency reuse, interference evaluations, availability, and reliability. Further, the
chapter also brings out the main objectives of the thesis to study the aforementioned
areas.
1.1 Traditional Broadband Satellite Systems
A traditional broadband satellite system provides communication services in remote
areas wherein the terrestrial infrastructures do not reach them, especially in rural
areas. The services offered by a broadband satellite system range from Internet access
to multimedia services [1, 2]. Figure 1.1 depicts the architecture of a traditional
broadband satellite system. Each network component accomplishes a specific role
in the broadband satellite architecture. However, the satellite GS network has not
undergone a major change in its architecture, but there are some parameters to take
into account which are explained in the following subsections in detail [3].
• Satellite: The broadband satellite is placed in a geostationary orbit (GEO)
in order to cover a wider area. Also, the GEO satellite is able to connect a
gateway station (GW) to the user terminals (UTs) using radio frequency (RF)
links (feeder and user links). This link connection is also known as bent-pipe
satellite architecture, however, it is important to note that the amplification
and frequency translations are only executed by satellite onboard processing.
1
CHAPTER 1. INTRODUCTION 2
Forward Link
Return Link
Satellite GEO
Gateway Station,
GW
User Terminals,
UT
Feeder Link User Link
Figure 1.1: Architecture of a broadband satellite system.
• Gateway Station: The gateway (GW) station is capable of transmitting and
receiving data. That is, the GW is responsible for controlling, managing and
operating data traffic to and from user terminals (UT), via satellite links. In
satellite networks, GWs are also known as gateway earth stations. Further, the
GS network architecture can be made up of multiple GWs when the gateway
diversity is considered necessary.
• User Terminal: The user terminal (UT), also referred to as user earth station,
is a two-way broadband terminal, i.e., the UT can both transmit and receive
data to and from the satellite employing a very small aperture terminal (VSAT).
Moreover, the UT is a device that consists of two main units, the indoor unit
(IDU) and the outdoor unit (ODU) [1]. The ODU is made up of the antenna, the
RF transmitter, and one or more RF receivers. Meanwhile, the IDU contains the
modulator and demodulator (modem), and the interface to the local network.
• Feeder Link: It is a RF link between the GW and satellite. The feeder link
consists of both uplink and downlink. Currently, feeder links work in Ka-band1
for broadband satellite systems.
• User Link: The RF link between the satellite and UTs is also known as the
user link, which is made up of both uplink and downlink. It is important to
mention that time division multiple access (TDMA) [1] and multi-frequency
time division multiple access (MF-TDMA) [1, 5] are employed to access from
satellite to UTs and vise versa. Not only do the feeder link uses Ka-band but
also the user link.
1The Ka-band is a frequency band of millimeter waves ranges from 17.70–21.20 GHz for downlink,
whereas from 24.75–30.00 GHz for uplink. Both frequency ranges are available for ITU Region 2
(R2), [4].
CHAPTER 1. INTRODUCTION 3
• Forward Link: The link from GW towards UTs is described as the forward
link. The forward link is also called the end-to-end link, which includes uplink
of the feeder link, satellite, and downlink of the user link. Therefore, a GW
transmits data to the UTs via the forward link. In this work, the digital video
broadcasting (DVB), with its standard (DVB-S2X), is employed by the forward
link to communicate from GW to UT [6].
• Return Link: On the other hand, the link from UT towards the GW is named
as return link, which is made up of the uplink of the user link, satellite, and
downlink of the feeder link. The UT transmits RF signals to the GW by using
the DVB-RCS2 standard, which has the specification for the return link where
the data traffic flow from UT to GW [7].
1.2 Multiple Spot Beam Satellite Systems
Satellite systems have been an efficient solution to cover remote and rural areas due
to their ubiquitous links. Nevertheless, it is very tough to compete with terrestrial
communications because the transmitted bit has a high cost so that it is necessary to
reduce the cost by increasing the total capacity of the satellite systems [8]. Most of
the satellite systems solely cover a large area by using a single beam, especially in L2
, C3 and Ku4 bands.
Although the area is entirely covered by a single beam, the system capacity is
limited as well as its efficiency. Currently, modern SatCom systems have multiple
spot beams which increase the system capacity by using frequency reuse among
beams. This architecture is also known as multiple beam architecture which is
inspired by traditional cellular networks. Therefore, the inter-beam interference must
be maintained within the typical values and admissible limits in order to achieve high
spectral efficiency on the satellite link.
It is important to note that the latest satellite systems are at the technological
cutting edge to cover a region by generating multiple spot beams. This means that
satellites have multiple beam antennas (MBA). Reflector MBAs are usually used
for SatComs due to their excellent RF performance in terms of coverage gain and
the carrier-to-interference ratio (CIR), payload simplicity, reduced cost, and mature
technology. These MBAs are classified into three types: (a) single reflector with
a single feed per beam, (b) single reflector with over-lapping feed clusters, and (c)
2The L band is a frequency band in the radio spectrum from 1–2 GHz. The mobile satellite
services use L band carriers where the downlink ranges from 1518.0–1559.0 MHz, 2180.0–2200.0
MHz, and 2483.5–2500.0 MHz, whereas the uplink ranges from 1610.0–1660.5 MHz, 1668.0–1675.0
MHz, and 2000.0–2025.0 MHz, [4].
3The C band in communication satellites uses the frequency band from 3.40–4.20 GHz and
4.50–4.80 GHz for the downlink, whereas the uplink ranges from 5.091–5.250 GHz and 5.850–7.075
GHz. Both frequency ranges are available for ITU region 2 (R2), [4].
4The Ku band in communication satellites uses the frequency band from 10.70–12.70 GHz for
the downlink, whereas the uplink ranges from 12.70–13.25 GHz, 13.75–14.80 GHz, and 15.43–15.63
GHz. Both frequency ranges are available for ITU region 2 (R2), [4].
CHAPTER 1. INTRODUCTION 4
multiple reflectors with a single element per beam, [9]. In particular, the reflectors
are generally offset-fed parabolic reflectors where the feed elements are horns.
Thus, Type (a) MBA has a straightforward architecture, where the reflector is
illuminated by a dedicated feed in order to generate a beam pattern due to radiating
the radio waves reflected from the antenna. This technique offers both good isolation
and high radiation efficiency between the different beams. Type (b) MBA design
needs low-level beamforming networks to provide element sharing among beams and
beam combining functions. In fact, Type (b) antenna is usually used for mobile
satellite services at low frequencies such as L-band and S-band5.
Finally, Type (c) MBA is made up of multiple reflectors where each reflector is
illuminated with its own feed array. These apertures are either 3 or 4 units, which are
independent of each other. Here, each feed horn generates a single beam, therefore,
the feed is able to provide optimal illumination on the reflector. Type (c) MBAs are
often used at frequency bands such as the Ku band, Ka-band, and Extremely High
Frequency (EHF) band.
At the satellite, each beam is generated by a specific feed horn, where the beams
with the same frequency sub-bands are radiated from the same reflector. That is, each
frequency sub-band corresponds to a segment (also known as a color), consequently,
the frequency-reuse scheme in the MBA is usually known as either three-color or
four-color schemes. Figure 1.2 details the architecture of a satellite system with a
4-color frequency-reuse scheme.
Particularly, a 4-color scheme is defined by partitioning the available frequency
and polarization resources into four segments (colors), i.e, each color belongs to
half the available bandwidth and one polarization. It is important to mention that
frequency-reuse schemes employ the circular polarization, i.e., right hand circularly
polarized (RHCP) and left hand circularly polarized (LHCP), therefore, it is available
2 polarization resources. Currently, these modern SatComs are also known as high
throughput satellite (HTS) systems.
1.3 Evolution of High Throughput Satellite Systems
The evolution of HTS systems has outperformed many stages since the appearance of
the first broadband satellite system in 2005 [10], improving impressively its capacity in
order to increase both the coverage area and the quality of service (QoS) requirements
of the end-user. Figure 1.3 compares the capacity of some HTS systems in different
years. For instance, Spaceway 3 with 24 spot beams has a capacity of 10 Gb/s.
Echostar XVII has 60 spot beams and a capacity of 120 Gb/s, whereas Echostar XIX
has 120 spot beams and a capacity of 200 Gb/s. Spaceway 3, Echostar XVII, and XIX
are part of the Hughes® HTS constellation which has the latest JupiterTM System
technology for broadband satellite access [11,12]. KA-SAT is part of the EUTELSAT
fleet which has a capacity of 90 Gb/s and 82 spot beams [13]. By 2021 the Echostar
5The S-band is part of the microwave band in the electromagnetic spectrum, whose frequency
band ranges from 2.0–4.0 GHz [4].
CHAPTER 1. INTRODUCTION 5
Figure 1.2: Architecture of a multiple spot beams satellite system with 4-color frequency
reuse scheme.
XXIV will have been launching with a capacity up to 500 Gb/s and multiple spot
beams in Ka-band.
G
ig
ab
it
p
er
s
ec
o
n
d
[
G
b
/s
]
0
100
200
300
400
500
600
HTS Generations
HTS, Spaceway 3, year: 2007 HTS, KA-SAT, year: 2010 HTS, Echostar XVII, year: 2012 HTS, Echostar XIX, year: 2016 HTS, Echostar XXIV, year: 2021
500
200
120
90
10
HTS, Spaceway 3, year: 2007
HTS, KA-SAT, year: 2010
HTS, Echostar XVII, year: 2012
HTS, Echostar XIX, year: 2016
HTS, Echostar XXIV, year: 2021
Figure 1.3: Capacity Evolution of HTS systems.
From 2021 onwards, it is estimated that next-generation HTS systems will require
capacity equal to or greater than 1 Tb/s, which will reduce the cost per transmitted
bit. However, the bandwidth is limited due to the lack of available spectrum in
the Ka-band. In the following section, three alternatives are detailed in order to
overcome the spectrum constraint and technical limitations, which are given an
excellent background for this research.
CHAPTER 1. INTRODUCTION 6
1.4 State-of-art in current HTS systems
1.4.1 Feeder Link Design in Q/V band
As stated above in Section 1.3, the capacity of HTS systems is continually increasing
due to the demand for data traffic. Therefore, it is necessary to design a reliable feeder
link in terms of higher bandwidth and frequency bands. A first alternative is to move
the feeder link from the Ka-band to the Q/V band6 whose bandwidths are available
for satellite communications [4, 8, 14]. It is important to mention that the feeder
link needs more available spectrum, for this reason, up to 5 GHz are available for
both feeder uplink and feeder downlink, respectively. As a result, the whole Ka-band
spectrum is feasible for the user link, i.e., 3.50 GHz are available for both user uplink
and user downlink, respectively. Indeed, it is a very important solution for satellite
operators, however, the feeder link turns into a very susceptible link due to the heavy
rain attenuation [8, 15]. Although the fade mitigation technique (FMT) is used as
an uplink power control (ULPC), it can only compensate a few decibels (dB) during
short-term fades.
For this reason, it is necessary to design a Q/V band feeder link with availability
in excess of 99.9% [16]. Thus, multiple GWs are used by using transmit diversity in
order to achieve the required availability [13,14]. This design of feeder links must be
reliable so that each feeder link can operate in its corresponding GW, therefore, it is
a key-feature in the development of this research.
1.4.2 Aggressive Frequency Reuse
As already mentioned, current HTS systems use a 4-color multiple spot beam
structure in order to increase the total capacity in the satellite system. Nevertheless,
each beam utilizes only 1/4 of the total resources in terms of frequency and
polarization, therefore, it is not an efficient scheme. Nowadays, a lot of research
aims to improve the conventional 4-color frequency-reuse scheme in order to achieve
full frequency reuse with advanced systems [14, 17–19]. Although the main idea is
that each beam can utilize the whole available resources, this leads to an increase in
the inter-beam interference among the adjacent beams.
In fact, it has been explored 3-color and 4-color frequency reuse scheme with
both simple and double polarization [18], as well as the incidence in the interference
among user and feeder links. To sum up, it is important to know the affectation of
these interferences in the Q/V band feeder link in order to guarantee communication
from/to the HTS system.
6Q/V band ranges from 37.50–42.50 GHz for the downlink (Q band), whereas the uplink (V
band) ranges from 42.50–51.40, GHz. These frequency bands are within the EHF band for ITU
region 2 (R2), [4].
CHAPTER 1. INTRODUCTION 7
1.4.3 Ground Segment Architecture
In Section 1.4.1, it was mentioned that the transmit diversity was necessary to
achieve an availability greater than 99.9% on the feeder link. That is, the ground
segment architecture is made up of a number of GWs, which are interconnected by
terrestrial links in order to generate an ease routing of feeder link data. Therefore,
this architecture is used in diversity way in order to mitigate fades on the feeder link
of each GW. The main advantage of this mechanism is when a GW experiences either
feeder link outage or reduced capacity, then an available GW receives all data traffic
by using terrestrial links [14].
Consequently, some diversity scheme models have been studied by several
researchers, one of them has been the 1 + 1 diversity scheme [20, 21], where a
nominal GW was backed up by another redundant GW. Although this architecture is
acceptable for low/medium throughput systems, it is not efficient for HTS systems,
where high capacities and tens of GWs are necessary.
Further, another transmit diversity scheme is also known as the N + P scheme,
which has been studied in [8, 14, 22–26]. This scheme is made up of N active or
nominal gateways (NGW), and P idle or redundant gateways (PGW). Whether one
of the NGWs is in outage, whole the traffic from the affected NGW is rerouted to one
of the PGWs by a switching mechanism. Furthermore, the smart gateway diversity
(SGD) is another scheme that has been studied in [22, 26]. The main difference
between the SGD and the N + P scheme diversity is that the SGD does not need
redundant PGWs, but its disadvantage is when one or more NGWs are in outage,
that is, the throughput of users served is reduced by either the GW or GWs affected.
However, this disadvantage can be overcome by adding more capacity for each
GW in order to support other GWs in case of an outage. Therefore, GWs need to
be oversized in terms of capacity. Moreover, the user terminals require a level of
intelligence, increasing the complexity of the network. In this thesis, the research is
only focused on the N + P scheme, therefore, the contributions are associated with
this scenario.
1.5 Objectives
The objectives of this thesis are to design, allocate, develop and optimize the GS by
cutting edge techniques for HTS systems. That is, the main focus on this thesis is
related to an adequate design of the Q/V band feeder uplink, allocating properly the
GW sites on the GS, and developing techniques to optimize the network architecture
on the GS. In this section, a brief description of specific objectives are given as follows,
• Q/V band Feeder Uplink Design: As already mentioned, the Q/V band is
very susceptible to both rain attenuation and propagation effects. Therefore,
the Q/V band feeder uplink must be designed to achieve high reliability in order
to provide new broadband services on the user link of an HTS system. Each
Q/V band feeder uplink provides a high bandwidth communication to/from the
HTS. Therefore, this thesis is focused on the design of a reliable feeder uplink for
CHAPTER 1. INTRODUCTION 8
each GW in order to operate over the typical availability of an N +P diversity
scheme on the GS.
• Allocating GW Sites on Ground Segment: An important part of this
research is to analyze if the site for each GW is adequate to be considered in
an N + P diversity scheme. The performance of this scheme is assessed by
a correlated rain fading channel and its impact on the distance between two
or more GWs. For this reason, the GW sites are modeled by implementing
a dynamic rain attenuation model. It is important to mention that the rain
attenuation model is brought by the recommendations of the International
Telecommunication Union (ITU). As a result, the total outage probability is
determined by a theoretical analysis for each N + P simulated scenario. Also,
the effects of other parameters related to the dynamic rain attenuation model
are observed and studied by several N + P simulations.
• Optimization Techniques for Ground Segment Architecture: Very few
works that studied the N + P scheme were developed by deep mathematical
analysis. However, it is very important to develop a rigorous mathematical
analysis in order to detect and to find the switching-rate performance between
N and P GWs. The switching parameter is a key-feature in the system
model, which relates directly to the outage probability in the HTS system.
Therefore, the lack of a rigorous analysis taking into account dynamic rain
attenuation characteristics, spatial correlation between GWs, and switching rate
requirements, it motives strong research on the GS for multiple GWs using the
Q/V band. In this context, machine learning techniques are implemented in
this study in order to optimize the total number of GWs, which are assured
efficient switching with an optimal N + P scheme.
1.6 Methodology
In this thesis, the quantitative method is employed by mathematical analysis,
statistical data, and computational techniques. The scientific rigor is based on the
reliability and validity of data. For this purpose, several models are developed by
statistical analysis to later be handled and submitted to an experimental method by
advanced computational techniques based on machine learning algorithms. In this
context, due to the lack of a historical rainfall database in the gateway locations to be
used in this thesis, the ITU Recommendations provides rain attenuation time-series by
statistical analysis. Thus, these rain attenuation time-series are employed to develop,
generate, train, and validate rain-attenuation models by machine learning algorithms
in order to predict rain attenuation events in advance. Currently, computational tools
offer up high performance and speed to run and solve complex algorithms.
All found data will be collected and their statistical/computational treatment
results as well as all relevant results in relation to the research problem of this
thesis. These outcomes are also employed to propose, design, and evaluate an efficient
CHAPTER 1. INTRODUCTION 9
switching mechanism for sizing and optimizing the ground network segment by smart
techniques based on the predictive rain-attenuation models and the Markov process.
Finally, all obtained data from this research are depicted and tabulated by figures
and tables which explain and show the nature and behavior of the outcomes from
each process in more detail. However, all results are discussed in their respective
sections by comparative and analytic methods with other models and researches.
1.7 Contributions of the Research
The contributions of this thesis include modeling the rain attenuation over satellite
feeder links, predictive modeling based on machine learning algorithms, and
developing optimization techniques for the Ground network segment. A brief
description of each of them is detailed as follows:
• The satellite feeder link is modeled in this thesis by mathematical analysis.
Further, other models are added to the feeder link model to generate an
artificial rain-attenuation dataset for specific geographic locations. This model
is essential to find and understand the availability of each gateways station in
addition to the ground network availability.
• The available electromagnetic spectrum, maximum capacity, and schemes of
the frequency reuse of the high throughput satellite system were approached by
theoretical and numerical analysis. For this reason, code scripts using Matlab
and Python were developed in this thesis in order to find and evaluate the
results in these processes.
• The rain-attenuation dataset is employed to train and validate predictive models
by the implementation of Machine Learning algorithms in order to determine in
advance when the feeder link is impaired by the heavy rain. For this purpose,
code scripts using Python/Tensorflow were developed in this thesis to find and
explain the model outcomes.
• Finally, an efficient switching scheme mechanism for the N+P gateway scenario
is presented in this thesis. By using predictive rain-attenuation models obtained
from Machine Learning algorithms and the Markov Chain process, it is possible
to define the number of redundant gateways is necessary to maintain the network
availability above 99.9%. These processes optimize the Ground network segment
as a function of the number of redundant gateways.
1.8 Thesis Structure
This thesis is organized as follows: Chapter 2 explains the main background concepts
and system models to implement in this thesis. In Chapter 3, the quantification of
multiple spot beams and the total capacity of the HTS system is detailed. Chapter
4 evaluates the interferences in frequency reuse by using offset-parabolic-reflector
CHAPTER 1. INTRODUCTION 10
antennas. The deep learning models, developments, implementations, results, and
discussions are provided in more detail by Chapter 5. In Chapter 6, the optimization
of the Ground Segment is presented and discussed by smart and predictive switching
mechanisms. Finally, Chapter 7 concludes this research and its proposals, in addition
to remarking about future research studies. Contributions and scientific productions
of this thesis are also detailed.
Chapter 2
System Models for the Ground
Segment
2.1 Introduction
In this chapter, system models that are used throughout this thesis are introduced and
explained. Systems models were briefly mentioned in Chapter 1, showing a problem
statement in each case. Therefore, each system model is described in mathematical
detail providing an excellent background to the different objectives of this thesis.
The feeder uplink model is presented in Section 2.1, where the mathematical
analysis and the main impairments in the feeder uplink are minutely discussed.
Moreover, the system parameters are provided to find and discuss the problems
that are associated with the Q/V band uplink budget. In Section 2.2, the spatial
correlation model is described. This model is based on the ITU-R recommendations,
which is used to determine a minimum separation between GWs in order to avoid the
rain attenuation in two sites or more at the same time. The dynamic rain attenuation
model is presented in Section 2.3, which also is based on an ITU-R recommendation.
This model is used to obtain the rain attenuation in each site of the GS. Finally, the
N + P diversity model is explained in Section 2.4 which is implemented on the GS
by simulated scenarios to understand the behavior of the network.
2.2 Feeder Uplink Model
Each gateway station (GW) is distributed and separated from other GWs by a
distance of D km on the ground segment (GS). Hence, each satellite feeder uplink
belongs to its respective i-th GW, transmitting the signal si(t), where si(t) is a
function that varies with time. It is important to mention that i indicates the i-th
GW site. In this thesis, the modeled satellite-link is the feeder uplink which is between
the i-th GW and the high throughput satellite (HTS). It is worth noting that the
feeder uplink operates at the EHF band because there is more available spectrum.
The frequency band is ranged from 42.50–51.40 GHz which is also known as the V
band. This model is based on the Gharanjik’s feeder link model [21], but with slight
11
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 12
differences according to the aim of this study.
To begin with, the average power of the feeder-uplink signal is defined as
Ei = E{|si(t)|
2}, for i = 1, 2, . . . , N, (2.1)
where N is the number of GWs on the ground network. Thus, the set of GWs,
{1, 2, . . . , N}, can also be represented by the set notation [N ]. The continuous signal
si(t) is sampled by measuring the value of the continuous function every Ts seconds or
minutes, i.e., it depends on the sampling period to be used. As a result, the sampled
signal is given by si[k], where k is the index of each sample, being an integer value,
i.e., k = 1, 2, . . . , K, whereas K is the total number of samples, and Ts is the time
interval between samples or sampling period. In this case, the numeric value of the
k-th number in the sequence is equal to the value of the continuous signal, si(t), at
time t = kTs, i.e., si[k] = si(kTs). By using this analysis, the communication channel
between an i-th GW and the satellite at t = kTs can be represented by
hi[k] = |hi[k]|e
j↵i , for i = 1, 2, . . . , N, (2.2)
where ↵i is the phase component. Moreover, the channel amplitude |hi[k]| is estimated
by a signal beacon from the satellite. The channel expression is obtained from [25],
which is explained in more detail. Further, the clear sky carrier-to-noise ratio (CNR)
for each feeder uplink at the satellite receiver is expressed as
csi =
Ci
Ni
=
Ei
Ni
1
1 + ro
, for i = 1, 2, . . . , N, (2.3)
where Ni is the noise variance at satellite interface and ro is the roll-off parameter.
Here, the roll-off factor is a measure of the excess bandwidth of the filter, that is,
the bandwidth occupies beyond the Nyquist bandwidth. For this purpose, the roll-off
value, either 5% or 10%, is obtained from the Digital Video Broadcasting Second
Generation with Extensions (DVB-S2X) standard [6]. It is important to note that
the DVB-S2X standard is considered as the baseline air interface for the forward link.
Based on the previous equations, the actual CNR for the feeder uplink between an
i-th GW and the satellite at t = kTs for k = 1, 2, . . . , K, is denoted by
Γk,i = |hi[k]|
2Ci
Ni
= |hi[k]|
2 csi , for i = 1, 2, . . . , N, (2.4)
where each value of Γk,i can be stored in a column vector Γ :,i for i = 1, 2, . . . , N,
which defines a point in a K-dimensional space called Euclidean space denoted
by R
K , therefore, the matrix of CNR can be expressed by Γ 2 R
K⇥N . However,
this case is only when the feeder uplink is not affected by either interferences or
weather impairments, such as rain, clouds, and fog. Therefore, it is necessary to
obtain the carrier-to-noise and interference ratio (CNIR) using both the CNR and
the carrier-to-interference ratio (CIR). It is important to mention that CIR is one
of the most serious challenges due to the fact that interferences degrades the signal
quality, impairing directly to communication system. The CIR can also be represented
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 13
by the symbol ⇣.
The total CIR, ⇣T , directly impacts on the feeder uplink, which is made up
of the co-channel interference, ⇣co, adjacent channel interference, ⇣adj, and the
intermodulation interference, ⇣im, [8]. The ⇣T can be calculated by
1
⇣T
=
1
⇣co
+
1
⇣adj
+
1
⇣im
(2.5)
Finally, the CNIR for k = 1, 2, . . . , K is defined as
Ξk,i =
Γk,i · ⇣T
Γk,i + ⇣T
, for i = 1, 2, . . . , N, (2.6)
where each value of Ξk,i is also stored in a K-dimensional column vector, Ξ:,i 2 R
K
for i = 1, 2, . . . , N . The matrix of CNIR is given by Ξ 2 R
K⇥N . To sum up, the
signal quality is defined by the CNIR value for each k-th sample, therefore, this is
a good measure to know whether the feeder uplink can be able to transmit to the
satellite receiver, or if the weather conditions affect the feeder link, then it can not
transmit the signal to the satellite, degrading the spectral efficiency on the channel.
2.2.1 Parameters of the Feeder Uplink
Many parameters are involved in the performance of the Q/V band feeder uplink.
Table 2.1 shows the basic propagation parameters for the feeder uplink between a
GW station and the HTS system. These parameters were used in link simulations
for each GW station of the GS. Currently, the BRASILSAT B4 satellite is positioned
at 92.0° West. Although the satellite has exceeded its lifespan (19 years), the orbital
position is only used for HTS simulation purposes of HTS systems. Furthermore, the
specific parameters of the system and the link budget are detailed in Chapters 3 and
4.
Table 2.1: Basic Propagation Parameters
Feeder Uplink Values
Frequency 50 GHz
Orbital Position 92.0°W
Orbit GEO, d0 = 35786 km
Polarization LHCP, RHCP
Boltzmann’s constant =1.380 649⇥ 10 23 J/m
In order to determine elevation, azimuth, and range between each GW station and
the HTS system, it is essential to use geographical coordinates for each site (latitude
and longitude) [1,2]. With these parameters, the transmitter antenna of each GW is
adequately pointed towards the HTS receiver antenna.
It is important to mention that each feeder uplink is also affected by the user
downlinks, therefore, each forward link (Feeder Uplink + User Downlink) is analyzed
and calculated. Thus, it is assumed as a mixed Ka and Q/V band solution [8,14,18].
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 14
Although the study of the Ka-band user link is not the aim of this thesis, it is necessary
to know its implication in this system. The Ka-band implications are approached in
more detail in Chapter 4.
2.3 Spatial Correlation Model
Spatial correlation is a measure that analyzes the relationship between near spatial
units. For this purpose, the Recommendation ITU-R P.1815 indicates the method to
obtain the spatial correlation model, which is able to estimate the correlation of rain
rate to determine differential rain attenuation on satellite paths between the HTS and
multiple locations on the surface of the Earth [27]. The Recommendation ITU-R P.618
includes the spatial correlation model [28], (See §2.2.4.1), whose objective is generally
to calculate propagation data for the design of Earth-space telecommunication
systems. Here, a set of ITU Recommendations were neatly employed to implement the
obtained spatial-correlation-model for our study [29–35]. It is important to mention
that the use of Recommendations ITU-R 1815 and ITU-R 618 are valid for antenna
elevations above 10 , frequencies up to 55 GHz, and site separations D between 0 and
at least 250 km.
Log-normal distribution of rain intensity and rain attenuation level are assumed
by the differential rain attenuation method [27, 28]. Thus, this method determines
the joint probability (%) that the attenuation on the path to the first site is greater
than the attenuation threshold (a1) and the attenuation on the path to the second
site is greater than the attenuation threshold (a2). The probability is the product of
two joint probabilities and is expressed as,
Pr{A1 a1, A2 a2} = 100⇥ Pr ⇥ Pa, (2.7)
where Pr is the joint probability that it is raining at both sites, Pa is the conditional
joint probability that the A1 and A2 attenuations exceed a1 and a2, respectively [27].
In general terms, the attenuation threshold for every i-th site can be denoted as ai.
The discrete rain attenuation process and the channel gain, for each i-th site and
k = 1, 2, . . . , K, are related as
Ai[k] = 10 log10 |hi[k]|
2, for i = 1, 2, . . . , N (2.8)
Both probabilities (Pr and Pa) are complementary bivariate normal distributions,
which are expressed as
Pr =
1
2⇡
p
1 ⇢2r
Z 1
R1
Z 1
R2
exp
✓
r21 2⇢rr1r2 + r22
2(1 ⇢2r)
◆
dr2 dr1, (2.9)
Pa =
1
2⇡
p
1 ⇢2a
Z 1
ln a1−mlnA1
σlnA1
Z 1
ln a2−mlnA2
σlnA2
exp
✓
b21 2⇢ab1b2 + b22
2(1 ⇢2a)
◆
db2 db1, (2.10)
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 15
where r1 and r2 are independent variables of the function Pr whereas b1 and b2 are
independent variables of the function Pa. The correlation coefficients ⇢r and ⇢a are
respectively denoted as
⇢r = 0.7 exp
✓
D
60
◆
+ 0.3 exp
✓
D
700
◆2⌫
, (2.11)
⇢a = 0.94 exp
✓
D
30
◆
+ 0.06 exp
✓
D
500
◆2⌫
. (2.12)
In Eq. (2.11) and Eq. (2.12), D is the distance between two sites in km, whereas
in Eq. (2.9), the thresholds R1 and R2 are the solutions of rain probability (P rain
i )
for a particular i location, which is calculated from step 3 of Annex 1 in [29] by the
following equation:
P rain
i = 100⇥Q(Ri) = 100⇥ 1p
2⇡
Z 1
Ri
exp
✓
r2
2
◆
dr, for i = 1, 2 . . . , N. (2.13)
Once obtained P rain
i (%), it is possible to calculate the threshold Ri using the inverse
complementary cumulative normal distribution, Q 1, as follows Ri = Q 1(P rain
i /100).
Finally, the parameters mlnA1 , mlnA2 , lnA1 and lnA2 are computed by fitting each
i-th site rain attenuation, Ai, versus probability of occurrence, Pi, to the log-normal
distribution:
Pi = P rain
i Q
✓
lnAi mlnAi
lnAi
◆
, for i = 1, 2, . . . , N. (2.14)
This method is described in §2.2.1.1 from [28].
Nevertheless, the inclusion of spatial correlation makes the analysis very complex
so that it is necessary to assume spatially independent and identically distributed
(i.i.d) links. Indeed, a typical diversity scheme of multiple gateways has a few
tens of GWs dispersed over the GS, where each i-th site is only separated a few
tens kilometers from each other. In a sense, the rain attenuation at each i-th site
must be uncorrelated to avoid link outages in two or more GWs at the same time.
Therefore, applying the correlation coefficient, ⇢a, by Eq. (2.12), it is possible to find
uncorrelated distances for ⇢a 0.1 [25]. Figure 2.1 depicts the correlation coefficient
⇢a between two sites, where the correlation drops to 0.1 at a distance of 80 km,
therefore, it takes solely a few tens of kilometers to carry through the decorrelation
of rain attenuation.
2.4 The Dynamic Rain Attenuation Model
In the Q/V band, the link impairment is mainly related to the rain attenuation, which
is traditionally modeled and validated by using a log-normal distribution [36]. Eq.
(2.8) expresses the relation between the rain attenuation and the channel gain, where
hi[k] is the communication channel between a GW and the HTS. However, the rain
attenuation is a time-varying process so that it is necessary to model the dynamic
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 16
Figure 2.1: The correlation coefficient, ⇢a, between two sites separated by a distance of D
km.
behavior of the rain attenuation.
Despite that several time-series models to synthesize rain attenuation samples
with temporal properties have been studied and proposed, only one model has been
stood out [3]. This model was based on the Maseng-Bakken model [37], which
was implemented as a new recommendation by the International Telecommunication
Union (ITU), ITU-R P.1853 [38].
The ITU Recommendation P.1853 details how the time series synthesis method
generates a time series that reproduces the spectral characteristics, fade slope and
fade duration statistics of rain attenuation events. Also, the statistics of the inter-fade
duration are reproduced but solely within individual events of the rain-attenuation.
Figure 2.2 depicts the block diagram of this process, for this purpose, the rain
attenuation, A(t), is synthesized from the white-Gaussian-noise discrete process, n(t).
Here, it is important to mention that A(t) is a continuous signal. In this model,
the white-Gaussian-noise is filtered by a low-pass filter, which is transformed from
a normal distribution to a log-normal distribution in a memoryless nonlinearity, in
addition to being calibrated to match the desired rain-attenuation statistics [38].
The time-series synthesizer is made up of five features, where m is the mean of the
log-normal rain attenuation distribution, is the standard deviation of the log-normal
rain attenuation distribution, p is the probability of rain, is the parameter that
describes the time dynamics (s 1), and the Aoffset that adjusts the time-series to
match the probability of rain (dB). This time-series synthesizer is employed in this
thesis, therefore, the following procedure is the basis for effective implementation.
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 17
k
p+
Low-pass
Filter
exp (m+ [X(t)])
Memoryless Non-linear
Device
Aoffset
Calibration
n(t)
White
Gaussian
Noise
X(t) A(t)
Rain
Attenuation
(dB)
Figure 2.2: Block diagram of the rain attenuation time series synthesizer, ITU-R
P.1853.
2.4.1 Procedure to Synthesize the Rain Attenuation
Time-series
The following step-by-step method is used to synthesize the attenuation time-series
Ai(kTs) for k = 1, 2, . . . , K, and for i = 1, 2, . . . , N , which is based on the ITU
Recommendation P.1853 [38]. The suggested method is apportioned as follows:
A. Estimation of m and σ
The parameters m and are obtained by the cumulative distribution of rain
attenuation vs. probability of occurrence. In most cases, the development process
is extremely difficult due to the lack of rain attenuation statistics from measured
data in situ. However, the ITU Recommendation P.618 [28] provides excellent
rain attenuation data prediction for Earth-space paths, which are also known as
artificial rain attenuation data. Therefore for the path and frequency of interest,
it was performed a log-normal fit of the rain attenuation vs. probability of
occurrence by using the ITU Recommendation P.1057, where the step-by-step
procedure to approximate a complementary cumulative distribution by a log-normal
complementary cumulative distribution is described in detail [39]. As a result of this
procedure, it was possible to obtain m and .
B. Low-pass filter parameter
The parameter is configured as, = 2⇥ 10 4 s 1
C. Attenuation offset
The attenuation offset for each i-th site is estimated by
Aioffset = exp
m+ Q 1
✓
P rain
i
100
◆
, for i = 1, 2, . . . , N (2.15)
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 18
D. Time-series synthesis
The rain attenuation time-series, Ai(kTs), for k = 1, 2, . . . , K, and for i = 1, 2, . . . , N ,
is synthesized as follows:
• Step D1: A white-Gaussian-noise time-series, ni(kTs), is synthesized for each
i-th GW site, where k = 1, 2, . . . , K, with m{ni(kTs)} = 0 and Var{ni(kTs)} = 1
at a sampling period, Ts. It is important to mention that the sampling period
value can be modified, however, the default value is 1 s.
• Step D2: Set X(0) = 0
• Step D3: The noise time-series, ni(kTs), for k = 1, 2, . . . , K, is filtered by a
recursive low-pass filter, which is expressed by
Xi(kTs) = ⇢X[(k 1)Ts] + ni(kTs)
p
1 + ⇢2, for i = 1, 2, . . . , N, (2.16)
where ⇢ = exp ( Ts).
• Step D4: Yi(kTs) for k = 1, 2, . . . , K, and for each i-th GW site, is defined as
Yi(kTs) = exp [m+ Xi(kTs)], for i = 1, 2, . . . , N (2.17)
• Step D5: The dynamic rain attenuation, Ai(kTs) (dB) for k = 1, 2, . . . , K, and
for each i-th GW site, is computed as follows
Ai(kTs) = max bYi(kTs) Aioffset , 0c, for i = 1, 2, . . . , N (2.18)
• Step D6: The first 200000 samples are discarded from the synthesized
time-series which correspond to the filter transient. Here, rain attenuation
events are represented by sequences where, for a consecutive number of samples,
values are above 0 dB. Consequently, Eq. (2.8) can be replaced by Eq. (2.18) in
order to simulate a dynamic rain attenuation. It is necessary to mention that
Eq. (2.18) can be expressed as a discrete signal, Ai[k], where each k-element
of rain attenuation, Ak,i, can be stacked in a K-dimensional column vector,
A:,i 2 R
K for i = 1, 2, . . . , N,. Finally, the matrix of rain-attenuation is given
by A 2 R
K⇥N .
2.5 N + P Diversity Model on Ground Segment
As was previously mentioned in Chapter 1, an important drawback was the seriously
increased rain attenuation in the Q/V band compared with Ka-band. Therefore, site
diversity is introduced in order to mitigate the impact of rain attenuation [22,25,26].
This kind of diversity model uses generally spatial diversity of the feeder links, which
has feeder links spatially independent and identically distributed, i.i.d, as already
mentioned in Section 2.2. As a result, there is a very small probability that all the
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 19
GWs experience heavy rain fading at the same time. However, if a GW is affected
by the rain attenuation, another GW under clear sky conditions can take over the
capacity from that affected GW [8,14,18,26].
For this process to be possible, the network must have extra resources, e.g. extra
GWs, which are also known as redundant or backup PGWs. In other words, this
process is also known as the switching mechanism between NGWs and PGWs. The
switching mechanism can be performed if a certain switching threshold is reached,
which can be a capacity threshold, a carrier-to-noise and interference ratio (CNIR)
threshold or a combination of these two cases. In summary, when the switching
threshold is reached, the capacity of the NGW affected is transferred (handover) to
another PGW with clear sky or in much better conditions [21,23,24].
Figure 2.3 shows the gateway diversity scheme to be used in this thesis, which is
known as the smart gateway diversity technique 3 and is gathered from [26]. When
there are good weather conditions, the nominal gateways (NGW1 and NGW2) use
the total available bandwidth allowed by the ITU radio regulations [4], as shown in
Figure 2.3a. However, when the feeder link of the NGW1 is affected by a rain fading
event, that capacity is taken over by the redundant gateway PGW1, as illustrated in
Figure 2.3b. Therefore, this model expresses adequately the N + P diversity model,
which has the following properties:
• from the system aspect: the system capacity does not degrade unless
too many gateways are impacted at the same time. It is thus important to
carefully select switching thresholds and carefully perform system dimensioning.
End-to-end availability is improved;
• from the user aspect: user terminals served by the impacted gateway do not
need to switch to a different carrier. They experience a short switching time
because of switching traffic from one GW to the other [26].
Altogether, the N+P model is used to establish a solid base for this thesis in order
to approach from a mathematical point of view to the optimization of this model.
CHAPTER 2. SYSTEM MODELS FOR THE GROUND SEGMENT 20
Beam 3
Beam 2
Beam 1
Beam 6
Beam 5
Beam 4
Empty
Empty
Empty
HTS
NGW
1
NGW
2
PGW
1
Beam
1
Beam
2
Beam
3
Beam
4
Beam
5
Beam
6
(a) Operation in clear sky conditions
HTS
NGW
1
NGW
2
PGW
1
Beam
1
Beam
2
Beam
3
Beam
4
Beam
5
Beam
6
N/A
N/A
N/A
Beam 6
Beam 5
Beam 4
Beam 3
Beam 2
Beam 1
(b) Operation under rain conditions
Figure 2.3: Gateway diversity scheme with a redundant GW.
Chapter 3
Quantification of Multiple Spot
Beams for HTS Systems in Ka and
Q/V bands
3.1 Introduction
In this Chapter, it is proposed a methodology to quantify the necessary number of
HTS spot beams in order to achieve a capacity of 1 Tb/s for the Latin America
region by using an HTS system under ideal conditions. Firstly, a mix solution is
proposed in order to obtain more available spectrum. Therefore, the Q/V band is
used by the Feeder link whereas Ka-band is used by the User link. In this sense,
two mechanisms are introduced to this system to find the precise number of spot
beams for both Feeder and User links respectively. Moreover, traffic configurations
are analyzed by simulations in order to project future needs in terms of capacity. The
system capacity is important to supply multiple data services to/from UT such as
HD streaming, Big Data, IoT, M2M, etc. At the end of this Chapter, it was assured
the theoretical capacity for the HTS system in order to face the main challenges by
2021 and beyond.
3.2 Connectivity in Latin America: A Review
Latin America is one of the biggest regions in the world either in population
or territory, however, there is a strong social, economic, and technological gap,
influencing especially in the Internet connectivity in urban and rural areas with
noteworthy differences. Based on the above, an HTS system could cover this region,
close the current gap, and reach remotes areas where terrestrial infrastructures do
not deploy their services due to economic and technological factors.
The Internet penetration rate is crucial to size any communication system so that
it is also useful in an HTS system. In the last decade, the percentage of inhabitants
that use the Internet has increased by 36% on average. In 2016 the 56% of inhabitants
accessed to the Internet in Latin America [40]. Moreover, the Latin American region
21
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .22
is characterized by a high heterogeneity between countries. For instance, the number
of households connected to the Internet grown by about 103% between 2010 and 2016,
however, more than 50% of total households have no Internet connections. For the
sake of brevity, the big economic differences between social classes are the reason for
the lack of internet connections in Latin America, however, the thesis does not delve
into this topic.
Thus, the countries with the highest growth rate of households connected to the
Internet in the period 2010–2016 were Nicaragua, Guatemala, Honduras, and Bolivia,
which had a lower Internet penetration rate at the beginning of the period. Costa
Rica had the highest growth in the total Internet penetration of households (from 24%
to 65%), as it is shown in Figure 3.1. There is a lot of difference between urban and
P
er
ce
n
ta
g
e
o
f
H
o
u
se
h
o
ld
s
[%
]
0
5
10
15
20
25
30
35
40
45
50
55
60
65
C
R
I
U
R
Y
C
H
L
A
R
G
B
R
A
P
A
N
C
O
L
M
E
X
V
E
N
E
C
U
P
R
Y
B
O
L
P
E
R
H
N
D
G
T
M
S
L
V
N
IC
C
U
B
H
T
I
Year 2016
Year 2010
Figure 3.1: Percentage of households with Internet access per country [40].
rural zones in terms of households with an Internet connection. Figure 3.2 illustrates
the high inequality between urban and rural households due to social and economic
differences in the Latin America region in addition to higher costs of infrastructure
deployment in rural areas than urban areas.
As a result, there were 18 million rural households without Internet access in the
year 2016, whereas by the year 2020 there will be 20 million rural households without
Internet access for all Latin America region. For this reason, it is necessary to reduce
the social gap between urban and rural zones generating more opportunities in these
areas [40].
Based on above analysis, it is convenient to use high throughput satellite (HTS)
systems in order to cover remote zones where the terrestrial infrastructure cannot
reach them and be able to face to the access demand (getting higher and higher).
This satellite infrastructure can be used either as a high-performance backhaul or
be part of a non-geostationary (NGSO) satellite fleet. No doubt this leads to new
services and markets in satellite communications and other service providers, being
mainly integrators of both Shared and 5G networks by the year 2020 [41–48].
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .23
P
er
ce
n
ta
g
e
o
f
H
o
u
se
h
o
ld
s
[%
]
0
5
10
15
20
25
30
35
40
45
50
55
60
65
70
C
H
L
C
R
I
P
A
N
U
R
Y
C
O
L
M
E
X
B
R
A
JA
M
E
C
U
P
R
Y
P
E
R
S
L
V
B
O
L
Rural zone, year 2015
Urban zone, year 2015
Figure 3.2: Percentage of households with Internet access per geographical zone [40].
3.3 Available Spectrum
Nowadays, the Ku-band does not have sufficient available spectrum for an HTS
system, therefore, it is necessary to go to upper bands either the Ka-band or the
novel Q/V band, providing enough spectrum in order to achieve a capacity of
satellite system above 1 Tb/s. There are some satellite services assigned to the
radio spectrum, for instance, Fixed Satellite Service (FSS), Mobile Satellite Service
(MSS), Inter-satellite links, Space Research, Radioastronomy, Earth Exploration by
Satellite, Standard Frequency and Time Signal-Satellite (SFTSS), and Broadcast
Satellite Service (BSS). Also, the satellite services are sharing spectrum with other
terrestrial services as Mobile Service (MS) and Fixed Service (FS) [4].
3.3.1 Ka-band Radio Regulations
The current generation of HTS systems is exclusively dedicated to using the Ka-band,
where the downlink ranges from 19.70–20.20 GHz and the uplink ranges from
29.50–30.00 GHz. That is, there is an available bandwidth of 2⇥ 500 MHz given by
ITU Region 2 (R2) [4]. The rest of the frequency allocations in Ka-band are sharing
with other services that depend on regulations in each country and coordination
between operators.
• Downlink Frequency Bands: This frequency band ranges from 17.70–19.70
GHz and shares services such as FSS, MSS, and BSS, however, in some
countries could be different assignments so that it is necessary to have adequate
coordination. The frequency band of 20.20–21.20 GHz is assigned to military
services, therefore, it could be restricted in some countries. It is important
to mention that both frequency bands of 18.30–19.30 GHz and 19.70–20.20
GHz are available for High-Density application in the Fixed Satellite Services
(HDFSS), which enables to deploy a large number of earth stations (VSATs).
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .24
Table 3.1 indicates the frequency allocations for satellite downlink in Ka-band
given by ITU Region R2 [4].
• Uplink Frequency Bands: In this case, the frequency band ranges from
24.75–25.25 GHz that is exclusively for FSS, therefore, this frequency band
is not shared with other services. The frequency band of 27.00–29.50 GHz is
shared with other services such as FS, MSS, and Space Exploration. Finally,
both frequency bands of 28.35–29.10 GHz and 29.25–30.00 GHz are available
for HDFSS. Table 3.2 indicates the frequency allocations for satellite uplink in
Ka-band given by ITU Region R2 [4].
3.3.2 Q/V band Radio Regulations
In this case, it does not exist exclusive assignments for FSS services, but it requires
coordination for each portion of the spectrum since the ITU suggests that each portion
must be assigned to services such as FSS, BSS, MSS, FS, MS, radioastronomy, space
research, and exploration.
• Downlink Frequency Bands (Q band): This Q band spectrum is
divided into two frequency bands, where the first frequency band ranges
from 37.50–39.50 GHz, sharing services such as FSS, FS, research, and space
exploration. Meanwhile, the frequency band ranges from 40.00–42.50 GHz is
sharing services such as FS, FSS, MSS, BSS, research, and space exploration so
that this frequency band needs a lot of coordination between regulatory entities
in each country. Table 3.3 shows the frequency band portions for satellite
downlinks in the Q band for ITU Region R2 [4].
• Uplink Frequency Bands (V band): In contrast to the Q band, the V
band is divided into three frequency bands, therefore, the first frequency band
ranges from 42.50–43.50 GHz, which is shared by other services such as FSS,
FS, and radioastronomy. The second frequency band is assigned to both FSS
and FS, which ranges from 47.20–50.20 GHz, however, each regulatory entity
could issue restrictions in this frequency band. Finally, the third frequency
band ranges from 50.40–51.40 GHz, where its spectrum assignment is similar
to the previous frequency band. Table 3.4 denotes the frequency band portions
for satellite uplinks in the V band for ITU Region R2 [4].
As a result, in the Ka-band are available of 3.500 GHz of the spectrum for downlink
and 3.500 GHz for uplink, whereas in the Q/V band are available 5.000 GHz of the
spectrum for downlink and 5.000 GHz for uplink, respectively. In both cases, the
availability must be subject to coordination from regulatory entities.
As stated above, a mixed solution was proposed in this Thesis, i.e., Ka-band was
used for the User Link and the Q/V band was used for the Feeder Link. Nevertheless,
it was not possible to use the entire available spectrum, therefore, it was assumed
that the analysis was totally theoretical so that 3.000 GHz were available for both
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .25
T
ab
le
3.
1:
F
re
qu
en
cy
A
ll
oc
at
io
n
s
fo
r
S
at
el
li
te
D
ow
n
li
n
k
in
K
a-
b
an
d
,
IT
U
R
eg
io
n
R
2
[4
].
K
a
-b
a
n
d
D
o
w
n
li
n
k
17
.7
0
–
17
.8
0
[G
H
z
]
17
.8
0–
18
.4
0
[G
H
z
]
18
.4
0–
18
.6
0
[G
H
z
]
18
.6
0–
18
.8
0
[G
H
z
]
18
.8
0–
1
9.
3
0
[G
H
z
]
1
9.
3
0–
1
9.
7
0
[G
H
z
]
1
9.
7
0
–
2
0.
2
0
[G
H
z
]
2
0
.2
0–
2
1.
2
0
[G
H
z
]
B
W
[G
H
z
]
F
S
S
(s
-E
)/
B
S
S
/F
S
0.
10
0
0
.1
0
0
F
S
S
(s
-E
)/
F
S
0.
60
0
0.
20
0
0
.5
0
0
0
.4
0
0
1
.7
0
0
F
S
S
(s
-E
)/
E
ar
th
E
x
p
.
S
at
./
F
S
/M
ob
il
e/
S
p
ac
e
R
es
ea
rc
h
0
.2
00
0
.2
0
0
F
S
S
(s
-E
)
0
.5
0
0
0
.5
0
0
F
S
S
(s
-E
)/
M
S
S
(s
-E
)/
S
F
T
S
S
1
.0
0
0
1
.0
0
0
A
v
a
il
a
b
le
B
a
n
d
w
id
th
3
.5
0
0
N
o
t
e
:
S
o
m
e
ba
n
d
w
id
th
s
re
qu
ir
e
co
o
rd
in
a
ti
o
n
w
it
h
o
th
er
se
rv
ic
es
,
es
pe
ci
a
ll
y
F
ix
ed
S
er
vi
ce
s
(F
S
).
T
ab
le
3.
2:
F
re
qu
en
cy
A
ll
oc
at
io
n
s
fo
r
S
at
el
li
te
U
p
li
n
k
in
K
a-
b
an
d
,
IT
U
R
eg
io
n
R
2
[4
].
K
a
-b
a
n
d
U
p
li
n
k
24
.7
5
–
25
.2
5
[G
H
z
]
27
.0
0–
27
.5
0
[G
H
z
]
27
.5
0–
28
.5
0
[G
H
z
]
28
.5
0–
29
.1
0
[G
H
z
]
2
9.
1
0–
2
9.
5
0
[G
H
z
]
2
9.
5
0–
2
9.
9
0
[G
H
z
]
2
9.
9
0–
3
0.
0
0
[G
H
z
]
B
W
[G
H
z
]
F
S
S
(E
-s
)
0
.5
00
0.
5
0
0
F
S
S
(E
-s
)/
F
S
/I
n
te
r-
S
at
el
li
te
s/
M
S
S
0
.5
00
0.
5
0
0
F
S
S
(E
-s
)/
F
S
/M
S
S
1.
00
0
1.
0
0
0
F
S
S
(E
-s
)/
E
ar
th
E
x
p
.
S
at
./
F
S
0.
60
0
0.
4
0
0
1.
0
0
0
F
S
S
(E
-s
)/
E
ar
th
E
x
p
.
S
at
.
0.
4
0
0
0.
1
0
0
0.
5
0
0
A
v
a
il
a
b
le
B
a
n
d
w
id
th
3
.5
0
0
N
o
t
e
:
S
o
m
e
ba
n
d
w
id
th
s
re
qu
ir
e
co
o
rd
in
a
ti
o
n
w
it
h
o
th
er
se
rv
ic
es
,
es
pe
ci
a
ll
y
F
ix
ed
S
er
vi
ce
s
(F
S
).
T
ab
le
3.
3:
F
re
qu
en
cy
A
ll
oc
at
io
n
s
fo
r
S
at
el
li
te
D
ow
n
li
n
k
in
Q
b
an
d
,
IT
U
R
eg
io
n
R
2
[4
].
Q
b
a
n
d
D
o
w
n
li
n
k
37
.5
0–
38
.0
0
[G
H
z
]
38
.0
0–
39
.5
0
[G
H
z
]
39
.5
0
–
40
.0
0
[G
H
z
]
40
.0
0–
4
0
.5
0
[G
H
z
]
4
0.
5
0–
4
1.
0
0
[G
H
z
]
4
1.
0
0
–
4
2.
5
0
[G
H
z
]
B
W
[G
H
z
]
F
S
S
(s
-E
)/
F
S
/S
p
ac
e
R
es
ea
rc
h
/M
S
S
/E
ar
th
E
x
p
.
S
at
.
0.
50
0
0
.5
0
0
F
S
S
(s
-E
)/
F
S
/M
S
/E
ar
th
E
x
p
.
S
at
1.
50
0
1
.5
0
0
F
S
S
(s
-E
)/
F
S
//
M
S
/M
S
S
/E
ar
th
E
x
p
.
S
at
.
0.
50
0
0
.5
0
0
F
S
S
(s
-E
)/
F
S
//
M
S
/M
S
S
/E
ar
th
E
x
p
.
S
at
./
S
p
ac
e
R
es
ea
rc
h
0
.5
0
0
0
.5
0
0
F
S
S
(s
-E
)/
F
S
/B
S
S
/M
S
/M
S
S
0
.5
0
0
0
.5
0
0
F
S
S
(s
-E
)/
F
S
/B
S
S
/M
S
1
.5
0
0
1
.5
0
0
A
v
a
il
a
b
le
B
a
n
d
w
id
th
5
.0
0
0
N
o
t
e
:
S
o
m
e
ba
n
d
w
id
th
s
re
qu
ir
e
co
o
rd
in
a
ti
o
n
w
it
h
o
th
er
se
rv
ic
es
,
es
pe
ci
a
ll
y
F
ix
ed
S
er
vi
ce
s
(F
S
).
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .26
Table 3.4: Frequency Allocations for Satellite Uplink in V band, ITU Region R2 [4].
V band
Uplink
42.50–43.50
[GHz]
47.20–50.20
[GHz]
50.40–51.40
[GHz]
BW
[GHz]
FSS(E-s)/FS/MSS/Radioastronomy 1.000 1.000
FSS(E-s)/FS/MS 3.000 1.000 4.000
Available Bandwidth 5.000
Note: Some bandwidths require coordination with other services, especially Fixed Services (FS).
the downlink and uplink in Ka-band, respectively. Finally, the feeder link employed
only 4.000 GHz for the uplink (V band) and 4.000 GHz for the downlink (Q band),
taking into account that it was an ideal case.
3.4 Satellite Link Design in Ka and Q/V bands
The satellite links can be designed and calculated by using traditional satellite-links
calculations, i.e., the equations for free space loss (FSL), effective isotropic radiated
power (EIRP), Range, CNR, etc., are the same for bands C, Ku, Ka, and Q/V
[1,2, 49, 50].
For this satellite link design, it must be taking into account that the sky is in clear
conditions. This assumption is necessary to find the maximum limit of the satellite
system. However, there are some losses on the path such as attenuation due to
fog, clouds, and gases, but without considering the rain attenuation. Moreover, both
directions forward link and the return link are analyzed separately and independently.
Despite the sky is in clear conditions, there are some losses on the path that affect
the satellite link performance. For this reason, the following ITU-R Recommendations
are included in this study:
• ITU-R P.676: Attenuation by atmospheric gases and related effects [51].
• ITU-R P.834: Effects of tropospheric refraction on radiowave propagation [52].
• ITU-R P.840: Attenuation due to clouds and fog [53].
It is important to mention that the found losses on the Earth-space (E-s) satellite
links range from 0.5–4.0 dB, which do not affect strongly to satellite links. For this
case, satellite links on the E-s path, i.e., satellite uplinks, operate at the V-band.
As a result, the carrier-to-noise ratios (CNR) have reasonable power levels so that it
is possible to estimate the carrier-to-noise and interference ratio (CNIR) using the
carrier-to-interference ratio (CIR), as a value of 20 dB [18].
An initial estimation of capacity and system configuration can only be based on the
bandwidth, spectral efficiency, frequency reuse, and polarization reuse. Therefore, the
total bandwidth available (BWT ) and the total capacity (CpT ) are obtained from [18]
and are expressed as
BWT =
NbNp
Nsb
Savail, (3.1)
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .27
and
CpT = BWTSeff , (3.2)
where Nb is the number of the beams, Np is the potential for polarization reuse, Nsb
is the number of the frequency sub-bands, Savail is the total available spectrum in the
chosen direction, and Seff is the spectral efficiency that is obtained from DVB-S2X.
Table 3.5: Frequency Reuse and Polarization Schemes [18].
Conf. Nsb Np Comments Scheme
1F2P 1 2
Each spot beam is
dual-polarization but it
has no frequency reuse.
3F1P 3 1
Beams are
single-polarization and
3 frequency sub-bands
(3-color scheme).
3F2P 3 2
Beams are
dual-polarization and
3 frequency sub-bands
(3-color scheme).
4F1P 4 1
Beams are
single-polarization and
4 frequency sub-bands
(4-color scheme).
4F2P 4 2
Beams are
dual-polarization and
4 frequency sub-bands
(4-color scheme).
Table 3.5 indicates the schemes that use different frequency sub-bands or/and
polarizations for multiple spot beams (adjacent beams) on the HTS system. The use
of these schemes is popular thereby reducing the interference levels by appropriate
filtering and permitting reuse of a sub-band in a non-adjacent beam. The degree of
frequency reuse is often described numerically by using a frequency reuse factor (FRF)
[18]. It is conventional in frequency reuse to divide the frequency band into several
sub-bands, which use different frequencies, usually 3, 4, 7, etc. These frequency
sub-bands can be referred to as colors for simplicity, therefore, in this study, the
system can have either three or four colors. Moreover, the polarization discrimination
between the beams can be used in various patterns. In [18], the author has developed
the nomenclature mFnP, who has described the use of m colors and n polarizations,
of course, this nomenclature is employed throughout this thesis.
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .28
Furthermore, the antenna circular polarization is employed in this study, as
a result, the module of the field is constant. It is important to note that
circular polarizations are made up of two directions of the rotation field, left
(counter-clockwise) and right (clockwise), also known as left hand circularly polarized
(LHCP) and right hand circularly polarized (RHCP) respectively.
3.5 Radio Interfaces
In order to achieve high capacities for HTS systems, it is essential to have advanced
radio interfaces that permit high performance in satellite links. Thereby, the
DVB-S2X standard is used on the forward link [6], whereas the DVB-RCS2 standard
is used on the return link [7].
The DVB-S2X standard has 39 modulation and coding schemes (MODCOD),
whose modulations varying from QPSK (M-ary Phase-Shift Keying, M-PSK) to
256-APSK (M-ary Amplitud and Phase-Shift Keying, M-APSK) with different coding
rates. Also, the standard uses Low-Density Parity-Check (LDPC) which is a linear
error-correcting code for transmitting a message over a noisy transmission channel
[54,55]. The roll-off can be either 5% or 10%, therefore, it is possible to have excellent
spectral efficiency, that is, DVB-S2X has a 51.44% more spectral efficiency than the
DVB-S2 standard [56].
On the other hand, the DVB-RCS2 standard uses M-ary Quadrature Amplitude
Modulation (M-QAM) and M-PSK schemes, i.e., varying from QPSK to 16-QAM with
two bursts, short and long. Moreover, the roll-off takes on 4 different values, which
are 20%, 25%, 30%, and 35%. It is important to mention that each MODCOD has
a particular coding rate and its error-correcting code is based on turbo-codes [57].
Table 3.6 indicates and compares the maximum spectral efficiencies for each DVB
standard where both DVB-S2X and DVB-RCS2 MODCODs are used on the forward
link and return link respectively.
Table 3.6: Maximum Spectral Efficiencies
Standard MODCOD Roll-off
Spectral
Efficiency,
[b/s/Hz]
DVB-S2X
vs.
DVB-S2
DVB-S2 32 APSK 9/10 20% 3.7109 –
DVB-S2X 256 APSK 3/4 5% 5.6199 +51.44%
DVB-RCS2 16 QAM 5/6 20% 2.1417 –
3.6 Methods to Calculate the Performance of
Satellite Antennas
The trend is to have satellites of large dimensions, especially in communication
capacity as well as size and weight. Indeed, in this thesis were implemented two
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .29
methods to calculate the performance of the satellite antenna, aiming to reduce the
antenna size, therefore, the antenna is also reduced in weight and assured the desired
cover. It is important to mention that if the beamwidth is larger, the power amplifier
needs more power and a lower diameter of the antenna. However, the proposal is
to use MBAs so that it is possible to reduce the beamwidth and to cover the same
region using multiple spot beams. Therefore, there is lower consumption of energy on
the satellite’s power subsystem but higher coordination in the onboard processing for
managing the satellite spot beams, in addition to higher size antennas. Consequently,
the two evaluation methods for antenna performance are presented in the following
subsections, where the first method is focused on the antenna size evaluation, whereas
the second method is approached in the power amplifier performance in Ka-band.
3.6.1 First Method: Antenna Size Evaluation
In this process, the satellite-antenna size (Dant) was numerically calculated by a loop
process. The Q band is used in the space-Earth satellite-communication link (s-E),
which is denoted by Freqs E. The s-E satellite downlink is established between the
HTS and the GW (return link).
To begin with the method execution, the Dant was either increased or decreased
in steps of 1 mm until to find the desired beamwidth (✓BW ). Here, it was suggested
that the ✓BW of the satellite antenna might be defined by a value of 0.20 in direction
to GW [18]. As a result, the best diameter of the satellite antenna was fitted in 2.63
m by using the illumination law [19], therefore, it was assured a narrow beamwidth
with a coverage radius (rc) about 62.5 km. It is possible to compute rc by using ✓BW
and the s-E range (R). Finally, Method 1 is detailed in the pseudocode of the process
to size the satellite antenna.
This method can be also applied to size the satellite antenna diameter in Ka-band,
Dant, which is a function of antenna beamwidth (✓BW ). In this case, the ✓BW must
be less than 0.5 , therefore, the Dant fluctuates in steps of 1 mm, which is very
similar to the calculation process of antenna size in the Q band. In particular, the
coverage radius (rc) of the spot beams can be modified from 80–125 km by using
different beamwidths that range from 0.26 –0.40 according to the ground-segment
requirements.
3.6.2 Second Method: Calculations for Antenna Gain and
EIRP
Here, it is possible to find the transmitter-antenna gain (GTX) and the effective
isotropic radiated power (EIRPsat) for each spot beam in Ka-band. For this purpose,
High Power Amplifier (HPA) was either increased or decreased by steps of 100 mW
until reaching the maximum EIRP as a system threshold, whereas the second option
was to modify the Dant by steps of 1 mm until reaching the maximum EIRP allowed.
As a result, the diameter of the satellite antenna (Dant) was ranged from 2.62–4.10
m, and the HPA was set in 53.0 W, therefore, it was possible to calculate in different
coverage cases and to choose the best option for the proposed system. Method 2 is
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .30
Method 1 Calculation of the Satellite Antenna Diameter
Input: Dant, R, Freqs E
Output: Dant, rc
1: ✓BW compute(Freqs E, Dant);
{For Ka-band: 0.26 –0.40 }
2: if (✓BW 0.20 ) then
3: while (✓BW 0.20 ) do
4: Dant Dant 0.001;
5: ✓BW compute(Freqs E, Dant);
6: end while
7: else
8: while (✓BW 0.20 ) do
9: Dant Dant + 0.001;
10: ✓BW compute(Freqs E, Dant);
11: end while
12: end if
13: rc compute(R, ✓BW );
14: return [Dant, rc]
detailed in the pseudocode of the process to calculate the satellite antenna size as
a function of HPA, EIRP, Freqs E, and ✓BW . Moreover, if the EIRP levels exceed
the permitted limit in the Q band, this method can be also implemented for satellite
antennas in that frequency band.
3.7 Analysis and Quantification of Multiple Spot
Beams for HTS Systems in Ka and Q/V bands:
Numerical Results and Discussion
Primarily, the analysis to quantify the number of spot beams was based on the
bandwidth, spectral efficiency, frequency reuse, polarization, EIRP levels, and data
traffic rates. Generally, this first analysis provided a theoretical quantification for the
HTS system, however, it was considered neither payload nor power subsystem of the
HTS system, since it was not the purpose of this thesis.
As stated above, the scheme N + P was implemented on the Ground Segment,
also, the mixed solution was proposed where both Ka and Q/V bands were assigned
to user and feeder links, respectively. It is important to mention that 1 Tb/s is the
maximum capacity for this numerical analysis. Furthermore, the quantification of
multiple spot beams was also evaluated by data-traffic-ratio scenarios at the feeder
link. The traffic scenarios can be denoted by [DL:UL], where DL is the download
traffic and UL is the upload traffic.
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .31
Method 2 Calculations for Antenna Gain and EIRP
Input: Dant, R, option, Freqs E, EIRPmax, HPAsat
Output: Dant, rc, HPAsat
1: GTX compute(Dant, F reqs E);
2: EIRPsat compute(GTX , HPAsat);
3: switch (option)
4: case 1:
5: while (EIRPsat < EIRPmax) do
6: HPAsat HPAsat + 0.1;
7: EIRPsat compute(GTX , HPAsat);
8: end while
9: case 2:
10: while (EIRPsat < EIRPmax) do
11: Dant Dant + 0.001;
12: GTX compute(Dant, F reqs E);
13: EIRPsat compute(GTX , HPAsat);
14: ✓BW compute(Freqe T , Dant);
15: end while
16: default:
17: {This option does not exist}
18: end switch
19: rc compute(R, ✓BW );
20: return [Dant, rc, HPAsat]
3.7.1 Orbital Position Analysis
In this thesis, several GWs were distributed in the Latin America region, that is,
Mexico, Central America, The Caribbean, and South America. For this reason, the
orbital position for the HTS system was crucial for this evaluation. The list of possible
geographic coordinates for GW locations is provided in Appendix A.1.
First, the energy per symbol to noise power spectral density (Es/N0) must be
calculated for each GW feeder uplink by ranging the orbital position from 0 –180
(west longitude). To find the best orbital position, the Monte Carlo Method is
employed in this assessment [58], where a domain of possible inputs (window of orbital
positions) is used to perform a deterministic computation of the inputs. Thus, each
Es/N0 is computed regarding the range of orbital positions. As a result, the best
orbital position can be interpreted according to the best power levels, which are
represented by a set-theoretic intersection. Table 3.7 shows the parameters used for
the feeder uplink calculations which were obtained from [8, 13, 14, 18] in addition to
the previously mentioned Methods.
In order to obtain the best spectral efficiency for each feeder uplink, the Es/N0
levels must be above 20 dB due to the fact that the best MODCOD (256 APSK 3/4)
of the DVB-S2X standard requires 19.57 dB of Es/N0. Furthermore, the found Es/N0
have a positive link margin of 2.5 dB above the reference of 13.5 dB [14]. Figure 3.3
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .32
Table 3.7: Feeder Uplink Parameters
HTS System User Terminal (UT) Gateway Station (GW)
Parameter Value Parameter Value Parameter Value
Dant (Q band) 2.63 m Dant (Ka-band) 0.75 m Dant (V band) 5 m
Dant (Ka-band) 4.10 m LNA (Ka-band) 8.8 W HPA (V band) 16 W
HPA (Q band) 15 W G/T 20 dB/K
HPA (Ka-band) 53 W
depicts the Es/N0 20 dB of each feeder uplink by using the Monte Carlo Method.
Therefore, several valid orbital-positions were found from -102.0 to -50.0 , but only
one orbital-position encompassed the best Es/N0 levels of each feeder uplink.
-180 -160 -140 -120 -100 -80 -60 -40 -20
20
22
24
26
28
30
32
Figure 3.3: Es/N0 for each feeder uplink using the Monte Carlo Method.
For this purpose, the Central Limit Theorem was applied to the orbital-position
observations, which were obtained from the Monte Carlo Method. It is important to
mention that each observation does not depend on the values of the other observations,
therefore, the central limit theorem says that the distribution of the average is closely
approximated by a normal distribution. Figure 3.4 depicts the normal distribution
by using the central limit theorem, providing the best orbital position that is -74.1 .
Even though -74.1 is the best orbital position in this evaluation, the range from
-80.0 to -70.0 is also a good option for orbital positions due to the orbital positions
are true 68% of the time within the 1- events. Finally, It is worth noting that the
orbital positions must be negotiated with regulatory entities, however, these orbital
positions are studied for theoretical purposes. All these calculations are made using
computer codes, i.e., Matlab and Python scripts.
3.7.2 Quantification of Multiple Spot Beams
In order to find the number of spot beams, Eq. 3.1 and Eq. 3.2 were used by
taking as reference the total capacity of 1 Tb/s. Thus, the traffic behavior was
considered in this analysis where 5 data-traffic-ratio scenarios were proposed and
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .33
Figure 3.4: Normal distribution of the orbital-position observations.
studied. Initially, symmetric traffic between forward and return links was assumed
as [1 : 1] data traffic ratio, meaning 50% (500 Gb/s) of the data throughput in the
outbound and 50% (500 Gb/s) in the inbound. However, a more realistic traffic
scenario could be 60%:40% which can be expressed as [3 : 2] data traffic ratio. The
other three traffic scenarios employed in this thesis are: [4 : 1] (80%:20%), [7 : 3]
(70%:30%), and [9 : 1] (90%:10%).
Figure 3.5 shows the results of the number of beams using DVB-S2X and DVB-S2
standards for the feeder uplink. For this purpose, the 1F2P scheme was implemented
for both trials, in addition to the satellite links were computed under ideal conditions.
Due to the best spectral efficiency, the best results were found by using the DVB-S2X
standard, where the number of beams for the feeder uplink is always less than
the number of beams using the DVB-S2 standard, for all data traffic scenarios.
That is, each feeder uplink beam corresponds to a gateway station (GW) on the
ground segment (GS). For this reason, a fewer number of beams are necessary for
infrastructure savings on terrestrial segment deployments. Throughout this study,
the DVB-S2X standard is used only for assessing and calculating satellite forward
links.
Figure 3.6 illustrates the results of the number of carriers in the feeder downlink
(return link). Each carrier has a bandwidth of 10 MHz [8], hence, whether the return
data traffic is greater, there are more carriers on the feeder downlink. Particularly
with 50% of data traffic, i.e. [1 : 1] traffic ratio, there are 11664 carriers. On the
other hand, with 10% of data traffic, which is [9 : 1] traffic ratio, there are only
2331 carriers. Therefore, the choice of data-traffic-ratio can be either a positive or a
negative factor in order to synchronize all data carriers in the satellite return links
towards GWs.
Figure 3.7 depicts the behavior of total capacity vs. data traffic ratio, where the
upper limit of capacity is always 1 Tb/s. Nevertheless, there is a slight increase in
total capacity due to the greater number of beams, generally, one more beam in the
set of feeder uplinks. Table 3.8 shows the increase in the throughput of feeder uplinks
whereas the set of feeder downlinks is easier to fit because each carrier has only a
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .34
N
u
m
b
er
o
f
B
ea
m
s
0
7
14
21
28
35
Data Traffic Ratio
[1:1] [3:2] [7:3] [4:1] [9:1]
31
27
24
21
17
21
18
16
14
12
Feeder Uplink, DVB-S2X [1F2P]
Feeder Uplink, DVB-S2 [1F2P]
Figure 3.5: Comparison of the obtained number of beams using DVB-S2X and DVB-S2
standards.
N
u
m
b
er
o
f
C
ar
ri
er
s
0
1250
2500
3750
5000
6250
7500
8750
10000
11250
12500
Data Traffic Ratio
[1:1] [3:2] [7:3] [4:1] [9:1]
2331
4662
6992
9338
11664
Feeder Downlink, DVB-RCS2 [1F2P]
Figure 3.6: The number of carriers in the feeder downlink.
bandwidth of 10 MHz. That is, each feeder-uplink beam has a theoretical capacity of
44.96 Gb/s and each feeder-downlink carrier has a capacity of 42.83 Mb/s.
Figure 3.8 shows the number of beams and carriers in the user link using the
3-color frequency-reuse scheme. Figure 3.8a displays the number of beams using both
3F1P and 3F2P schemes where it is possible to visualize the difference between these
two configurations. The number of beams for the 3F1P scheme is approximately
twice regarding the 3F2P scheme, for all traffic-ratio cases. However, the theoretical
throughput of each user downlink for the 3F2P scheme is twice regarding the 3F1P
scheme, that is, 11.24 Gb/s versus 5.62 Gb/s respectively.
On the other hand, Figure 3.8b illustrates the number of carriers for both
3-color frequency-reuse schemes. The number of carriers for the 3F1P scheme is
approximately two times greater than the 3F2P scheme, for all traffic-ratio cases.
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .35
T
o
ta
l
C
ap
ac
it
y
[
T
b
/s
]
0.000
0.100
0.200
0.300
0.400
0.500
0.600
0.700
0.800
0.900
1.000
1.100
1.200
1.300
Data Traffic Ratio
[1:1] [3:2] [7:3] [4:1] [9:1]
Feeder Downlink Traffic, DVB-RCS2
Feeder Uplink Traffic, DVB-S2X
Figure 3.7: Total capacity comparison between different data traffic ratios.
Table 3.8: Throughput Comparisons
Traffic
Ratio
Feeder
Uplink
Throughput
[Gb/s]
Feeder
Downlink
Throughput
[Gb/s]
Total
Increase
[1 : 1] 539.51 499.62 3.9%
[3 : 2] 629.43 399.98 2.9%
[7 : 3] 719.35 299.50 1.9%
[4 : 1] 809.27 199.69 0.9%
[9 : 1] 944.14 99.85 4.4%
Conversely, the theoretical throughput of each user uplink for the 3F2P scheme is
two times greater than the 3F1P scheme, that is, 14.28 Mb/s versus 7.14 Mb/s
respectively. It is important to mention that it is tough to manage and to synchronize
several user-uplink carriers. For this reason, the sophisticated multiple-frequency
time-division multiple access (MF-TDMA) is a great choice in order to manage
adequately user traffic, in addition to increasing the number of users on the GS [5].
The analysis for 4-color schemes is similar to 3-color schemes, thus, Figure 3.9
shows the number of beams and carriers using both 4F1P and 4F2P schemes. The
common factor between the number of beams for both 4-color schemes was about 2.0
due to the double polarization antenna, reducing the number of beams by a factor of
2.0 but increasing the throughput in the same proportion. Figure 3.9a describes the
number of beams in both 4-color schemes where the throughput of the 4F1P scheme
is about 4.21 Gb/s and 8.42 Gb/s for the 4F2P scheme in all data traffic cases.
Figure 3.9b illustrates the number of carriers for the user uplink in all data traffic
cases, where 4F1P schemes have a number of carriers greater than 4F2P schemes by a
factor about 2.0. Finally, the throughput of carrier uplinks for 4F2P schemes is about
10.70 Mb/s whereas the throughput of carrier uplinks for 4F1P schemes is about 5.35
Gb/s.
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .36
N
u
m
b
er
o
f
B
ea
m
s
0
19
38
57
76
95
114
133
152
171
190
Data Traffic Ratio
[1:1] [3:2] [7:3] [4:1] [9:1]
81
72
63
54
45
161
143
125
107
89
User Downlink, DVB-S2X [3F1P]
User Downlink, DVB-S2X [3F2P]
(a) The number of beams in the user downlink for the 3-color schemes.
N
u
m
b
er
o
f
C
ar
ri
er
s
0
7500
15000
22500
30000
37500
45000
52500
60000
67500
75000
Data Traffic Ratio
[1:1] [3:2] [7:3] [4:1] [9:1]
6,966
13,968
20,979
27,972
35,010
14,007
27,885
42,000
55,961
69,954
User Uplink, DVB-RCS2 [3F1P]
User Uplink, DVB-RCS2 [3F2P]
(b) The number of carriers in the user uplink for the 3-color schemes.
Figure 3.8: The number of beams and carriers using both 3F1P and 3F2P schemes.
To sum up, the sizing of the number of beams depends on the coverage area and
the data traffic ratio, that is, to find an adequate beamwidth size in order to fit in
the desired coverage area. Moreover, it is important to understand the data traffic
behavior for each user so that it can choose properly the data traffic scenario. This
last parameter is essential to design the throughput in both beams and carriers for
the HTS system and to be able to size the right capacity for the end-user. At last
but not least, the rain attenuation was not considered in this initial study, however, a
complete analysis of rain attenuation and interferences sources are taken into account
in the following Chapters.
CHAPTER 3. QUANTIFICATION OF MULTIPLE SPOT BEAMS FOR HTS . . .37
N
u
m
b
er
o
f
B
ea
m
s
0
50
100
150
200
250
Data Traffic Ratio
[1:1] [3:2] [7:3] [4:1] [9:1]
107
95
84
72
60
214
190
167
143
119
User Downlink, DVB-S2X [4F1P]
User Downlink, DVB-S2X [4F2P]
(a) The number of beams in the user downlink for the 4-color schemes.
N
u
m
b
er
o
f
C
ar
ri
er
s
0
20000
40000
60000
80000
100000
Data Traffic Ratio
[1:1] [3:2] [7:3] [4:1] [9:1]
9,309
18,620
27,972
37,296
46,680
18618
37240
55945
74646
93296
User Uplink, DVB-RCS2 [4F1P]
User Uplink, DVB-RCS2 [4F2P]
(b) The number of carriers in the user uplink for the 4-color schemes.
Figure 3.9: The number of beams and carriers using both 4F1P and 4F2P schemes.
3.8 Contributions of the Research
This chapter offers an initial study about the quantification of multiple beams and
carriers for the HTS systems, therefore, the results have been published in IEEE Latin
America Transactions journal and titled:
• Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez, and Jose M.
Matias. Analysis to Quantify and Optimize Spot Beams for a High Throughput
Satellite in Ka and Q/V Bands. IEEE Lat. Am. Trans., 17(02):219–227, feb
2019
Chapter 4
Interference Evaluations in Frequency
Reuse by Using Offset-
Parabolic-Reflector Antennas
4.1 Introduction
This chapter analyzes the features and parameters of an offset-fed parabolic reflector
antenna for the HTS system. Indeed, some antenna parameters are important
for satellite communication links, so that it is necessary to evaluate the effect
of frequency-reuse configurations and spot-beam interferences by designing an
offset-parabolic-reflector antenna for this satellite system. That is, the interferences
such as cross-polarization (cross-polar) and co-polarization (co-polar) are vital for
the system performance and its feasibility. To be precise, the main aim is to adjust
and to reduce the interference levels to ensure satellite links for the HTS system,
obtaining the maximum advantage for frequency reuse configurations, the number of
beams, and the circular polarization. The evaluations of both Multiple Spot Beams
and the Carrier-to-Interference employ well-known methods [9,18,59], but with slight
modifications to this proposal. Finally, the designed antenna is theoretically and
numerically assessed by both implemented methods. Consequently, it is possible to
find a satellite reflector antenna suitable for the proposed HTS system.
4.2 Geometric Design of an Offset-Parabolic-Reflector
Antenna
Usually, the offset-parabolic-reflector antennas are a great option to reduce sidelobe
levels. Motivated by this, moving the feed out of the aperture, some problems with
axisymmetrical reflectors can reduce. Hence, diffraction-caused sidelobes, blockage
losses, and cross-polarization can reduce or even disappear. For this reason, a
proper antenna’s design for the proposed HTS system is necessary to accomplish
the above-mentioned requirements. Thereby, some parameters were obtained from
38
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .39
geometry and illumination laws by procedures from [19, 60]. In this section, the
antenna geometry and its parameters are obtained by the following method.
To begin with, the aperture plane diameter, also known as antenna diameter
(Dant), depends on a coefficient obtained from the illumination law, whereby a
coefficient with a typical value of 70 introduces some tapering at the edge of the
reflector [19]. Consequently, the aperture plane diameter can be expressed by
Dant = 70
✓3dB
, (4.1)
where is the wavelength of the frequency band used in a specific satellite link,
and the ✓3dB is the angle subtended by the half-power points of the main lobe [19].
This last parameter was used to characterize the beamwidth ✓BW (see Section 3.6),
therefore, it was possible to assume that ✓BW = ✓3dB.
Figure 4.1 shows the geometric parameters that are part of the
parabolic-reflector-antenna [60, 61]. As the reflector antenna indicates, there is
an offset distance that provides a blockage-free region for structures in the focal
region [60]. The offset (h) is the distance from the axis of symmetry to the lower
reflector edge. Generally, the offset distance is popular in VSAT applications. In
this case, this h distance can be obtained from [60] and denoted by
h =
Dant
8
, h > 0 (4.2)
The distance from the symmetry axis to the center of the reflector is known as
the offset of the reflector center (H) which is described as
H =
Dant
2
+ h (4.3)
In this procedure, one choice of the antenna design is to keep the ratio of focal
length (F ) to the antenna diameter (Dant) constant, i.e., the ratio is constant at
1 [62]. Therefore, F/Dant = 1. It is worth to mention that the F/Dant choice also
impacts on cross-polarization performance [63].
The main contribution to this study was to modify Dant regarding different
beamwidths (✓3dB) for both Ka and Q/V bands. Indeed, this is possible as the antenna
diameter is a flexible parameter, in addition to exploring different cover areas with
multiple spot beams. According to [61], the half of the angle subtended ( s) by the
reflector as viewed from the focal point can be expressed as
s = tan 1
✓
8FDant
16F 2 + 4H2 D2
ant
◆
(4.4)
Meanwhile, the angle which bisects the reflector subtended angle is defined by
B = tan 1
✓
16FH
16F 2 +D2
ant 4H2
◆
(4.5)
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .40
xx
s
z y
UU
L
C
FA
B
P
C
L
xf
zf
Figure 4.1: The geometry for the offset-parabolic-reflector antenna.
The lower angle ( L) can be described as
L = B s (4.6)
Thus, the upper angle ( U) is determined by the sum of the subtended angle and
the lower angle, which is expressed by
U = 2 s + L (4.7)
It is important to note that the feed is directed at an angle ( f ), i.e., the angle of
the feed-antenna-pattern peak relative to reflector axis of symmetry, where the feed
is directed toward the point P . The f is denoted by
f = 2 tan 1
✓
H
2F
◆
(4.8)
Furthermore, the angle C is similar to the angle f when the feed aims at the
reflector point C, corresponding to the aperture center.
The angle from the lower edge of the dish to feed pointing direction is defined by
P = f L (4.9)
For a parabolic reflector, the spherical wave nature of the feed radiation is referred
to as spherical spreading loss (SSL), for both upper angle ( U) and lower angle ( L)
edges [63], which is given by
SSL( ) = 20 log
cos2
2
(4.10)
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .41
The feed-pointing angle gives a difference in edge illuminations (EI) [60], which is
expressed by
∆EI = EIU EIL (4.11)
The negative of the edge illumination (EI) is the sum of the feed taper (FT) and
the spherical-spreading loss (SSL) [60] so that it can be denoted by
FTL + SSLL = FTU + SSLU +∆EI (4.12)
Finally, substituting Eq. (4.10) into Eq. (4.12) outputs the following design
equation,
∆FT = FTL FTU = 40 log
8
>
>
>
>
<
>
>
>
>
:
cos
L
2
cos
U
2
9
>
>
>
>
=
>
>
>
>
;
+∆EI (4.13)
Figure 4.1 illustrates the points of FTU , FTL, EIU , and EIL. For the case of balanced
aperture illumination, then ∆EI = 0. To sum up, each parameter was rigorously
evaluated and computed in order to obtain the best performance of the antenna
in terms of reduced sidelobes with an acceptable antenna gain according to the
focal-length-to- diameter ratio, F/Dant [63]. Therefore, this procedure is able to
find a suitable offset-parabolic-reflector antenna for the HTS system.
4.3 Multiple Spot Beams Antenna for HTS Systems
The multiple-aperture-antenna with a single element per beam was chosen to study
in this thesis. This antenna is either 3 or 4 apertures, that is, it depends on
the frequency-reuse schemes either 3 or 4 colors, respectively. Hereafter, this
multiple-aperture-antenna is also known as the multiple beam antenna (MBA) or
Type (c) MBA [9]. The MBAs have carried out important researches in this area
[9, 13,59,62], which have provided some advantages such as:
• The effective spectral bandwidth increases in several folds due to the
frequency-reuse channels over numerous spot beams,
• The beam has a smaller size so that the antenna has higher gain, resulting
in enhanced effective isotropic radiated power (EIRP) for the downlink and
improving the antenna gain-to-noise temperature (G/T ) for the uplink,
• The MBAs allows the use of much smaller ground terminals such as GWs and
VSATs.
• The satellite reflector could be lower than 5.0 m diameter, which could lead to
the best accommodation of multiple reflectors in the spacecraft. Naturally, the
antenna diameter depends on its gain and the employed frequency.
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .42
θ
θ
θ
θ
θ
θ
θ
θ
(a) 3-cell frequency-reuse
θ
θ
θ
θ
θ
θ
θ
θ
(b) 4-cell frequency-reuse
Figure 4.2: The hexagonal-grid layouts of both 3-cell and 4-cell for frequency reuse,
illustrating the beam parameters.
4.3.1 Multiple Spot Beams: Design and Analysis
Adjacent beams are generated from different apertures, which are formed in an
interleaved multiple-spot-beam coverage on the GS. Therefore, the design of a
multiple-spot-beam antenna is depended on the beam size, which is related to the
minimum-coverage-area directivity requirement. Figure 4.2 depicts hexagonal grid
layouts for both 3-reflector and 4-reflector antennas, where the minimum directivity
occurs at the triple beam crossover of the three adjacent beams for MBAs with
uniform-size beams.
Furthermore, typical beam overlaps used for MBA designs are approximately
3 dB for two adjacent beams, and 4 dB for three adjacent beams [64]. The
optimum overlap level depends on the minimum-coverage directivity, which requires
the co-polar isolation among reuse beams and the frequency-reuse scheme (3-color
scheme, 4-color scheme, etc.). Consequently, the spacing between adjacent-beam
centers (✓s) determines the number of beams for the desired coverage. Figure 4.2a
and Figure 4.2b illustrate the hexagonal-grid layout of the beams for 3-reflector and
4-reflector schemes, respectively. As a result, the ✓s can be expressed by
✓s =
p
3
2
✓0 =
p
3rc, (4.14)
where ✓0 is the beam diameter at the triple beam crossover and rc is the coverage
radius of the spot beam.
Thus, the minimum number of beams (Nbmin
), for the desired coverage area is
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .43
given approximately by
Nbmin
⇡ Ac
Ah
, (4.15)
where Ac is the coverage area and the Ah represents the area of the hexagonal cell
associated with each spot beam. The area of the hexagon can be calculated by
Ah = 3
p
3
2
r2c . It is important to note that the actual number of beams (Nb) is usually
20% larger than Nbmin
for an efficient layout of the beams over the coverage area [59].
Moreover, the triple beam crossover levels for outer beams occur at the edge of the
beam coverage.
Further, the closest spacing between beam centers reusing the same frequency (✓c)
for 3-reflector (using the 3-color frequency-reuse scheme) and 4-reflector (using the
4-color frequency-reuse scheme) antenna systems can be respectively denoted by
✓c,3 =
p
3✓s = 3rc, (4.16)
and,
✓c,4 = 2✓s = 2
p
3rc (4.17)
Finally, the closest spacing between the reuse-beam edges (✓r), determines the
achievable carrier-to-interference ratio (CIR), which is expressed by ✓r = ✓c ✓0.
Therefore, the ✓r for both 3-reflector and 4-reflector hexagonal-grid layouts can be
respectively expressed by
✓r,3 =
✓s
2
= rc, (4.18)
and,
✓r,4 = ✓s(
p
3 1) = 2(
p
3 1)rc (4.19)
4.3.2 Spot Beam Pattern: Model Analysis
The analysis of this model has been very similar to the previous research [18], but
it has been more practical and simpler than from the aforecited method. Thus,
the simplified analysis has been performed and adapted to this study successfully.
As was previously mentioned in Section 4.2, the offset-parabolic-reflector antenna
was a good option in order to reduce the sidelobe levels. For this purpose, it was
necessary to taper the field distribution over the circular aperture. Thereby, an
adequate alternative was to use a parabolic taper on an offset-distance value at the
edge. In summary, the antenna performance can be determined analytically for integer
taper roll-off values.
For starting, the antenna beam is based on the Gaussian beam pattern obtained
from [18], where the antenna pattern function is given by
f(✓BW , n, h) =
hf(✓BW , n = 0) +
1 h
n+ 1
f(✓BW , n)
h+
1 h
n+ 1
(4.20)
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .44
so that,
f(✓BW , n) = 2n+1(n+ 1)!
Jn+1(U)
Un+1
(4.21)
where, for simplicity,
U =
2⇡
Dant sin ✓BW , (4.22)
where h is the offset distance of the parabolic antenna, ✓BW is the beamwidth, n is
an integer number ranging from 0–2, which represents the field taper roll-off factor.
Moreover, Jn+1 is the Bessel function of the n + 1 kind, is the wavelength of the
frequency band used, and Dant the diameter of the antenna aperture.
Applied to this study, the relative gain of the antenna obtained from [18], Gr(✓),
can be found by
Gr(✓BW ) = 20 log10(|f(✓BW ), n, h|) (4.23)
To sum up, the relative gain of the offset-parabolic-reflector antenna was an
important parameter in order to evaluate the CIR of the downlinks for both Q band
and Ka-band respectively.
4.4 Carrier-to-Interference: Evaluation Model
In this Section, the methodology to evaluate the carrier-to-interference ratio (CIR)
was only determined for both feeder uplinks and downlinks, and solely user downlinks.
Thus, it is important to mention that the CIR, also known as ⇣, is independent of the
propagation impairments, e.g., rain attenuation, as the useful and interfering signal
follows the same path from the HTS to the considered ground point [18]. To begin
with, a beam grid was defined by several points, where each of these points was
scanned in order to compute the ⇣. This methodology has been obtained from [18],
whose performances and results for the ⇣ evaluation have been successfully tested.
Some assumptions were proposed regarding the downlink interferences,
• Each beam transmitted identical signals
• At the center of each beam, the forward uplink was the same
• In the spectrum of interest, there was a unique carrier in each beam
• Each GW was located at the beam center, therefore, its signal was emitted from
there.
With these considerations, the ⇣ can be computed by
⇣ =
Gr(!)
PLT
n
P
i=1
Ii( )
PLi
(4.24)
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .45
where Gr(!) is the antenna relative gain of the wanted beam at the point of interest
by using the Eq. (4.23), ! is the angle that ranges from ! = 0 (beam center) to
! = ✓BW/2 (beam edge), PLT is the path loss from HTS to the user terminal (UT)
at the point T , Ii( ) is a function that describes the nature of the interference, is
the angle between the point T and the center of the i-th beam, and PLi is the path
loss from the HTS to the center of the i-th beam [18,50].
The evaluation method was divided into 4 cases. Firstly, the method started when
the i-th beam evaluated was the same color, (frequency), and the same polarization,
then Ii would become the relative gain of the antenna at the point of interest.
Secondly, when the i-th beam was of the same frequency but different polarization,
then Ii would match to the beam relative gain of the cross-polar antenna at the point
of interest. For the third case, if the i-th beam was in a different frequency but
the same polarization, then Ii would be similar to the antenna relative gain of the
beam at the point of interest but reduced by the Adjacent Carrier Isolation (ACI)
factor. Lastly, if the i-th beam was neither in the same frequency-reuse nor the same
polarization, then Ii would become the beam relative gain of the cross-polar antenna
at the point of interest but reduced by the Cross-Polar Isolation (XPiso) factor.
Figure 4.3 illustrates the complete interference geometry of the downlink used for
the evaluation method. However, for the feeder downlink, the evaluation method
only had 2 cases because the feeder link not only did it uses double polarization but
it did not use frequency reuse. Hence, when the i-th beam evaluated was the same
frequency and the same polarization, then Ii would be the relative gain of the antenna
at the point of interest. Meanwhile, when the i-th beam was the same frequency but
different polarization, then Ii would equal to the beam relative gain of the cross-polar
antenna at the point of interest.
The CIR evaluation mechanism used in this thesis was very similar to the method
employed in cellular networks. For instance, the ⇣ of the cellular networks ranges
from 13–15 dB [65]. By contrast, for satellite systems in Ka-band, the ⇣ ranging from
14.5–17.6 dB [66]. Meanwhile, the ⇣ for the downlink in the Q band must be above
20 dB, keeping an average of 29 dB [18]. This parameter was necessary to define in
order to know whether the GW beams had good isolation between each other. For
this reason, it was necessary to implement this method to find the minimum CIRs
for the proposed HTS system.
4.5 Offset-Parabolic-Reflector Antennas: Numerical
Results
4.5.1 Offset-Parabolic-Reflector Antenna: Geometric
Parameters
As was analyzed in Section 4.2, the Eqs. (4.1)–(4.9) were used to find each geometric
parameter of the offset-parabolic-reflector antenna. For the downlink in Ka-band, the
geometric parameters were calculated by 4 different beamwidths in order to explore
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .46
T
HTS
-th beami
Figure 4.3: The interference geometry of the downlink.
diverse antenna diameters, i.e., ✓BW = 0.26 , ✓BW = 0.32 , ✓BW = 0.40 , and ✓BW =
0.48 . Meanwhile, the geometric parameters in the Q band were calculated by only
one beamwidth, ✓BW = 0.20 . Table 4.1 shows the obtained geometric parameters of
the antenna by using the previous equations.
Furthermore, the antenna feed was modeled by a symmetric Gaussian radiation
pattern, which was given by a value of 10.0 dB at the edges [60, 63]. Hence, 10.0 dB
were defined for both lower (FTL) and upper (FTU) edge tapers. Table 4.2 indicates
the values of spherical spread losses and edge illuminations obtained by using Eqs.
(4.10)–(4.13).
It is important to note that the EI values are not recommended because the
difference between EIU and EIL must be zero. Thereby, it was necessary to point the
antenna feed with an angle f in order to balance the aperture illumination. Table
4.3 indicates the balanced parameters of the offset-parabolic-reflector antenna when
the feed is aiming to the point P with an angle f . The angle f was modified
by little steps until a high performance of cross-polar isolation(XPiso) was achieved.
With these results obtained, offset-parabolic-reflector antennas were simulated by a
computer program.
Thus, the computer program TICRA-GRASP was used to simulate the
offset-parabolic-reflector antenna [67], which, among other things, was configured
to operate in circular polarization, that is, LHCP and RHCP. Figure 4.4 depicts
the radiation-pattern diagram of the offset-parabolic-reflector antenna for the feeder
downlink, operating in the Q band.
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .47
Table 4.1: Geometric Parameters of the Offset-Parabolic-Reflector Antenna
Parameters Q band Ka-band Unit
Freqs E 40.00 20.00 20.00 20.00 20.00 GHz
✓BW 0.20 0.48 0.40 0.32 0.26 deg
Dant 2.63 2.19 2.63 3.28 4.10 m
h 0.33 0.27 0.33 0.41 0.51 m
H 1.64 1.37 1.64 2.05 2.56 m
F/D 1.00 1.00 1.00 1.00 1.00 -
B 32.93 32.93 32.93 32.93 32.93 deg
s 25.78 25.78 25.78 25.78 25.78 deg
L 7.15 7.15 7.15 7.15 7.15 deg
U 58.72 58.72 58.72 58.72 58.72 deg
f 34.71 34.71 34.71 34.71 34.71 deg
P 27.56 27.56 27.56 27.56 27.56 deg
Table 4.2: Spherical Spread Losses and Edges illuminations with a Symmetric
Gaussian Radiation Pattern
Parameters Q band Ka-band Unit
Freqs E 40.00 20.00 20.00 20.00 20.00 GHz
✓BW 0.20 0.48 0.40 0.32 0.26 deg
FTU 10.00 10.00 10.00 10.00 10.00 dB
FTL 10.00 10.00 10.00 10.00 10.00 dB
∆FT 0.00 0.00 0.00 0.00 0.00 dB
SSLU 2.39 2.39 2.39 2.39 2.39 dB
SSLL 0.03 0.03 0.03 0.03 0.03 dB
EIU 12.39 12.39 12.39 12.39 12.39 dB
EIL 10.03 10.03 10.03 10.03 10.03 dB
∆EI 2.35 2.35 2.35 2.35 2.35 dB
Table 4.3: Spherical Spread Losses and Edges illuminations, f = 34.71
Parameters Q band Ka-band Unit
Freqs E 40.00 20.00 20.00 20.00 20.00 GHz
✓BW 0.20 0.48 0.40 0.32 0.26 deg
FTU 10.00 10.00 10.00 10.00 10.00 dB
FTL 12.35 12.35 12.35 12.35 12.35 dB
∆FT 2.35 2.35 2.35 2.35 2.35 dB
SSLU 2.39 2.39 2.39 2.39 2.39 dB
SSLL 0.03 0.03 0.03 0.03 0.03 dB
EIU 12.39 12.39 12.39 12.39 12.39 dB
EIL 12.39 12.39 12.39 12.39 12.39 dB
∆EI 0.00 0.00 0.00 0.00 0.00 dB
Demonstrating the pointing of the antenna feed, this parameter is very sensitive
so that it influences directly to the cross-polarization.
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .48
Figure 4.4: Normalized pattern of the offset-parabolic-reflector antenna, 40 GHz.
Table 4.4 shows the XPiso simulation results for both feeder and user downlinks.
Table 4.4: Cross-polar isolation XPiso for both feeder and user downlinks.
Parameters Q band Ka-band Unit
Freqs E 40.00 20.00 20.00 20.00 20.00 GHz
✓BW 0.20 0.48 0.40 0.32 0.26 deg
XPiso 27.78 27.39 27.33 27.25 27.16 dB
The values of XPiso are very similar to each other because the antenna geometry
is the same for all cases. Experiments have been carried out to find the most
appropriate range of XPiso, which have been ranged from 20–40 dB [1,2,8,17,18,50,61].
Recommending a value of 25 dB, the results are totally convincing and viable for the
final design of offset-parabolic-reflector antennas in this study.
4.5.2 Sizing of Spot Beams over the Coverage Area
Initially, it is worth to notice that the Latin American region is the area of interest,
where it was possible to find the approximate number of spot beams by using the Eq.
(4.15). For this purpose, the Latin American region was divided into three coverage
regions: Mexico, Central America & The Caribbean, and South America. Moreover,
this evaluation was solely performed by both 3 and 4 frequency-reuse schemes, i.e.,
for user downlinks in Ka-band, as it is the service area. Table 4.5 displays the
approximate number of beams for each coverage area regarding the beamwidths.
It is important to note that the number of Nb is approximately 20% larger than the
Nbmin
because it is a more realistic scenario for an efficient layout of the spot beams
over the coverage area [59]. Therefore, this is an excellent approximation for sizing
the spot beams over the coverage area without considering the capacity of the HTS
system.
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .49
Table 4.5: The Number of Spot Beams (Nbmin
/Nb) for the Latin American Regions.
Beamwidths, ✓BW
✓ = 0.48 ✓ = 0.40 ✓ = 0.32 ✓ = 0.26
rc ⇡ 150
km
rc ⇡ 125
km
rc ⇡ 100
km
rc ⇡ 80
km
Region Ac Nbmin
/Nb
Mexico 1.973⇥ 106 km2 34/41 49/59 76/92 119/143
Central America &
The Caribbean
7.328⇥ 105 km2 13/16 19/23 29/35 45/54
South America 1.801⇥ 107 km2 309/371 444/533 694/833 1084/1300
Total Beams 356/428 512/615 799/960 1248/1497
4.5.3 CIR Evaluation Results
Primarily one of the advantages of the Q/V band was able to generate narrower beams
than other frequency bands. Applying the method of Section 4.4, Table 4.6 shows
the CIR results in the Q band for the feeder downlink. Also, it is important to note
that this method evaluates and gives the co-channel interference (⇣co) as well as the
adjacent-channel interference (⇣adj).
Table 4.6: The CIR Evaluation for Feeder Downlink, 40 GHz.
Parameters Results
Denom. Values Denom. Values Denom. Values
mFnP 1F2P Dant 2.63 m ⇣co 36.09 dB
position 92.0 W EI 12.39 dB ⇣adj 38.17 dB
Freqs E 40.00 GHz n 2 ⇣1/⇣co+1/⇣adj 33.99 dB
✓BW 0.20 ACI 30.00 dB
rc ⇡ 65.50 km XPiso 27.78 dB
The ACI is considered a constant value, which is obtained from [66]. As a result,
the ⇣1/⇣co+1/⇣adj evaluation gives a value of 33.99 dB, which is higher than the average
value of 29.00 dB [18]. Therefore, the offset-parabolic-reflector antenna design is
robustness in terms of isolation between other feeder links, assuring an excellent
performance for the proposed HTS system.
Figure 4.5 depicts the spot beams of each feeder downlink projected over the Latin
America area. Details of geographic locations for the spot beams are in Appendix
A.1. This choice was based on the fact that those major cities could access multiple
infrastructures of various types, guaranteeing the fast deployment of facilities for
gateways stations. Furthermore, it was analyzed the reaching of the HTS system in
the entire Latin American region, theoretically.
Meanwhile, Table 4.7 indicates the CIR results for the 3-color scheme, evaluating
both simple and double polarization. In this evaluation, the frequency band was
of 20.00 GHz in the Ka-band, the HTS orbital position was 92.0 W, as was
discussed in subsection 2.2.1, and the taper roll-off, n, was of zero for all cases. The
co-polar evaluation was simplified by considering the closest six interferences [9,59,62].
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .50
Figure 4.5: Spot beams of feeder downlinks over the Latin America region.
Furthermore, the crossover (triple-point) level was calculated by using the method
from [64].
To sum up, the CIR results are very promising where each evaluation is higher
than the range between 14.5–17.6 dB. Therefore, this scheme can be used as the main
layout in order to cover the ground segment in Ka-band. Figure 4.6 illustrates the
3-color frequency-reuse schemes over the Mexico area. For the 4-color frequency-reuse
scheme, the evaluation method was the same as the 3-color scheme, for this reason,
the parameters were exactly the same. Table 4.8 shows the CIR results for the 4-color
scheme, assessing both simple and double polarization.
The 4-color frequency-reuse scheme is more aggressive than the 3-color scheme,
therefore, the CIR results for the 4-color scheme are lower than CIR results for 3-color
schemes. Although the CIR values are a few dBs lower than CIR values for 3-color
schemes, it does not mean bad results. In fact, the CIR results for 4-color schemes
are within the range value between 14.5–17.5 dB, where the scheme 4F1P overcomes
slightly that range. Figure 4.7 illustrates the 4-color frequency-reuse scheme over the
Mexico area.
In conclusion, the antenna results are very encouraging, giving an excellent outlook
on the design of the offset-parabolic-reflector antennas for HTS systems. The double
polarization increases the capacity of the link without degrading the signal at the
reception antenna. It is also important to mention that the larger beam spacing
enables a proportionate increase in the feed horn size, which improves the antenna
gain by reducing spillover loss, i.e., the reflector illuminates optimally increasing the
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .51
Table 4.7: The CIR Evaluation for the 3-color Frequency-reuse Scheme.
Denom. Values
✓BW 0.48 0.40 0.32 0.26
rc ⇡ 150 km ⇡ 125 km ⇡ 100 km ⇡ 80 km
Dant 2.19 m 2.63 m 3.28 m 4.10 m
EI 12.39 dB 12.39 dB 12.39 dB 12.39 dB
ACI 30.00 dB 30.00 dB 30.00 dB 30.00 dB
XPiso 27.39 dB 27.33 dB 27.25 dB 27.16 dB
Results
Denom. 3F1P Scheme
⇣co 21.36 dB 21.29 dB 21.20 dB 21.03 dB
⇣adj 25.49 dB 25.45 dB 25.40 dB 25.34 dB
⇣1/⇣co+1/⇣adj 19.94 dB 19.88 dB 19.80 dB 19.66 dB
Denom. 3F2P Scheme
⇣co 21.36 dB 21.29 dB 21.20 dB 21.03 dB
⇣adj 24.32 dB 24.19 dB 24.07 dB 23.99 dB
⇣1/⇣co+1/⇣adj 19.58 dB 19.49 dB 19.39 dB 19.25 dB
Crossover level 2.88 dB 2.97 dB 3.05 dB 3.10 dB
beam End-of-Coverage (EoC) gain and reducing sidelobe levels.
Furthermore, these results demonstrate that both 3-color and 4-color frequency
reuse schemes accomplish the interference requirements in order to generate
spot beams from offset-parabolic-reflector antennas, standing out the 3-color
frequency-reuse schemes so that it is possible to recommend these layouts for the
HTS systems.
Finally, the Latin America region requires a lot of spot beams in order to cover
the entire area. For this purpose, an HTS fleet is necessary to supply the demand for
spot beams in the region.
4.5.4 Link Margin
According to [8], the aim of link margin evaluation is to determine the positive margin
of operation between the feeder uplink and the user downlink by using the combined
CNIR, i.e., this analysis is only for the forward link in clear-sky conditions. The
average value of the CNR for feeder uplinks in the V band was about 32.36 dB, whereas
the average value of the CNR for user downlinks in the Ka-band was about 24.61 dB,
as was discussed in Chapter 3. Furthermore, the intermodulation interference (⇣im)
is a constant value of 25.00 dB [8]. Table 4.9 shows the CNIR for both cases by using
the interferences values obtained from simulations. For user downlinks, it was used
the 3F1P scheme with a beamwidth (✓BW ) of 0.32 .
As a result, the combined CNIR for the forward link was 17.43 dB for the
maximum co-channel CIR (⇣co) for both feeder uplink and user downlink, which meant
a margin of 3.93 dB over the required value of 13.50 dB [8]. Figure 4.8 depicts the
link margin as a function of co-channel CIR that ranges from 12.00 to 25.00 dB.
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .52
(a) ✓BW = 0.26 (b) ✓BW = 0.32
(c) ✓BW = 0.40 (d) ✓BW = 0.48
Figure 4.6: The 3-color frequency-reuse schemes over the Mexico region with 4 different
beamwidths.
It is important to note that around 17.60 dB of co-channel CIR is required to
obtain a positive margin. The co-channel CIR for the user downlink is much lower
than the feeder uplink CIR, for this reason, it is a significant challenge and limiting
factor so that it is necessary to further studies to improve the CIR antenna. In
conclusion, these evaluations of offset-parabolic-antennas give excellent results but it
is imperative to carry out real trials in order to obtain a more propitious scenario. All
calculations in this thesis were performed on an Intel Core i5 2⇥ 2.70 GHz machine
with 8 GB RAM.
4.6 Contributions of the Research
The results of this Chapter were published for The International Journal
Of Engineering and Science and Congreso Internacional de Computacion y
Telecomunicaciones, COMTEL, as follows. However, It is important to mention that
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .53
Table 4.8: The CIR Evaluation for the 4-color Frequency-reuse Scheme.
Denom. Values
✓BW 0.48 0.40 0.32 0.26
rc ⇡ 150 km ⇡ 125 km ⇡ 100 km ⇡ 80 km
Dant 2.19 m 2.63 m 3.28 m 4.10 m
EI 12.39 dB 12.39 dB 12.39 dB 12.39 dB
ACI 30.00 dB 30.00 dB 30.00 dB 30.00 dB
XPiso 27.39 dB 27.33 dB 27.25 dB 27.16 dB
Results
Denom. 4F1P Scheme
⇣co 18.95 dB 18.95 dB 18.95 dB 18.94 dB
⇣adj 25.49 dB 25.45 dB 25.40 dB 25.34 dB
⇣1/⇣co+1/⇣adj 18.07 dB 18.07 dB 18.06 dB 18.04 dB
Denom. 4F2P Scheme
⇣co 18.95 dB 18.95 dB 18.95 dB 18.94 dB
⇣adj 23.14 dB 23.09 dB 23.03 dB 22.96 dB
⇣1/⇣co+1/⇣adj 17.54 dB 17.53 dB 17.51 dB 17.49 dB
Crossover level 3.84 dB 3.96 dB 4.07 dB 4.13 dB
Table 4.9: The CNIR Values for the Forward Link.
Feeder Uplink User Downlink
Parameters Values Parameters Values
CNR 32.36 dB CNR 24.61 dB
⇣co 36.09 dB ⇣co 21.20 dB
⇣adj 38.17 dB ⇣adj 25.40 dB
⇣im 25.00 dB
CNIRup 30.09 dB CNIRdn 17.67 dB
the conference proceedings were published in "Revista de Tecnología e Información"
by Universidad Inca Garcilaso de la Vega, Lima-Peru.
• Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez-Rodriguez, and
Jose M Matias. Interference Evaluations in Frequency Reuse by Using Offset-
Parabolic-Reflector Antennas for a UHTS System. Int. J. Eng. Sci., 7(8):34–45,
2018.
• Andres Cornejo, Salvador Landeros-Ayala, Ramon Martínez Rodríguez-Osorio,
and Jose M Matias. A method for designing an offset-parabolic-reflector
antenna for a ultra-high throughput satellite. In Rev. Tecnol. e Inf., volume
16, pages 84–87, Lima, Peru, 2018. Universidad Inca Garcilaso de la Vega.
CHAPTER 4. INTERFERENCE EVALUATIONS IN FREQUENCY REUSE . . .54
(a) ✓BW = 0.26 (b) ✓BW = 0.32
(c) ✓BW = 0.40 (d) ✓BW = 0.48
Figure 4.7: The 4-color frequency-reuse schemes over the Mexico region with 4 different
beamwidths.
L
in
k
M
ar
g
in
[
d
B
]
-5.00
-3.75
-2.50
-1.25
0.00
1.25
2.50
3.75
5.00
Co-channel CIR [dB]
12 13 14 15 16 17 18 19 20 21 22 23 24 25
Figure 4.8: Forward link margin vs. co-channel CIR.
Chapter 5
Method of Rain Attenuation
Prediction Based On Long-Short
Term Memory Network
5.1 Introduction
Satellite communication links operate at different frequency bands such as L band, C
band, and higher. Despite the higher frequency bands have more available spectrum,
the satellite links are more susceptible to weather impairments, especially the rain
attenuation. For this reason, a lot of research has been carried out in order to
find accurate models of rain attenuation, whose methods have been developed by
stochastic processes [37,68,69]. The Maseng-Bakken model is currently employed for
the International Telecommunication Union (ITU) and its Recommendation ITU-R
P.1853 [38]. Other researchers have been performed experimental models, such as
the Prévision d’Ensemble ARPEGE (PEARP) system by Météo-France [15] and fade
mitigation experiments for the ALPHASAT Satellite [70]. However, the main problem
is the lack of historical rain database for each location and very few experimental
models so that the stochastic models are the only theoretical methods in order to
determine the main implications of the rain attenuation over satellite links.
Nowadays, Machine Learning techniques are employed by several real-world
applications as well as theoretical problems, especially in clustering, classification,
and regression problems. To be specific, Machine learning is the study of computer
algorithms that improve automatically through experience. It is related to a subfield
of computer sciences and is considered a subset of artificial intelligence. In this
context, a novel method is proposed in this thesis by a machine learning method,
to be precise, a deep learning technique. Thereby, the Long-short Term Memory
(LSTM) network is a deep learning technique and predicts, accurately, events of
rain-attenuation without relying on either mathematical or stochastic models.
The main objective is to train and validate the proposed models based on LSTM
networks by using artificial rain attenuation time-series at the input. That is, the
model learns from experience where the deep learning algorithms employ advanced
55
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 56
computational methods in order to learn directly from input data without appealing
to predetermined models or equations. Moreover, the rain-attenuation outcomes are
useful to determine the status-in-advance of the satellite link and to anticipate a
sudden link outage due to heavy rain. Indeed, it is important to obtain reliable
results with the aim of guaranteeing the availability of satellite links and improving
mechanisms such as the satellite link switching [71–73], site diversity on the ground
segment [14, 22, 26], and uplink power control (ULPC) on satellite links [74–76],
especially in extremely high-frequency bands (EHF).
Finally, a comparison with other methods/models is employed to determine the
best model in terms of performance, generalization, and accuracy.
5.2 Implementation of the Proposed Deep Learning
Method Based on LSTM Network
5.2.1 Deep Learning Network: Model Architecture
LSTM is based on a recurrent architecture network, which is designed to overcome
error back-flow problems. An LSTM layer is able to support sequence data and
time series in a network. Despite the input sequence data can be either noisy
or incompressible, the LSTM network can learn to join time intervals in excess of
1000 steps without losing short-time-capabilities [77]. This is possible by an efficient
gradient-based algorithm. In this thesis, there is a need to predict the values of future
time steps of the input time-series. Therefore, the input time-series are trained by the
proposed LSTM network with sequence-to-sequence mode, where the output data are
sequences with values shifted by a one time step. In other words, the LSTM network
learns to predict the value of the next time step, t + 1, for each time step, t, of the
input time-series sequence.
A. The Architecture of a Single LSTM Cell
Consider a dataset x = (x(1),x(2), . . . ,x(m)), where each element is a vector (sequence)
x(m) 2 R
K , and m is the number of examples of each dataset. For simplicity and
readability, this general approach only considers m = 1 examples in dataset x. Figure
5.1 depicts the architecture of a single LSTM cell, providing two recurrent features
(LSTM cell outputs). The outputs are also known as the hidden (ht) and cell (ct)
states.
For this purpose, let U = [0, 1] indicate the unit interval, whereas let ±U = [ 1, 1].
Denoted by L, the LSTM cell is a mathematical function that enables three inputs
and generates two outputs. The LSTM cell function can be expressed as
(ht, ct) = L(ht 1, ct 1,xt), (5.1)
where ht, ht 1, ct, ct 1 2 ±U and the input vector xt 2 R
K [77, 78]. In order to
compute the first output and updated cell state, the initial state of the cell and the
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 57
Forget Update Output
ht−1 ∈ ±U
AAACAXicbVBNS8NAEN34WetX1IvgZbEIXixJFfRY9OKxgmkLTQib7aZdursJuxuhhHrxr3jxoIhX/4U3/42bNgdtfTDweG+GmXlRyqjSjvNtLS2vrK6tVzaqm1vbO7v23n5bJZnExMMJS2Q3QoowKoinqWakm0qCeMRIJxrdFH7ngUhFE3GvxykJOBoIGlOMtJFC+3AY5vrMnfhUQD/lPkd6GEW5NwntmlN3poCLxC1JDZRohfaX309wxonQmCGleq6T6iBHUlPMyKTqZ4qkCI/QgPQMFYgTFeTTDybwxCh9GCfSlNBwqv6eyBFXaswj01lcqOa9QvzP62U6vgpyKtJME4Fni+KMQZ3AIg7Yp5JgzcaGICypuRXiIZIIaxNa1YTgzr+8SNqNunteb9xd1JrXZRwVcASOwSlwwSVoglvQAh7A4BE8g1fwZj1ZL9a79TFrXbLKmQPwB9bnD1d3ltM=
ct−1 ∈ ±U
AAACAXicbVBNS8NAEN34WetX1IvgZbEIXixJFfRY9OKxgmkLTQib7bZdursJuxuhhHjxr3jxoIhX/4U3/42bNgdtfTDweG+GmXlRwqjSjvNtLS2vrK6tVzaqm1vbO7v23n5bxanExMMxi2U3QoowKoinqWakm0iCeMRIJxrfFH7ngUhFY3GvJwkJOBoKOqAYaSOF9iEOM33m5j4V0E+4z5EeRVHm5aFdc+rOFHCRuCWpgRKt0P7y+zFOOREaM6RUz3USHWRIaooZyat+qkiC8BgNSc9QgThRQTb9IIcnRunDQSxNCQ2n6u+JDHGlJjwyncWFat4rxP+8XqoHV0FGRZJqIvBs0SBlUMewiAP2qSRYs4khCEtqboV4hCTC2oRWNSG48y8vknaj7p7XG3cXteZ1GUcFHIFjcApccAma4Ba0gAcweATP4BW8WU/Wi/Vufcxal6xy5gD8gfX5A09wls4=
×
AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0mqoMeiF48V7Ae0oWy2m3btZhN2J0IJ/Q9ePCji1f/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6wEnC/YgOlQgFo2ilVg9FxE2/XHGr7hxklXg5qUCORr/81RvELI24QiapMV3PTdDPqEbBJJ+WeqnhCWVjOuRdSxW1S/xsfu2UnFllQMJY21JI5urviYxGxkyiwHZGFEdm2ZuJ/3ndFMNrPxMqSZErtlgUppJgTGavk4HQnKGcWEKZFvZWwkZUU4Y2oJINwVt+eZW0alXvolq7v6zUb/I4inACp3AOHlxBHe6gAU1g8AjP8ApvTuy8OO/Ox6K14OQzx/AHzucPt3mPOA==
ct ∈ ±U
AAAB/nicbVBPS8MwHE39O+e/qnjyEhyCp9FOQY9DLx4n2G2wlpJm6RaWpCVJhVEKfhUvHhTx6ufw5rcx3XrQzQeBx3u/H7+XF6WMKu0439bK6tr6xmZtq769s7u3bx8cdlWSSUw8nLBE9iOkCKOCeJpqRvqpJIhHjPSiyW3p9x6JVDQRD3qakoCjkaAxxUgbKbSPcZjrwqfCT7nPkR5HUe4Vod1wms4McJm4FWmACp3Q/vKHCc44ERozpNTAdVId5Ehqihkp6n6mSIrwBI3IwFCBOFFBPotfwDOjDGGcSPOEhjP190aOuFJTHpnJMqFa9ErxP2+Q6fg6yKlIM00Enh+KMwZ1Assu4JBKgjWbGoKwpCYrxGMkEdamsbopwV388jLptpruRbN1f9lo31R11MAJOAXnwAVXoA3uQAd4AIMcPINX8GY9WS/Wu/UxH12xqp0j8AfW5w8RtpYy
ht ∈ ±U
AAAB/nicbVBPS8MwHE39O+e/qnjyEhyCp9FOQY9DLx4n2G2wlpJm6RaWpCVJhVEKfhUvHhTx6ufw5rcx3XrQzQeBx3u/H7+XF6WMKu0439bK6tr6xmZtq769s7u3bx8cdlWSSUw8nLBE9iOkCKOCeJpqRvqpJIhHjPSiyW3p9x6JVDQRD3qakoCjkaAxxUgbKbSPx2GuC58KP+U+R3ocRblXhHbDaTozwGXiVqQBKnRC+8sfJjjjRGjMkFID10l1kCOpKWakqPuZIinCEzQiA0MF4kQF+Sx+Ac+MMoRxIs0TGs7U3xs54kpNeWQmy4Rq0SvF/7xBpuPrIKcizTQReH4ozhjUCSy7gEMqCdZsagjCkpqsEI+RRFibxuqmBHfxy8uk22q6F83W/WWjfVPVUQMn4BScAxdcgTa4Ax3gAQxy8AxewZv1ZL1Y79bHfHTFqnaOwB9Ynz8ZrpY3
+
AAAB6HicbVDLSgNBEOyNrxhfUY9eBoMgCGE3CnoMevGYgHlAsoTZSW8yZnZ2mZkVQsgXePGgiFc/yZt/4yTZgyYWNBRV3XR3BYng2rjut5NbW9/Y3MpvF3Z29/YPiodHTR2nimGDxSJW7YBqFFxiw3AjsJ0opFEgsBWM7mZ+6wmV5rF8MOME/YgOJA85o8ZK9YteseSW3TnIKvEyUoIMtV7xq9uPWRqhNExQrTuemxh/QpXhTOC00E01JpSN6AA7lkoaofYn80On5MwqfRLGypY0ZK7+npjQSOtxFNjOiJqhXvZm4n9eJzXhjT/hMkkNSrZYFKaCmJjMviZ9rpAZMbaEMsXtrYQNqaLM2GwKNgRv+eVV0qyUvctypX5Vqt5mceThBE7hHDy4hircQw0awADhGV7hzXl0Xpx352PRmnOymWP4A+fzB3LTjLM=
×
AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0mqoMeiF48V7Ae0oWy2m3btZhN2J0IJ/Q9ePCji1f/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6wEnC/YgOlQgFo2ilVg9FxE2/XHGr7hxklXg5qUCORr/81RvELI24QiapMV3PTdDPqEbBJJ+WeqnhCWVjOuRdSxW1S/xsfu2UnFllQMJY21JI5urviYxGxkyiwHZGFEdm2ZuJ/3ndFMNrPxMqSZErtlgUppJgTGavk4HQnKGcWEKZFvZWwkZUU4Y2oJINwVt+eZW0alXvolq7v6zUb/I4inACp3AOHlxBHe6gAU1g8AjP8ApvTuy8OO/Ox6K14OQzx/AHzucPt3mPOA==
xt ∈ R
K
AAACBXicbVC7TsMwFHXKq5RXgBEGiwqJqUoKEowVLEgsBdGH1ITIcZ3WquNEtoOooiws/AoLAwix8g9s/A1OmwFajmTp+Jx7de89fsyoVJb1bZQWFpeWV8qrlbX1jc0tc3unLaNEYNLCEYtE10eSMMpJS1HFSDcWBIU+Ix1/dJH7nXsiJI34rRrHxA3RgNOAYqS05Jn7TojU0A/Sh8xTDuVw+vfTm+zuyjOrVs2aAM4TuyBVUKDpmV9OP8JJSLjCDEnZs61YuSkSimJGsoqTSBIjPEID0tOUo5BIN51ckcFDrfRhEAn9uIIT9XdHikIpx6GvK/Md5ayXi/95vUQFZ25KeZwowvF0UJAwqCKYRwL7VBCs2FgThAXVu0I8RAJhpYOr6BDs2ZPnSbtes49r9euTauO8iKMM9sABOAI2OAUNcAmaoAUweATP4BW8GU/Gi/FufExLS0bRswv+wPj8AQnkmOg=
×
AAAB7XicbVBNS8NAEJ3Ur1q/qh69LBbBU0mqoMeiF48V7Ae0oWy2m3btZhN2J0IJ/Q9ePCji1f/jzX/jts1BWx8MPN6bYWZekEhh0HW/ncLa+sbmVnG7tLO7t39QPjxqmTjVjDdZLGPdCajhUijeRIGSdxLNaRRI3g7GtzO//cS1EbF6wEnC/YgOlQgFo2ilVg9FxE2/XHGr7hxklXg5qUCORr/81RvELI24QiapMV3PTdDPqEbBJJ+WeqnhCWVjOuRdSxW1S/xsfu2UnFllQMJY21JI5urviYxGxkyiwHZGFEdm2ZuJ/3ndFMNrPxMqSZErtlgUppJgTGavk4HQnKGcWEKZFvZWwkZUU4Y2oJINwVt+eZW0alXvolq7v6zUb/I4inACp3AOHlxBHe6gAU1g8AjP8ApvTuy8OO/Ox6K14OQzx/AHzucPt3mPOA==
tanh
AAAB7HicbVBNS8NAEJ34WetX1aOXxSJ4KkkV9Fj04rGCaQttKJvtpl262YTdiVBCf4MXD4p49Qd589+4bXPQ1gcDj/dmmJkXplIYdN1vZ219Y3Nru7RT3t3bPzisHB23TJJpxn2WyER3Qmq4FIr7KFDyTqo5jUPJ2+H4bua3n7g2IlGPOEl5ENOhEpFgFK3k95CqUb9SdWvuHGSVeAWpQoFmv/LVGyQsi7lCJqkxXc9NMcipRsEkn5Z7meEpZWM65F1LFY25CfL5sVNybpUBiRJtSyGZq78nchobM4lD2xlTHJllbyb+53UzjG6CXKg0Q67YYlGUSYIJmX1OBkJzhnJiCWVa2FsJG1FNGdp8yjYEb/nlVdKq17zLWv3hqtq4LeIowSmcwQV4cA0NuIcm+MBAwDO8wpujnBfn3flYtK45xcwJ/IHz+QPcCI63
fgt
AAAB7nicbVBNS8NAEJ34WetX1aOXxSJ4KkkV9Fj04rGC/YA2hM120y7dbMLuRCihP8KLB0W8+nu8+W/ctjlo64OBx3szzMwLUykMuu63s7a+sbm1Xdop7+7tHxxWjo7bJsk04y2WyER3Q2q4FIq3UKDk3VRzGoeSd8Lx3czvPHFtRKIecZJyP6ZDJSLBKFqpEwX5MMBpUKm6NXcOskq8glShQDOofPUHCctirpBJakzPc1P0c6pRMMmn5X5meErZmA55z1JFY278fH7ulJxbZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif95vQyjGz8XKs2QK7ZYFGWSYEJmv5OB0JyhnFhCmRb2VsJGVFOGNqGyDcFbfnmVtOs177JWf7iqNm6LOEpwCmdwAR5cQwPuoQktYDCGZ3iFNyd1Xpx352PRuuYUMyfwB87nD5ZWj7s=
igt
AAAB7nicbVBNS8NAEJ34WetX1aOXxSJ4KkkV9Fj04rGC/YA2hM120y7dbMLuRCihP8KLB0W8+nu8+W/ctjlo64OBx3szzMwLUykMuu63s7a+sbm1Xdop7+7tHxxWjo7bJsk04y2WyER3Q2q4FIq3UKDk3VRzGoeSd8Lx3czvPHFtRKIecZJyP6ZDJSLBKFqpI4J8GOA0qFTdmjsHWSVeQapQoBlUvvqDhGUxV8gkNabnuSn6OdUomOTTcj8zPKVsTIe8Z6miMTd+Pj93Ss6tMiBRom0pJHP190ROY2MmcWg7Y4ojs+zNxP+8XobRjZ8LlWbIFVssijJJMCGz38lAaM5QTiyhTAt7K2EjqilDm1DZhuAtv7xK2vWad1mrP1xVG7dFHCU4hTO4AA+uoQH30IQWMBjDM7zCm5M6L86787FoXXOKmRP4A+fzB5r0j74=
cut
AAAB7nicbVBNS8NAEJ34WetX1aOXxSJ4KkkV9Fj04rGC/YA2hM120y7dbMLuRCihP8KLB0W8+nu8+W/ctjlo64OBx3szzMwLUykMuu63s7a+sbm1Xdop7+7tHxxWjo7bJsk04y2WyER3Q2q4FIq3UKDk3VRzGoeSd8Lx3czvPHFtRKIecZJyP6ZDJSLBKFqpw4I8C3AaVKpuzZ2DrBKvIFUo0AwqX/1BwrKYK2SSGtPz3BT9nGoUTPJpuZ8ZnlI2pkPes1TRmBs/n587JedWGZAo0bYUkrn6eyKnsTGTOLSdMcWRWfZm4n9eL8Poxs+FSjPkii0WRZkkmJDZ72QgNGcoJ5ZQpoW9lbAR1ZShTahsQ/CWX14l7XrNu6zVH66qjdsijhKcwhlcgAfX0IB7aEILGIzhGV7hzUmdF+fd+Vi0rjnFzAn8gfP5A6caj8Y=
ogt
AAAB7nicbVBNS8NAEJ34WetX1aOXxSJ4KkkV9Fj04rGC/YA2hM120y7dbMLuRCihP8KLB0W8+nu8+W/ctjlo64OBx3szzMwLUykMuu63s7a+sbm1Xdop7+7tHxxWjo7bJsk04y2WyER3Q2q4FIq3UKDk3VRzGoeSd8Lx3czvPHFtRKIecZJyP6ZDJSLBKFqpkwT5MMBpUKm6NXcOskq8glShQDOofPUHCctirpBJakzPc1P0c6pRMMmn5X5meErZmA55z1JFY278fH7ulJxbZUCiRNtSSObq74mcxsZM4tB2xhRHZtmbif95vQyjGz8XKs2QK7ZYFGWSYEJmv5OB0JyhnFhCmRb2VsJGVFOGNqGyDcFbfnmVtOs177JWf7iqNm6LOEpwCmdwAR5cQwPuoQktYDCGZ3iFNyd1Xpx352PRuuYUMyfwB87nD6Qwj8Q=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
Figure 5.1: A single LSTM cell diagram [78].
first time step of the sequence are used by the LSTM cell. That is, the current state
of the cell, at time step t, is employed by the LSTM cell (ht 1, ct 1,), whereas the
next time step of the sequence is used to compute the output state ht, in addition to
updating the cell state ct. Therefore, the hidden and cell states are part of the state
of the cell.
At time step t, ht and ct states are left at the output so that the same cell can
be fed back by these states at time step t + 1. It is important to mention that an
element of the input sequence x ∈ R
K is also fed into the cell any time step t [78].
Meanwhile, the learned information from previous time steps is contained in the cell
state. In summary, the cell appends information to or deletes information from the
cell state at each time step, in other words, the cell is able to control these updates
employing gates.
For this purpose, the flow of data plays an important role in the LSTM cell,
where the hidden state, ht 1 ∈ ±U, and the input vector, xt ∈ R
K , feed three gates
(functions), as can be seen in Figure 5.1. At the time t, the three gates (forget, input,
and output functions) are respectively given by
fgt(xt, ht 1) = g(wx
T
f xt + whf
ht 1 + bf ) ∈ U, (5.2)
igt(xt, ht 1) = g(wx
T
i xt + whi
ht 1 + bi) ∈ U, (5.3)
ogt(xt, ht 1) = g(wx
T
o xt + who
ht 1 + bo) ∈ U, (5.4)
where g is the gate activation function, wxf ,wxi,wxo ∈ R
K are the weight vectors,
and whf
, whi
, who
, bf , bi, bo ∈ R are the recurrent weights and biases, respectively.
During the training process of the cell, the weights are capable of learning. By default,
g is based on the sigmoid function to compute the gate activation function. For this
reason, the gates of a single LSTM cell produce scalar values. Figure 5.1 stands out
how the gates forget, update, and output the cell and hidden states. These three
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 58
gates can be interpreted as switches, that is, if their outputs are near 1, then they
are on. Otherwise, when the outputs are close to 0, then they are off.
Furthermore, the cell update function is built from a single neuron, which, at time
step t, is expressed as
cut
(xt, ht 1) = tanh(wx
Txt + whht 1 + b) 2 ±U, (5.5)
where tanh is the state activation function, wx 2 R
K is the weight vector, and wh, b 2
R are the recurrent weights and bias, respectively. Also, these weight parameters
are learned in the training process. The state activation function is based on the
hyperbolic tangent function, which is also set by default. Specifically, the tanh and
sigmoid activation functions are discussed in Section 5.2.2.
Finally, the forget gate, fg, controls the level of the cell state reset (forget), whereas
the input gate, ig, controls the level of the cell state update, cu, that adds it to the
cell state. Moreover, the output gate, og, controls the modified cell state to become
the next hidden state [77, 78]. As a result, the new cell and hidden states, at time
step t, can be expressed as
ct = fgt · ct 1 + igt · cut
2 ±U, (5.6)
ht = ogt · tanh(ct) 2 ±U, (5.7)
where both values are scalars. For readability, it is important to note that all of these
functions and parameters are encapsulated in the function L, as can be seen in Eq.
(5.1).
B. The Architecture of a Layer of LSTM Cells
Figure 5.2 illustrates a layer composed of several LSTM cells U , also known as the
number of hidden units, and features C. For this purpose, the LSTM layer, denoted
by LU , is the concatenation of LSTM cells, L1,L2, . . . ,LU , creating an LSTM Block.
At time step t, both hidden, ht, and cell, ct, states indicate the LSTM Block output.
Each cell has a different set of weight parameters, thus, the mathematical expression
of the LSTM layer is given by
(h1t , c1t) = L1(h1(t−1)
, c1(t−1)
,xt),
(h2t , c2t) = L2(h2(t−1)
, c2(t−1)
,xt),
...
(hUt
, cUt
) = LU(hU(t−1)
, cU(t−1)
,xt),
therefore, the expression can be also rewritten as
(ht, ct) = LU(ht 1, ct 1,xt), (5.8)
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 59
Number of
Hidden Units
Number of
Features
Number of
Time Steps
LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
c
t−1
h
t−1
c
t
h
t
LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
Initial
State
Final
State
LSTM
Block
LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
LSTM
Block
LSTM
Block
LSTM
Layer
……LSTM
Block
x11
x21
...
xC1
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
x12
x22
.
.
.
xC2
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
x1t
x2t
.
.
.
xCt
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
x1K
x2K
.
.
.
xCK
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
h11
h21
.
.
.
hU1
AAACOXicbZDLSsNAFIYn3o23qks3wSK4KkkVdFl047KC1UITwmRy0g5OJmHmpFBCX8uNb+FOcONCEbe+gNM2C7UeGPj4z23OH+WCa3TdZ2thcWl5ZXVt3d7Y3Nreqe3u3eqsUAw6LBOZ6kZUg+ASOshRQDdXQNNIwF10fznJ3w1BaZ7JGxzlEKS0L3nCGUUjhbW2LyDBnu1H0OeypErR0bhkY3sQll7ojX1/Qs0Z+cM4Qz2TOkayfZBx1WP7ivcHGIS1uttwp+HMg1dBnVTRDmtPfpyxIgWJTFCte56bY2CmImcCzNxCQ07ZPe1Dz6CkKeignF4+do6MEjtJpsyT6EzVnx0lTbUepZGpTCkO9N/cRPwv1yswOQ9KLvMCQbLZoqQQDmbOxEYn5goYipEByhQ3f3XYgCrK0JhtGxO8vyfPw22z4Z00mten9dZFZccaOSCH5Jh45Iy0yBVpkw5h5IG8kDfybj1ar9aH9TkrXbCqnn3yK6yvbyvLrSQ=
h12
h22
.
.
.
hU2
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
h1t
h2t
.
.
.
hUt
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
h1K
h2K
.
.
.
hUK
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
Figure 5.2: The architecture of an LSTM layer [77].
where ht,ht 1, ct, ct 1 2 ±U
U , and xt 2 R
K . As a result, the cell and hidden states,
at time step t, can be respectively given by
ct = fgt ct 1 + igt cut
2 ±U
U , (5.9)
ht = ogt tanh(ct) 2 ±U
U , (5.10)
where is the Hadamard product, that is the element-wise multiplication of
vectors. Moreover, the dot products become matrix-vector products. Here, the
individual weight vectors and biases from each LSTM cells can be stored in matrices.
Therefore, the three gates and the cell update function contain weight matrices,
Wxf ,Wxi,Wxo,Wx 2 R
U⇥K , respectively. Meanwhile, the stacked hidden state
is a U -dimensional vector, therefore, the gates contain recurrent weight matrices,
Whf ,Whi,Who,Wh 2 R
U⇥U , and bias vectors bf ,bi,bo,b 2 R
U [78].
As a shorthand, all parameters of the LSTM network are usually concatenated
and expressed as a whole by parameter ΘL [79]. For notational convenience, the
parameter ΘL can be written as
ΘL ⌘
8
>
>
<
>
>
:
Wxf , Whf , bf
Wxi, Whi, bi
Wxo, Who, bo
Wx, Wh, b
9
>
>
=
>
>
;
(5.11)
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 60
C. The Architecture of an LSTM Multi-Layer Network
In this case, multiple layers could improve the output accuracy, increasing the LSTM
network complexity. At time step t, the function LU has two outputs ht, ct 2 U
U ,
thereby, the next LSTM layer can be sequentially fed by the hidden state from the
previous LSTM layer. Figure 5.3 depicts how the states pass through an LSTM
multi-layer network.
LU1
AAAB+XicbVBNS8NAFHzxs9avqEcvi0XwVJIq6LHoxYOHCqYttCFstpt26WYTdjeFEvpPvHhQxKv/xJv/xk2bg7YOLAwz7/FmJ0w5U9pxvq219Y3Nre3KTnV3b//g0D46bqskk4R6JOGJ7IZYUc4E9TTTnHZTSXEcctoJx3eF35lQqVginvQ0pX6Mh4JFjGBtpMC2+zHWI4J5/jALcs+dBXbNqTtzoFXilqQGJVqB/dUfJCSLqdCEY6V6rpNqP8dSM8LprNrPFE0xGeMh7RkqcEyVn8+Tz9C5UQYoSqR5QqO5+nsjx7FS0zg0k0VOtewV4n9eL9PRjZ8zkWaaCrI4FGUc6QQVNaABk5RoPjUEE8lMVkRGWGKiTVlVU4K7/OVV0m7U3ct64/Gq1rwt66jAKZzBBbhwDU24hxZ4QGACz/AKb1ZuvVjv1sdidM0qd07gD6zPH6Btk6Y=
LU2
AAAB+XicbVBNS8NAFHzxs9avqEcvi0XwVJIq6LHoxYOHCqYttCFsttt26WYTdjeFEvJPvHhQxKv/xJv/xk2bg7YOLAwz7/FmJ0w4U9pxvq219Y3Nre3KTnV3b//g0D46bqs4lYR6JOax7IZYUc4E9TTTnHYTSXEUctoJJ3eF35lSqVgsnvQsoX6ER4INGcHaSIFt9yOsxwTz7CEPMq+RB3bNqTtzoFXilqQGJVqB/dUfxCSNqNCEY6V6rpNoP8NSM8JpXu2niiaYTPCI9gwVOKLKz+bJc3RulAEaxtI8odFc/b2R4UipWRSaySKnWvYK8T+vl+rhjZ8xkaSaCrI4NEw50jEqakADJinRfGYIJpKZrIiMscREm7KqpgR3+curpN2ou5f1xuNVrXlb1lGBUziDC3DhGppwDy3wgMAUnuEV3qzMerHerY/F6JpV7pzAH1ifP6Hyk6c=
LU3
AAAB+XicbVBNS8NAFHypX7V+RT16CRbBU0laQY9FLx48VDBtoQ1hs920SzebsLsplJB/4sWDIl79J978N27aHLR1YGGYeY83O0HCqFS2/W1UNja3tnequ7W9/YPDI/P4pCvjVGDi4pjFoh8gSRjlxFVUMdJPBEFRwEgvmN4Vfm9GhKQxf1LzhHgRGnMaUoyUlnzTHEZITTBi2UPuZ24r98263bAXsNaJU5I6lOj45tdwFOM0IlxhhqQcOHaivAwJRTEjeW2YSpIgPEVjMtCUo4hIL1skz60LrYysMBb6cWUt1N8bGYqknEeBnixyylWvEP/zBqkKb7yM8iRVhOPloTBlloqtogZrRAXBis01QVhQndXCEyQQVrqsmi7BWf3yOuk2G06r0Xy8qrdvyzqqcAbncAkOXEMb7qEDLmCYwTO8wpuRGS/Gu/GxHK0Y5c4p/IHx+QOjd5Oo
xt ∈ R
K
AAACBnicbVDLSsNAFJ3UV62vqEsRBovgqiRV0GXRjeCmin1AE8NkOmmHTiZhZiKWkJUbf8WNC0Xc+g3u/BsnbRbaeuDC4Zx7ufceP2ZUKsv6NkoLi0vLK+XVytr6xuaWub3TllEiMGnhiEWi6yNJGOWkpahipBsLgkKfkY4/usj9zj0Rkkb8Vo1j4oZowGlAMVJa8sx9J0Rq6AfpQ+Yp6FAOp4Kf3mR3V55ZtWrWBHCe2AWpggJNz/xy+hFOQsIVZkjKnm3Fyk2RUBQzklWcRJIY4REakJ6mHIVEuunkjQweaqUPg0jo4gpO1N8TKQqlHIe+7sxvlLNeLv7n9RIVnLkp5XGiCMfTRUHCoIpgngnsU0GwYmNNEBZU3wrxEAmElU6uokOwZ1+eJ+16zT6u1a9Pqo3zIo4y2AMH4AjY4BQ0wCVoghbA4BE8g1fwZjwZL8a78TFtLRnFzC74A+PzB2aumRI=
hU1(t−1)
AAACAHicbVBNS8NAEN3Ur1q/oh48eFksQj1YkirosejFYwXTFtoQNttNu3Tzwe5EKCEX/4oXD4p49Wd489+4aXvQ1gcDj/dmmJnnJ4IrsKxvo7Syura+Ud6sbG3v7O6Z+wdtFaeSMofGIpZdnygmeMQc4CBYN5GMhL5gHX98W/idRyYVj6MHmCTMDckw4gGnBLTkmUf9kMDID7JR7mWO7WU1OLfP8twzq1bdmgIvE3tOqmiOlmd+9QcxTUMWARVEqZ5tJeBmRAKnguWVfqpYQuiYDFlP04iETLnZ9IEcn2plgINY6ooAT9XfExkJlZqEvu4szlWLXiH+5/VSCK7djEdJCiyis0VBKjDEuEgDD7hkFMREE0Il17diOiKSUNCZVXQI9uLLy6TdqNsX9cb9ZbV5M4+jjI7RCaohG12hJrpDLeQginL0jF7Rm/FkvBjvxsestWTMZw7RHxifPyUtlho=
hU1t
AAAB/HicbVDLSsNAFJ34rPUV7dLNYBFclaQKuiy6cVnBtIU2hMl00g6dPJi5EUKIv+LGhSJu/RB3/o2TNgttPTBwOOde7pnjJ4IrsKxvY219Y3Nru7ZT393bPzg0j457Kk4lZQ6NRSwHPlFM8Ig5wEGwQSIZCX3B+v7stvT7j0wqHkcPkCXMDckk4gGnBLTkmY1RSGDqB/m08HLH9nIoCs9sWi1rDrxK7Io0UYWuZ36NxjFNQxYBFUSpoW0l4OZEAqeCFfVRqlhC6IxM2FDTiIRMufk8fIHPtDLGQSz1iwDP1d8bOQmVykJfT5ZR1bJXiv95wxSCazfnUZICi+jiUJAKDDEum8BjLhkFkWlCqOQ6K6ZTIgkF3Vddl2Avf3mV9Not+6LVvr9sdm6qOmroBJ2ic2SjK9RBd6iLHERRhp7RK3oznowX4934WIyuGdVOA/2B8fkDaaaVQw==
cU1t
AAAB/HicbVDLSsNAFJ34rPUV7dLNYBFclaQKuiy6cVnBtIU2hMl00g6dPJi5EUKIv+LGhSJu/RB3/o2TNgttPTBwOOde7pnjJ4IrsKxvY219Y3Nru7ZT393bPzg0j457Kk4lZQ6NRSwHPlFM8Ig5wEGwQSIZCX3B+v7stvT7j0wqHkcPkCXMDckk4gGnBLTkmY1RSGDqBzktvNyxvRyKwjObVsuaA68SuyJNVKHrmV+jcUzTkEVABVFqaFsJuDmRwKlgRX2UKpYQOiMTNtQ0IiFTbj4PX+AzrYxxEEv9IsBz9fdGTkKlstDXk2VUteyV4n/eMIXg2s15lKTAIro4FKQCQ4zLJvCYS0ZBZJoQKrnOiumUSEJB91XXJdjLX14lvXbLvmi17y+bnZuqjho6QafoHNnoCnXQHeoiB1GUoWf0it6MJ+PFeDc+FqNrRrXTQH9gfP4AYeCVPg==
cU1(t−1)
AAACAHicbVBNS8NAEN3Ur1q/oh48eFksQj1YkirosejFYwXTFtoQNttNu3Tzwe5EKCEX/4oXD4p49Wd489+4aXvQ1gcDj/dmmJnnJ4IrsKxvo7Syura+Ud6sbG3v7O6Z+wdtFaeSMofGIpZdnygmeMQc4CBYN5GMhL5gHX98W/idRyYVj6MHmCTMDckw4gGnBLTkmUf9kMDIDzKae5lje1kNzu2zPPfMqlW3psDLxJ6TKpqj5Zlf/UFM05BFQAVRqmdbCbgZkcCpYHmlnyqWEDomQ9bTNCIhU242fSDHp1oZ4CCWuiLAU/X3REZCpSahrzuLc9WiV4j/eb0Ugms341GSAovobFGQCgwxLtLAAy4ZBTHRhFDJ9a2YjogkFHRmFR2CvfjyMmk36vZFvXF/WW3ezOMoo2N0gmrIRleoie5QCzmIohw9o1f0ZjwZL8a78TFrLRnzmUP0B8bnDx1TlhU=
hU1t
AAAB/HicbVDLSsNAFJ34rPUV7dLNYBFclaQKuiy6cVnBtIU2hMl00g6dPJi5EUKIv+LGhSJu/RB3/o2TNgttPTBwOOde7pnjJ4IrsKxvY219Y3Nru7ZT393bPzg0j457Kk4lZQ6NRSwHPlFM8Ig5wEGwQSIZCX3B+v7stvT7j0wqHkcPkCXMDckk4gGnBLTkmY1RSGDqB/m08HLH9nIoCs9sWi1rDrxK7Io0UYWuZ36NxjFNQxYBFUSpoW0l4OZEAqeCFfVRqlhC6IxM2FDTiIRMufk8fIHPtDLGQSz1iwDP1d8bOQmVykJfT5ZR1bJXiv95wxSCazfnUZICi+jiUJAKDDEum8BjLhkFkWlCqOQ6K6ZTIgkF3Vddl2Avf3mV9Not+6LVvr9sdm6qOmroBJ2ic2SjK9RBd6iLHERRhp7RK3oznowX4934WIyuGdVOA/2B8fkDaaaVQw==
cU2(t−1)
AAACAHicbVBNS8NAEJ3Ur1q/oh48eAkWoR4sSRX0WPTisYJpC20Im+2mXbrZhN2NUEIu/hUvHhTx6s/w5r9x0/agrQ8GHu/NMDMvSBiVyra/jdLK6tr6RnmzsrW9s7tn7h+0ZZwKTFwcs1h0AyQJo5y4iipGuokgKAoY6QTj28LvPBIhacwf1CQhXoSGnIYUI6Ul3zzqR0iNgjDDuZ+5DT+rqXPnLM99s2rX7SmsZeLMSRXmaPnmV38Q4zQiXGGGpOw5dqK8DAlFMSN5pZ9KkiA8RkPS05SjiEgvmz6QW6daGVhhLHRxZU3V3xMZiqScRIHuLM6Vi14h/uf1UhVeexnlSaoIx7NFYcosFVtFGtaACoIVm2iCsKD6VguPkEBY6cwqOgRn8eVl0m7UnYt64/6y2ryZx1GGYziBGjhwBU24gxa4gCGHZ3iFN+PJeDHejY9Za8mYzxzCHxifPx7glhY=
hU2(t−1)
AAACAHicbVBNS8NAEJ3Ur1q/oh48eAkWoR4sSRX0WPTisYJpC20Im+2mXbrZhN2NUEIu/hUvHhTx6s/w5r9x0/agrQ8GHu/NMDMvSBiVyra/jdLK6tr6RnmzsrW9s7tn7h+0ZZwKTFwcs1h0AyQJo5y4iipGuokgKAoY6QTj28LvPBIhacwf1CQhXoSGnIYUI6Ul3zzqR0iNgjAb5X7mNvysps6dszz3zapdt6ewlokzJ1WYo+WbX/1BjNOIcIUZkrLn2InyMiQUxYzklX4qSYLwGA1JT1OOIiK9bPpAbp1qZWCFsdDFlTVVf09kKJJyEgW6szhXLnqF+J/XS1V47WWUJ6kiHM8WhSmzVGwVaVgDKghWbKIJwoLqWy08QgJhpTOr6BCcxZeXSbtRdy7qjfvLavNmHkcZjuEEauDAFTThDlrgAoYcnuEV3own48V4Nz5mrSVjPnMIf2B8/gAmupYb
hU2t
AAAB/HicbVBNS8NAFNz4WetXtEcvwSJ4KkkV9Fj04rGCaQttCJvtpl262YTdFyGE+Fe8eFDEqz/Em//GTZuDtg4sDDPv8WYnSDhTYNvfxtr6xubWdm2nvru3f3BoHh33VJxKQl0S81gOAqwoZ4K6wIDTQSIpjgJO+8HstvT7j1QqFosHyBLqRXgiWMgIBi35ZmMUYZgGYT4t/Nxt+zkUhW827ZY9h7VKnIo0UYWub36NxjFJIyqAcKzU0LET8HIsgRFOi/ooVTTBZIYndKipwBFVXj4PX1hnWhlbYSz1E2DN1d8bOY6UyqJAT5ZR1bJXiv95wxTCay9nIkmBCrI4FKbcgtgqm7DGTFICPNMEE8l0VotMscQEdF91XYKz/OVV0mu3nItW+/6y2bmp6qihE3SKzpGDrlAH3aEuchFBGXpGr+jNeDJejHfjYzG6ZlQ7DfQHxucPay+VRA==
cU2t
AAAB/HicbVDLSsNAFJ34rPUV7dLNYBFclaQKuiy6cVnBtIU2hMl00g6dPJi5EUKIv+LGhSJu/RB3/o2TNgttPTBwOOde7pnjJ4IrsKxvY219Y3Nru7ZT393bPzg0j457Kk4lZQ6NRSwHPlFM8Ig5wEGwQSIZCX3B+v7stvT7j0wqHkcPkCXMDckk4gGnBLTkmY1RSGDqBzktvNxpezkUhWc2rZY1B14ldkWaqELXM79G45imIYuACqLU0LYScHMigVPBivooVSwhdEYmbKhpREKm3HwevsBnWhnjIJb6RYDn6u+NnIRKZaGvJ8uoatkrxf+8YQrBtZvzKEmBRXRxKEgFhhiXTeAxl4yCyDQhVHKdFdMpkYSC7quuS7CXv7xKeu2WfdFq3182OzdVHTV0gk7RObLRFeqgO9RFDqIoQ8/oFb0ZT8aL8W58LEbXjGqngf7A+PwBY2mVPw==
hU2t
AAAB/HicbVBNS8NAFNz4WetXtEcvwSJ4KkkV9Fj04rGCaQttCJvtpl262YTdFyGE+Fe8eFDEqz/Em//GTZuDtg4sDDPv8WYnSDhTYNvfxtr6xubWdm2nvru3f3BoHh33VJxKQl0S81gOAqwoZ4K6wIDTQSIpjgJO+8HstvT7j1QqFosHyBLqRXgiWMgIBi35ZmMUYZgGYT4t/Nxt+zkUhW827ZY9h7VKnIo0UYWub36NxjFJIyqAcKzU0LET8HIsgRFOi/ooVTTBZIYndKipwBFVXj4PX1hnWhlbYSz1E2DN1d8bOY6UyqJAT5ZR1bJXiv95wxTCay9nIkmBCrI4FKbcgtgqm7DGTFICPNMEE8l0VotMscQEdF91XYKz/OVV0mu3nItW+/6y2bmp6qihE3SKzpGDrlAH3aEuchFBGXpGr+jNeDJejHfjYzG6ZlQ7DfQHxucPay+VRA==
cU3t
AAAB/HicbVBNS8NAFNzUr1q/oj16CRbBU0laQY9FLx4rmLbQlrDZbtqlm03YfRFCiH/FiwdFvPpDvPlv3LQ5aOvAwjDzHm92/JgzBbb9bVQ2Nre2d6q7tb39g8Mj8/ikp6JEEuqSiEdy4GNFORPUBQacDmJJcehz2vfnt4Xff6RSsUg8QBrTcYinggWMYNCSZ9ZHIYaZH2Qk9zK37WWQ557ZsJv2AtY6cUrSQCW6nvk1mkQkCakAwrFSQ8eOYZxhCYxwmtdGiaIxJnM8pUNNBQ6pGmeL8Ll1rpWJFURSPwHWQv29keFQqTT09WQRVa16hfifN0wguB5nTMQJUEGWh4KEWxBZRRPWhElKgKeaYCKZzmqRGZaYgO6rpktwVr+8TnqtptNutu4vG52bso4qOkVn6AI56Ap10B3qIhcRlKJn9IrejCfjxXg3PpajFaPcqaM/MD5/AGTylUA=
hU3t
AAAB/HicbVBNS8NAFHypX7V+RXv0EiyCp5K0gh6LXjxWMG2hLWGz3bRLN5uwuxFCiH/FiwdFvPpDvPlv3LQ5aOvAwjDzHm92/JhRqWz726hsbG5t71R3a3v7B4dH5vFJT0aJwMTFEYvEwEeSMMqJq6hiZBALgkKfkb4/vy38/iMRkkb8QaUxGYdoymlAMVJa8sz6KERq5gfZLPcyt+1lKs89s2E37QWsdeKUpAElup75NZpEOAkJV5ghKYeOHatxhoSimJG8NkokiRGeoykZaspRSOQ4W4TPrXOtTKwgEvpxZS3U3xsZCqVMQ19PFlHlqleI/3nDRAXX44zyOFGE4+WhIGGWiqyiCWtCBcGKpZogLKjOauEZEggr3VdNl+Csfnmd9FpNp91s3V82OjdlHVU4hTO4AAeuoAN30AUXMKTwDK/wZjwZL8a78bEcrRjlTh3+wPj8AWy4lUU=
hU3t
AAAB/HicbVBNS8NAFHypX7V+RXv0EiyCp5K0gh6LXjxWMG2hLWGz3bRLN5uwuxFCiH/FiwdFvPpDvPlv3LQ5aOvAwjDzHm92/JhRqWz726hsbG5t71R3a3v7B4dH5vFJT0aJwMTFEYvEwEeSMMqJq6hiZBALgkKfkb4/vy38/iMRkkb8QaUxGYdoymlAMVJa8sz6KERq5gfZLPcyt+1lKs89s2E37QWsdeKUpAElup75NZpEOAkJV5ghKYeOHatxhoSimJG8NkokiRGeoykZaspRSOQ4W4TPrXOtTKwgEvpxZS3U3xsZCqVMQ19PFlHlqleI/3nDRAXX44zyOFGE4+WhIGGWiqyiCWtCBcGKpZogLKjOauEZEggr3VdNl+Csfnmd9FpNp91s3V82OjdlHVU4hTO4AAeuoAN30AUXMKTwDK/wZjwZL8a78bEcrRjlTh3+wPj8AWy4lUU=
hU3(t−1)
AAACAHicbVBNS8NAEJ34WetX1IMHL4tFqAdL0gp6LHrxWMG0hbaEzXbTLt18sLsRSsjFv+LFgyJe/Rne/Ddu2hy09cHA470ZZuZ5MWdSWda3sbK6tr6xWdoqb+/s7u2bB4dtGSWCUIdEPBJdD0vKWUgdxRSn3VhQHHicdrzJbe53HqmQLAof1DSmgwCPQuYzgpWWXPO4H2A19vx0nLmp03DTqrqwz7PMNStWzZoBLRO7IBUo0HLNr/4wIklAQ0U4lrJnW7EapFgoRjjNyv1E0hiTCR7RnqYhDqgcpLMHMnSmlSHyI6ErVGim/p5IcSDlNPB0Z36uXPRy8T+vlyj/epCyME4UDcl8kZ9wpCKUp4GGTFCi+FQTTATTtyIyxgITpTMr6xDsxZeXSbtesxu1+v1lpXlTxFGCEziFKthwBU24gxY4QCCDZ3iFN+PJeDHejY9564pRzBzBHxifPyhHlhw=
cU3(t−1)
AAACAHicbVBNS8NAEJ34WetX1IMHL8Ei1IMlaQU9Fr14rGDaQhvCZrtpl242YXcjlJCLf8WLB0W8+jO8+W/ctD1o64OBx3szzMwLEkalsu1vY2V1bX1js7RV3t7Z3ds3Dw7bMk4FJi6OWSy6AZKEUU5cRRUj3UQQFAWMdILxbeF3HomQNOYPapIQL0JDTkOKkdKSbx73I6RGQZjh3M/chp9V1YVznue+WbFr9hTWMnHmpAJztHzzqz+IcRoRrjBDUvYcO1FehoSimJG83E8lSRAeoyHpacpRRKSXTR/IrTOtDKwwFrq4sqbq74kMRVJOokB3FufKRa8Q//N6qQqvvYzyJFWE49miMGWWiq0iDWtABcGKTTRBWFB9q4VHSCCsdGZlHYKz+PIyaddrTqNWv7+sNG/mcZTgBE6hCg5cQRPuoAUuYMjhGV7hzXgyXox342PWumLMZ47gD4zPHyBtlhc=
...
AAAB7XicbVBNS8NAEJ3Ur1q/qh69BIvgqSRV0GPRi8cKthbaUDabTbt2sxt2J4VS+h+8eFDEq//Hm//GbZuDtj4YeLw3w8y8MBXcoOd9O4W19Y3NreJ2aWd3b/+gfHjUMirTlDWpEkq3Q2KY4JI1kaNg7VQzkoSCPYbD25n/OGLacCUfcJyyICF9yWNOCVqp1R1FCk2vXPGq3hzuKvFzUoEcjV75qxspmiVMIhXEmI7vpRhMiEZOBZuWuplhKaFD0mcdSyVJmAkm82un7plVIjdW2pZEd67+npiQxJhxEtrOhODALHsz8T+vk2F8HUy4TDNkki4WxZlwUbmz192Ia0ZRjC0hVHN7q0sHRBOKNqCSDcFffnmVtGpV/6Jau7+s1G/yOIpwAqdwDj5cQR3uoAFNoPAEz/AKb45yXpx352PRWnDymWP4A+fzB8y9j0Y=
Figure 5.3: The architecture of an LSTM multi-layer [80].
For simplicity, an LSTM multi-layer network can be encapsulated by a function
S, where
...
(hU3t , cU3t) = LU3(hU3(t−1)
, cU3(t−1)
,hU2t),
(hU2t , cU2t) = LU2(hU2(t−1)
, cU2(t−1)
,hU1t),
(hU1t , cU1t) = LU1(hU1(t−1)
, cU1(t−1)
,xt),
can be given by
(Ht,Ct) = S(Ht 1,Ct 1,xt). (5.12)
It is important to mention that each layer can have a different number of LSTM
units U1, U2, U3 . . . , so that the dimensions of the hidden and cell states can be
uneven for each LSTM layer. At time step t, the variables Ht and Ct represent the
collection of all hidden and cell states, respectively [78]. As can be seen, the LSTM
multi-layer network is a larger model, containing a lot of stacked LSTM cells. For this
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 61
reason, the single and multiple layer LSTM networks can be employed in time-series
forecasting [77,78,80], which is the aim of this thesis.
D. The Architecture of the Proposed Deep Learning Network
Figure 5.4 describes in detail the proposed LSTM layer where each input dataset
obtained from ITU-R P.1853 [38], A0
i = (A0(1)
t,i ,A
0(2)
t,i , . . . ,A
0(m)
t,i ) for i = 1, 2, . . . , N ,
and t = 1, 2, . . . , K, is made up of m examples, A0(m)
t,i 2 R
K , where K is the number
of samples in each time-series, and N is the number of GW-sites on the ground
segment. Here, each rain-attenuation time-series must be normalized to improve the
convergence of the gradient descendent, for this reason, the input dataset of rain
attenuation is denoted by A
0
i. Indeed, the normalization method is discussed in more
detail in Section 5.2.2. In this thesis, each time series is a univariate sequence, that is,
it is a single series of observations. Furthermore, each time series has a single feature,
C = 1, and each dataset has 40 examples, i.e., m = {1, 2, . . . , 40}. Therefore, it is
a many-to-one sequence problem with a single feature. Motivated by this, an LSTM
network model is built in order to learn from the time series of past observations to
predict the next value in the sequence.
Number of
Hidden Units
c
t−1
h
t−1
c
t
h
t
Initial
State
Final
State
LSTM
Layer
……LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
LU
AAAB9HicbVBNS8NAFHypX7V+RT16WSyCp5JUQY9FLx48VDBtoQ1ls920SzebuLsplNDf4cWDIl79Md78N27aHLR1YGGYeY83O0HCmdKO822V1tY3NrfK25Wd3b39A/vwqKXiVBLqkZjHshNgRTkT1NNMc9pJJMVRwGk7GN/mfntCpWKxeNTThPoRHgoWMoK1kfxehPWIYJ7dz/pe3646NWcOtErcglShQLNvf/UGMUkjKjThWKmu6yTaz7DUjHA6q/RSRRNMxnhIu4YKHFHlZ/PQM3RmlAEKY2me0Giu/t7IcKTUNArMZB5SLXu5+J/XTXV47WdMJKmmgiwOhSlHOkZ5A2jAJCWaTw3BRDKTFZERlpho01PFlOAuf3mVtOo196JWf7isNm6KOspwAqdwDi5cQQPuoAkeEHiCZ3iFN2tivVjv1sditGQVO8fwB9bnD+Z3ki4=
A (m)
1,i
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
A (m)
2,i
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
A (m)
t,i
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
A (m)
K,i
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
h11
h21
...
hU1
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
h12
h22
.
.
.
hU2
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
h1t
h2t
.
.
.
hUt
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
h1K
h2K
.
.
.
hUK
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
… …
… …
A
(m)
t,i ∈ R
K , for
i = 1, 2, . . . , N
t = 1, 2, . . . ,K
C = 1
m = 1, 2 . . . , 40
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
Figure 5.4: The architecture of the proposed LSTM layer.
The LSTM network can be made up of a single layer, LU , i.e, S = {LU} with
hidden units, U , or multi-layer, S = {LU1,LU2,LU3, . . .} with different number of
hidden units, U1, U2, U3, . . ., in each layer. For a single LSTM layer, the hidden and
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 62
cell states, with m examples and LSTM units U , are given by
(ht, ct) = LU(ht 1, ct 1,A
0(m)
t,i ), for
8
>
<
>
:
i = 1, 2, . . . , N
t = 1, 2, . . . , K
m = 1, 2, . . . , 40
(5.13)
Meanwhile, the output function of the LSTM multi-layer network, S, is denoted
by
(Ht,Ct) = S(Ht 1,Ct 1,A
0(m)
t,i ), for
8
>
<
>
:
i = 1, 2, . . . , N
t = 1, 2, . . . , K
m = 1, 2, . . . , 40
(5.14)
As a result, the LSTM network presents K hidden states at the output,
h1,h2, . . . ,ht, . . . ,hK 2 ±U
U , increasing the network complexity by stacking several
hidden outputs. However, it is important to find a single sequence at the output. For
this purpose, dense layers can reduce from multiple inputs to a single output by an
activation function either linear or non-linear. Figure 5.5 illustrates the proposed deep
learning network, which connects the proposed LSTM network to the dense layer. To
be precise, the obtained output is the predicted rain attenuation time-series, at the
time step t+ 1, Â0
t+1,i 2 R
K .
S
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy4r2gdMh5JJM21oJhmSjFCGfoYbF4q49Wvc+Tdm2llo64HA4Zx7ybknTDjTxnW/ndLa+sbmVnm7srO7t39QPTzqaJkqQttEcql6IdaUM0HbhhlOe4miOA457YaT29zvPlGlmRSPZprQIMYjwSJGsLGS34+xGRPMs4fZoFpz6+4caJV4BalBgdag+tUfSpLGVBjCsda+5yYmyLAyjHA6q/RTTRNMJnhEfUsFjqkOsnnkGTqzyhBFUtknDJqrvzcyHGs9jUM7mUfUy14u/uf5qYmug4yJJDVUkMVHUcqRkSi/Hw2ZosTwqSWYKGazIjLGChNjW6rYErzlk1dJp1H3LuqN+8ta86aoowwncArn4MEVNOEOWtAGAhKe4RXeHOO8OO/Ox2K05BQ7x/AHzucPjPaRbQ==
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
L
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFNy5cVLAPmA4lk2ba0EwyJBmhDP0MNy4UcevXuPNvzLSz0NYDgcM595JzT5hwpo3rfjultfWNza3ydmVnd2//oHp41NEyVYS2ieRS9UKsKWeCtg0znPYSRXEcctoNJ7e5332iSjMpHs00oUGMR4JFjGBjJb8fYzMmmGf3s0G15tbdOdAq8QpSgwKtQfWrP5QkjakwhGOtfc9NTJBhZRjhdFbpp5ommEzwiPqWChxTHWTzyDN0ZpUhiqSyTxg0V39vZDjWehqHdjKPqJe9XPzP81MTXQcZE0lqqCCLj6KUIyNRfj8aMkWJ4VNLMFHMZkVkjBUmxrZUsSV4yyevkk6j7l3UGw+XteZNUUcZTuAUzsGDK2jCHbSgDQQkPMMrvDnGeXHenY/FaMkpdo7hD5zPH4JTkWY=
D
en
se
L
ay
er
AAAB+3icbVDLSgNBEJyNrxhfazx6GQyCp7AbBT0G9eDBQwTzgCSE2UknGTI7u8z0SsKSX/HiQRGv/og3/8bJ46CJBQ1FVTfdXUEshUHP+3Yya+sbm1vZ7dzO7t7+gXuYr5ko0RyqPJKRbgTMgBQKqihQQiPWwMJAQj0Y3kz9+hNoIyL1iOMY2iHrK9ETnKGVOm6+hTDC9BaUAXrPxqAnHbfgFb0Z6CrxF6RAFqh03K9WN+JJCAq5ZMY0fS/Gdso0Ci5hkmslBmLGh6wPTUsVC8G009ntE3pqlS7tRdqWQjpTf0+kLDRmHAa2M2Q4MMveVPzPaybYu2qnQsUJguLzRb1EUozoNAjaFRo4yrEljGthb6V8wDTjaOPK2RD85ZdXSa1U9M+LpYeLQvl6EUeWHJMTckZ8cknK5I5USJVwMiLP5JW8ORPnxXl3PuatGWcxc0T+wPn8AS/QlIg=
···
AAAB7XicbVBNS8NAEJ3Ur1q/qh69BIvgqSRV0GPRi8cK9gPaUDabTbt2sxt2J0Ip/Q9ePCji1f/jzX/jts1BWx8MPN6bYWZemApu0PO+ncLa+sbmVnG7tLO7t39QPjxqGZVpyppUCaU7ITFMcMmayFGwTqoZSULB2uHodua3n5g2XMkHHKcsSMhA8phTglZq9Wik0PTLFa/qzeGuEj8nFcjR6Je/epGiWcIkUkGM6fpeisGEaORUsGmplxmWEjoiA9a1VJKEmWAyv3bqnlklcmOlbUl05+rviQlJjBknoe1MCA7NsjcT//O6GcbXwYTLNEMm6WJRnAkXlTt73Y24ZhTF2BJCNbe3unRINKFoAyrZEPzll1dJq1b1L6q1+8tK/SaPowgncArn4MMV1OEOGtAECo/wDK/w5ijnxXl3PhatBSefOYY/cD5/AK+ljzM=
···
AAAB7XicbVBNS8NAEJ3Ur1q/qh69BIvgqSRV0GPRi8cK9gPaUDabTbt2sxt2J0Ip/Q9ePCji1f/jzX/jts1BWx8MPN6bYWZemApu0PO+ncLa+sbmVnG7tLO7t39QPjxqGZVpyppUCaU7ITFMcMmayFGwTqoZSULB2uHodua3n5g2XMkHHKcsSMhA8phTglZq9Wik0PTLFa/qzeGuEj8nFcjR6Je/epGiWcIkUkGM6fpeisGEaORUsGmplxmWEjoiA9a1VJKEmWAyv3bqnlklcmOlbUl05+rviQlJjBknoe1MCA7NsjcT//O6GcbXwYTLNEMm6WJRnAkXlTt73Y24ZhTF2BJCNbe3unRINKFoAyrZEPzll1dJq1b1L6q1+8tK/SaPowgncArn4MMV1OEOGtAECo/wDK/w5ijnxXl3PhatBSefOYY/cD5/AK+ljzM=
···
AAAB7XicbVBNS8NAEJ3Ur1q/qh69BIvgqSRV0GPRi8cK9gPaUDabTbt2sxt2J0Ip/Q9ePCji1f/jzX/jts1BWx8MPN6bYWZemApu0PO+ncLa+sbmVnG7tLO7t39QPjxqGZVpyppUCaU7ITFMcMmayFGwTqoZSULB2uHodua3n5g2XMkHHKcsSMhA8phTglZq9Wik0PTLFa/qzeGuEj8nFcjR6Je/epGiWcIkUkGM6fpeisGEaORUsGmplxmWEjoiA9a1VJKEmWAyv3bqnlklcmOlbUl05+rviQlJjBknoe1MCA7NsjcT//O6GcbXwYTLNEMm6WJRnAkXlTt73Y24ZhTF2BJCNbe3unRINKFoAyrZEPzll1dJq1b1L6q1+8tK/SaPowgncArn4MMV1OEOGtAECo/wDK/w5ijnxXl3PhatBSefOYY/cD5/AK+ljzM=
D
AAAB8nicbVDLSgMxFL1TX7W+qi7dBIvgqsxUQZdFXbisYB8wHUomzbShmWRIMkIZ+hluXCji1q9x59+YaWehrQcCh3PuJeeeMOFMG9f9dkpr6xubW+Xtys7u3v5B9fCoo2WqCG0TyaXqhVhTzgRtG2Y47SWK4jjktBtObnO/+0SVZlI8mmlCgxiPBIsYwcZKfj/GZkwwz+5mg2rNrbtzoFXiFaQGBVqD6ld/KEkaU2EIx1r7npuYIMPKMMLprNJPNU0wmeAR9S0VOKY6yOaRZ+jMKkMUSWWfMGiu/t7IcKz1NA7tZB5RL3u5+J/npya6DjImktRQQRYfRSlHRqL8fjRkihLDp5ZgopjNisgYK0yMbaliS/CWT14lnUbdu6g3Hi5rzZuijjKcwCmcgwdX0IR7aEEbCEh4hld4c4zz4rw7H4vRklPsHMMfOJ8/diuRXg==
···
AAAB7XicbVBNS8NAEJ3Ur1q/qh69BIvgqSRV0GPRi8cK9gPaUDabTbt2sxt2J0Ip/Q9ePCji1f/jzX/jts1BWx8MPN6bYWZemApu0PO+ncLa+sbmVnG7tLO7t39QPjxqGZVpyppUCaU7ITFMcMmayFGwTqoZSULB2uHodua3n5g2XMkHHKcsSMhA8phTglZq9Wik0PTLFa/qzeGuEj8nFcjR6Je/epGiWcIkUkGM6fpeisGEaORUsGmplxmWEjoiA9a1VJKEmWAyv3bqnlklcmOlbUl05+rviQlJjBknoe1MCA7NsjcT//O6GcbXwYTLNEMm6WJRnAkXlTt73Y24ZhTF2BJCNbe3unRINKFoAyrZEPzll1dJq1b1L6q1+8tK/SaPowgncArn4MMV1OEOGtAECo/wDK/w5ijnxXl3PhatBSefOYY/cD5/AK+ljzM=
h1 ∈ ±U
U
AAACCXicbVBNS8NAEN3Ur1q/oh69LBbBU0mqoMeiF48VTFtoYthsN+3S3U3Y3Qgl9OrFv+LFgyJe/Qfe/Ddu2hy09cHA470ZZuZFKaNKO863VVlZXVvfqG7WtrZ3dvfs/YOOSjKJiYcTlshehBRhVBBPU81IL5UE8YiRbjS+LvzuA5GKJuJOT1IScDQUNKYYaSOFNvQ50qMozkfT0PWp8FM+V6Lcm96bCu2603BmgMvELUkdlGiH9pc/SHDGidCYIaX6rpPqIEdSU8zItOZniqQIj9GQ9A0ViBMV5LNPpvDEKAMYJ9KU0HCm/p7IEVdqwiPTWVypFr1C/M/rZzq+DHIq0kwTgeeL4oxBncAiFjigkmDNJoYgLKm5FeIRkghrE17NhOAuvrxMOs2Ge9Zo3p7XW1dlHFVwBI7BKXDBBWiBG9AGHsDgETyDV/BmPVkv1rv1MW+tWOXMIfgD6/MHhBWa2w==
h2 ∈ ±U
U
AAACCXicbVBNS8NAEN3Ur1q/oh69LBbBU0mqoMeiF48VTFtoYthsN+3S3U3Y3Qgl9OrFv+LFgyJe/Qfe/Ddu2hy09cHA470ZZuZFKaNKO863VVlZXVvfqG7WtrZ3dvfs/YOOSjKJiYcTlshehBRhVBBPU81IL5UE8YiRbjS+LvzuA5GKJuJOT1IScDQUNKYYaSOFNvQ50qMozkfTsOlT4ad8rkS5N703Fdp1p+HMAJeJW5I6KNEO7S9/kOCME6ExQ0r1XSfVQY6kppiRac3PFEkRHqMh6RsqECcqyGefTOGJUQYwTqQpoeFM/T2RI67UhEems7hSLXqF+J/Xz3R8GeRUpJkmAs8XxRmDOoFFLHBAJcGaTQxBWFJzK8QjJBHWJryaCcFdfHmZdJoN96zRvD2vt67KOKrgCByDU+CCC9ACN6ANPIDBI3gGr+DNerJerHfrY95ascqZQ/AH1ucPha2a3A==
ht ∈ ±U
U
AAACCXicbVBNS8NAEN3Ur1q/oh69LBbBU0mqoMeiF48VTFtoYthsN+3S3U3Y3Qgl9OrFv+LFgyJe/Qfe/Ddu2hy09cHA470ZZuZFKaNKO863VVlZXVvfqG7WtrZ3dvfs/YOOSjKJiYcTlshehBRhVBBPU81IL5UE8YiRbjS+LvzuA5GKJuJOT1IScDQUNKYYaSOFNvQ50qMozkfTUPtU+CmfK1HuTe9NhXbdaTgzwGXilqQOSrRD+8sfJDjjRGjMkFJ910l1kCOpKWZkWvMzRVKEx2hI+oYKxIkK8tknU3hilAGME2lKaDhTf0/kiCs14ZHpLK5Ui14h/uf1Mx1fBjkVaaaJwPNFccagTmARCxxQSbBmE0MQltTcCvEISYS1Ca9mQnAXX14mnWbDPWs0b8/rrasyjio4AsfgFLjgArTADWgDD2DwCJ7BK3iznqwX6936mLdWrHLmEPyB9fkD7t2bHg==
hK ∈ ±U
U
AAACCXicbVBNS8NAEN34WetX1KOXxSJ4KkkV9Fj0InipYNpCE8Nmu2mX7m7C7kYooVcv/hUvHhTx6j/w5r9x0+agrQ8GHu/NMDMvShlV2nG+raXlldW19cpGdXNre2fX3ttvqySTmHg4YYnsRkgRRgXxNNWMdFNJEI8Y6USjq8LvPBCpaCLu9DglAUcDQWOKkTZSaEOfIz2M4nw4CW98KvyUz5Qo9yb3pkK75tSdKeAicUtSAyVaof3l9xOccSI0ZkipnuukOsiR1BQzMqn6mSIpwiM0ID1DBeJEBfn0kwk8Nkofxok0JTScqr8ncsSVGvPIdBZXqnmvEP/zepmOL4KcijTTRODZojhjUCewiAX2qSRYs7EhCEtqboV4iCTC2oRXNSG48y8vknaj7p7WG7dnteZlGUcFHIIjcAJccA6a4Bq0gAcweATP4BW8WU/Wi/Vufcxal6xy5gD8gfX5A62FmvU=
A
(m)
t,i ∈ R
K , for
i = 1, 2, . . . , N
t = 1, 2, . . . ,K
C = 1
m = 1, 2 . . . , 40
AAACcHicbZFNb9QwEIad8FXC10IvSIAwrBAtilbJUgkulQq9IFVCBbFtpc2ycpzJ1qrtRPYEsbJy5v9x40dw4Rfg7OZQWkay9eoZj2f8Oq+lsJgkv4LwytVr129s3Ixu3b5z997g/oMjWzWGw4RXsjInObMghYYJCpRwUhtgKpdwnJ/td/njb2CsqPQXXNYwU2yhRSk4Q4/mgx+ZYnial869e9m2c4exaL+6LbXdZkKvc7n77NFBG9MM4Tu6sjJtlOWwENpx39q2kdhN43GcyaJCG3/MsgjPgwMP9ndTv6sO93QniTLQRX/FfDBMRskq6GWR9mJI+jicD35mRcUbBRq5ZNZO06TGmWMGBZfgB2ws1IyfsQVMvdRMgZ25lWEtfeFJQf1D/NJIV/R8hWPK2qXK/cnOAnsx18H/5aYNlm9nTui6QdB83ahsJMWKdu7TQhjgKJdeMG6En5XyU2YYR/9HkTchvfjky+JoPEpfj8afdoZ773s7Nsgj8pxskZS8IXvkAzkkE8LJ72AzeBw8Cf6ED8On4bP10TDoazbJPxG++gsRkbpO
Â
t+1,i ∈ R
K
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
Figure 5.5: The architecture of the proposed deep learning network.
For readability and simplicity, the proposed LSTM network is analyzed as a single
LSTM layer, LU , where each output, also known as the hidden state at time step t,
is a U -dimensional vector, ht 2 ±U
U . The LSTM outputs become the inputs of the
dense layer. In this context, the dense layer is a deeply connected neural network,
where each connection can be typically considered as a multiple-input neuron [81,82].
Usually, the neuron’s output can be denoted by y = f(W · p + b), where W is
the weight matrix, b is the bias, f is the activation function, and p is the input
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 63
vector [81–83]. In this thesis, the dense layer can be mathematically expressed as
Â0
2,i = selu(wd
T
1 · h1 + bd1) 2 R
Â0
3,i = selu(wd
T
2 · h2 + bd2) 2 R
...
Â0
t+1,i = selu(wd
T
t · ht + bdt) 2 R
...
Â0
K+1,i = selu(wd
T
K · hK + bdK ) 2 R
where each value, Â0
t+1,i, is stored in the output vector. Therefore, the output of the
dense layer is encapsulated and rewritten as
Â0
t+1,i = D(ht) 2 R
K , for
(
i = 1, 2, . . . , N,
t = 1, 2 . . . , K,
(5.15)
where D is the dense layer function and ht 2 ±U
U for t = 1, 2, . . . , K, are the
input vectors of the dense layer. The scaled-exponential-linear-unit (SELU) activation
function is employed in this layer, which is discussed in more detail in Section 5.2.2.
Moreover, Wd 2 R
U⇥K is the weight matrix that contains each weight vector wdt 2
R
U and bd 2 R
K is the bias vector that includes each bias element bdt 2 R. Similarly
to the LSTM network, the weight parameters of the dense layer are also learned
during the training process. Furthermore, the weight parameters of the dense layer
can be combined under one symbol ΘD [79], which is given by
ΘD ⌘ {Wd,bd}. (5.16)
Finally, the output vectors obtained from the proposed deep learning network can
be stacked in the matrix, Â0 2 R
K⇥N , whose predicted values of rain attenuation
for each GW’s feeder uplink are contained within the matrix. To be specific, these
values allow the Network Control Center (NCC) to know the feeder uplink status in
advance, managing resources, e.g. gateway redundancy, to avoid link outages due to
the heavy rain.
5.2.2 Deep Learning Network: Model Description
A. Activation Functions
The activation function is used to determine the nature of the output of a neural
network, in addition to limiting the amplitude of the output of a neuron. That is,
it can map the outcome values between 0 to 1, -1 to 1, yes or no, etc., depending
upon the type of activation function. For this reason, the activation function can be
divided into two types, linear and non-linear activation functions. Generally, the most
frequent activation functions are non-linear functions because they can facilitate the
generalization and adaptation of the model with a lot of varied data. The activation
functions are also known as transfer functions and/or squashing functions [83, 84].
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 64
Usually, the most employed activation functions are in the following.
• Sigmoid Activation Function: the sigmoid function, also known as logistic
activation function, is especially used for models that are able to predict the
probability at the output. For this purpose, the sigmoid function always returns
a value between 0–1, therefore, it is ideal for probabilistic outputs [85]. The
sigmoid activation function, (z), is given by
(z) =
1
1 + exp( z) (5.17)
Figure 5.6 depicts the sigmoid function, where, for small values (z < 5), the
results of the sigmoid function get close to zero, and for large values (z > 5)
the sigmoid function returns values close to 1.
σ
(z
)
0.00
0.25
0.50
0.75
1.00
z
-5 -4 -3 -2 -1 0 1 2 3 4 5
Figure 5.6: Sigmoid activation function.
The sigmoid function is differentiable, therefore, the slope of the sigmoid curve
can be found at any two points. However, it is important to mention that the
sigmoid function could provoke a jam at the training process [83]. Also, it is
possible to assume that the sigmoid function is similar to a 2-element Softmax,
where the second element is zero. Indeed, the softmax activation function for
multiclass classification problems is based on the sigmoid function but with
n-elements [85].
• Hyperbolic Tangent Activation Function: the hyperbolic tangent
activation function (tanh) is similar to sigmoid function but with a wider range,
which varies from -1 to 1. In this case, the negative inputs are directly mapped
in the negative range whereas the zero inputs are mapped close to zero in the
tanh function. Naturally, this is the main advantage of the tanh activation
function. The tanh function is expressed as
tanh(z) =
sinh(z)
cosh(z)
=
1
1 + exp( 2z) 1 (5.18)
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 65
Figure 5.7 illustrates the tanh function, where it is possible to distinguish the
range between -1 to 1.
ta
n
h
(z
)
-1.00
-0.50
0.00
0.50
1.00
z
-5 -4 -3 -2 -1 0 1 2 3 4 5
Figure 5.7: Hyperbolic tangent activation function.
It is important to mention that the tanh function is differentiable, in addition
to being mainly used in classification problems, especially, between two classes.
Both sigmoid and tanh activation functions are used in feed-forward neural
networks [83,85].
• Rectifier-Linear-Unit Activation Function: it is also known as ReLU,
which is one of the most popular activation functions to use in machine learning
algorithms. This activation function is especially employed in deep learning and
convolutional neural networks [83]. The ReLU activation function expression is
given by
R(z) = max(0, z) =
(
0 for z < 0
z for z 0
(5.19)
Figure 5.8 shows the ReLU activation function, which ranges from 0 to 1. In
other words, the ReLU is half rectified, i.e., R(z) = 0 when z is less than zero
(z < 0) and R(z) = z when z is above or equal to zero (z 0) [85].
However, this activation function can become a problem if there are negative
input values, which turn into zero values. In a sense, this issue can affect the
ability of the model to train or fit from the data. For this reason, the input
data must be previously analyzed to apply the proper activation function.
• Scaled-Exponential-Linear-Unit Activation Function: it is also known
as SELU activation function, which is defined as
selu(z) = selu ·
(
z for z 0
↵selu[exp(z) 1] for z < 0
, (5.20)
where selu and ↵selu are pre-defined constants ( selu = 1.05070098 and ↵selu =
1.67326324) [85]. For the sake of simplicity, the SELU activation function
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 66
R
(z
)
0.00
1.00
2.00
3.00
4.00
5.00
z
-5 -4 -3 -2 -1 0 1 2 3 4 5
Figure 5.8: Rectifier-Linear-Unit activation function.
multiplies scale (↵selu > 1) with the exponential and the linear functions to
assure a slope larger than 1 for positive inputs.
Figure 5.9 depicts the SELU activation function. The values of selu and ↵selu
are chosen so that the mean and variance of the inputs are preserved between
two consecutive layers as long as the weights are initialized correctly, in addition
to the number of input units is large enough [85,86].
se
lu
(z
)
-2.00
0.00
2.00
4.00
6.00
z
-5 -4 -3 -2 -1 0 1 2 3 4 5
Figure 5.9: Scaled-Exponential-Linear-Unit activation function.
Self-normalizing neural networks (SNN) use the SELU activation function,
which is able to enable high-level abstract representations. SELU activation
functions induce self-normalizing properties, improving the convergence and
accuracy, in addition to being widely used in machine learning and deep learning
models. [86].
To sum up, the choice of the best activation function depends on the type and nature
of input data, which translates into better convergence, accuracy, and generalization
of the neural network model. Therefore, each activation function must be rigorously
tested to find a strong and functional model.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 67
B. Feature Scaling
To begin with, the rain attenuation time-series for each i-th GW, A
(m)
t,i 2 R
K ,
is pre-processed by using feature scaling, also known as normalization, which is
expressed by
A0(m)
t,i =
A
(m)
t,i min1tK A
(m)
t,i
max1tK A
(m)
t,i min1tK A
(m)
t,i
, for
8
>
<
>
:
i = 1, 2, . . . , N
t = 1, 2, . . . , K
m = 1, 2, . . . , 40
, (5.21)
where A0(m)
t,i 2 R
K and each component A0(m)
t,i corresponds to a time step, t. This
process is carried out for each example sequence (m) from dataset Ai. The main
advantage of feature scaling is that the gradient descent converges much faster with
normalization than without it. Normalization is often employed in neural networks
when the distribution of data is not a Gaussian distribution. In this case, the nature
of rain attenuation data follows a log-normal distribution [28,38].
In summary, the normalized rain-attenuation time-series are ready to be employed
at the input layer of the LSTM network.
C. Initialization of Biases and Weights
The values of weights are usually initialized by random values, whereas the default
value for the bias is 1. However, the initialization could manually be changed. The
weights can be represented by concatenated matrices (Θ), where their dimensions
depend on how deep is the neural network and the number of features of the hypothesis
function.
D. The Loss Function
The loss functions are a measure in order to understand how well a model is able
to predict the expected outcome [84]. It is important to mention that there is not
a unique loss function, therefore, the choice depends on multiple factors such as the
type of machine learning algorithm, presence of outliers, differentiable loss functions,
the efficiency of optimization algorithms, and accurate predictions. Depending
on the machine learning problem, loss functions must be for either regression or
classification [80]. In order to predict the rain attenuation time-series, the chosen
machine-learning-algorithm must be based on a regression problem, therefore, the
loss functions for regression are considered in this thesis.
To begin with, a model is built from the proposed deep learning network (Section
5.2.1), whose performance is evaluated by a loss function. In this context, the loss
function evaluates the differences between predicted and observed values, reaching the
minimum when the prediction is exactly equal to the observed (true) value. For this
purpose, the mean absolute error (MAE) is a loss function used for regression, which
is more robust to outliers [84]. That is, rain-attenuation datasets contain outliers due
to the nature of the rain, therefore, the MAE loss function is useful in these cases. In
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 68
general terms, the MAE loss function is given by
L(Θ)i =
1
m
m
X
j=1
|A0(j)
t,i Â0
t,i|,
8
>
<
>
:
i = 1, 2, . . . , N
j = 1, 2, . . . ,m
t = 1, 2, . . . , K
(5.22)
where the parameter Θ is implicit in the predicted values Â0
t,i. During the training
process, the learning of the parameter Θ is continuous so that the output can have
a better fit and the model can be more accurate. Furthermore, it is important to
note that the parameter Θ is comprised of ΘL and ΘD, where each parameter is
updated in its respective training process. The loss function can be applied to both
the training and validation process. Thus, the model can be evaluated by the training
loss function, L(Θ)train, and the validation loss function L(Θ)val.
E. Optimization Algorithms
Machine Learning algorithms depend on maximizing or minimizing a function. In
this case, the loss function is minimized by optimization algorithms in order to find
the best performance and the minimum of the function. The main optimization
algorithms are presented in the following items.
• Stochastic Gradient Descent Algorithm: this algorithm updates the
parameter Θ to minimize the loss function in the direction of its negative
gradient by taking small steps [82–84]. The update algorithm of the parameter
Θ is denoted by
Θ`+1 := Θ` ↵rΘL(Θ`), (5.23)
where ` is the iteration number and ↵ is the step size or learning rate. The
learning rate setting could be quite tricky. For instance, if the learning rate is
constant, ↵ > 0, but its value is too small, convergence can be very slow, but if
the value is too large, the method can fail to converge at all [84].
In general, this algorithm evaluates the function gradient and updates the
parameter Θ by using a subset derived from the training set. Usually, the subset
is also known as mini-batch [84], whose gradient evaluation using the mini-batch
is performed by each iteration `. During each iteration, the algorithm is one step
closer to minimizing the loss function. Finally, the full-pass when the training
algorithm ends up its process in all mini-batches is an epoch.
• Root Mean Square Propagation (RMSProp): this algorithm tries to
improve network training by using different learning rates, ↵. That is, the
RMSProp algorithm can automatically adapt to the loss function, optimizing
the training process in addition to updating the parameter Θ efficiently [87]. A
moving average of the element-wise squares of the parameter gradients is kept
by this algorithm, which is given by
v` = 2v` 1 + (1 2)[rΘL(Θ`)]
2, (5.24)
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 69
where 2 is the decay factor rate, whose value usually can be 0.9, 0.99, or
0.999. This moving average is used by the RMSProp algorithm to normalize
the updates of the parameter Θ, individually. Parameter Θ is normalized and
updated by the following expression
Θ`+1 := Θ`
↵rΘL(Θ`)p
v` + ✏
, (5.25)
where ✏ is a small constant in order to avoid division by zero. By using RMSProp
algorithms, the learning rate decreases with larger gradients and increases with
small gradients [87].
• Adaptive Moment Estimation Optimizer (ADAM): the adaptive
moment estimation algorithm is a stochastic gradient-descent algorithm with
momentum which keeps an element-wise moving average of both parameter
gradients and their squared values [87]. These average values can be expressed
by
m` = 1m` 1 + (1 1)rΘL(Θ`), (5.26)
v` = 2v` 1 + (1 2)[rΘL(Θ`)]
2, (5.27)
where gradient decay factors, 1 and 2, are linear and quadratic respectively.
Also,rΘL(Θ`) is the gradient of the loss function, and ` is the iteration number.
ADAM optimizer uses these averages to update the parameter Θ, using the
following expression
Θ`+1 := Θ` ↵
m`p
v` + ✏
, (5.28)
where ↵ > 0 is the learning rate, and ✏ is a small constant added in order to avoid
division by zero. In this algorithm, the moving average of the gradient enables
the parameter Θ updates, picking up the momentum in a definite direction
as long as gradients throughout many iterations are similar. Also, the moving
average and parameter Θ can become smaller whether the gradients are noisy.
In general, there is not a specific learning rate, therefore, different learning
rates are tried out to find the optimal values. It can be stipulated a different
learning rate for each layer. Finally, the ADAM optimizer is more efficient than
RMSProp and stochastic gradient descent algorithms, especially due to the fast
convergence and adaptability [87].
F. L2 Regularization
The overfitting can be reduced by adding a regularization term for the weights of the
loss function L(Θ) [87]. Therefore, the loss function with the regularization term is
defined as
L(Θ)R = L(Θ) + RΩ(Θ), (5.29)
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 70
where R is the regularization coefficient, and Ω(Θ) is the regularization function.
Thus, the regularization function is given by vector notation, which is represented by
Ω(Θ) =
1
2
Θ
|
Θ (5.30)
It is important to mention that the biases are not regularized. The R coefficient can
manually be modified, also, it is possible to specify different regularization coefficients
for different layers.
G. Dropout Regularization
Dropout is a regularization method that is capable of dropping out inputs to a layer
by probabilistically removing. To be specific, dropout simulates a large number of
networks but with different structures. Thus, nodes in the network are more robust
to the inputs. For this purpose, the probability of establishing each input to zero is
determined in the layer which is also known as the dropout rate.
For instance, if a dropout rate sets at 0.8, it means that 20% of inputs are to
zero. Therefore, dropout is a computationally-low-cost method to regularize a deep
learning network, to reduce the overfitting, and to improve the generalization error
at the LSTM network.
H. Training Process
By setting the network hyperparameters, the model, let F , is trained by the proposed
deep learning network to forecast rain attenuation time-series in K-time steps into
the future. In this case, the used method is many-to-one, where the input dataset,
A
0
i, is comprised of multiple time-series,
A0(1)
t,i = [A0(1)
1,i , A
0(1)
2,i , . . . , A
0(1)
t,i , . . . , A
0(1)
K,i] 2 R
K
A0(2)
t,i = [A0(2)
1,i , A
0(2)
2,i , . . . , A
0(2)
t,i , . . . , A
0(2)
K,i] 2 R
K
...
A0(m)
t,i = [A0(m)
1,i , A
0(m)
2,i , . . . , A
0(m)
t,i , . . . , A0(m)
K,i ] 2 R
K
, for
8
>
<
>
:
i = 1, 2, . . . , N
t = 1, 2, . . . , K
m = 1, 2, . . . , 40
,
and the single time-series obtained from the deep learning network at the output is
given by
Â0
t+1,i = [Â0
2,i, Â0
3,i, . . . , Â0
t+1,i, . . . , Â0
K+1,i] 2 R
K , for
(
i = 1, 2, . . . , N
t = 1, 2, . . . , K
Here, the feature scaling must be undone according to the feature range parameters
calculated earlier (sub-section 5.2.2-Feature Scaling) [88]. As a result, each
rain-attenuation time-series is denoted by Ât+1,i 2 R
K .
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 71
I. Metrics
The proposed regression model is based on supervised learning, which must be
estimated and evaluated in its performance and fit. Usually, three metrics are used
for this purpose, such as the root mean squared error (RMSE), mean absolute error
(MAE), and the coefficient of determination (R-Squared or R2).
• Root Mean Squared Error: this metric is the most commonly used in
regression problems, which indicates how well the model fits the data. That
is, how near the observed values are from the predicted values, providing an
absolute measure of fit. The RMSE is given by
RMSEi =
v
u
u
t
1
K
K
X
j=1
(Aj,i Âj,i)2, for i = 1, 2, . . . , N (5.31)
To be precise, the RMSE represents the square root of the differences between
observed and predicted values. In other words, it is the quadratic mean of these
differences [82–84]. For calculations over the data sample, these deviations are
known as residuals, whereas they are called errors for out-of-sample calculations.
In general, a lower RMSE value indicates a better fit and accuracy of the model
to predict, which is the most important criterion in prediction models. However,
it is important to mention that RMSE is sensitive to outliers [89].
• Mean Absolute Error: it is the arithmetic average of the absolute errors, i.e.,
between observed and predicted values. This statistical metric is often used as a
measure of forecast error in the time-series analysis [89]. The MAE is calculated
as
MAEi =
1
K
K
X
j=1
|Aj,i Âj,i|, for i = 1, 2, . . . , N (5.32)
Both RMSE and MAE metrics express the average prediction error for LSTM
models. Furthermore, both metrics are indifferent to the direction errors, i.e.,
RMSE and MAE are negatively-oriented scores, which means lower values are
much better. Finally, it is important to mention that MAE is a linear score in
addition to being more robust to outliers [84].
• Coefficient of Determination: it is also known as R-Squared, R2, which is
used in the context of statistical models. Indeed, the main purpose is either the
prediction of future results or the testing of hypotheses, on the basis of other
related information. That is, it is implemented as a measure of how well the
observed outcomes are replicated by the model, based on the proportion of total
variation of outcomes explained by the model [90]. The R-Squared is computed
as
R2
i = 1
PK
j=1(Aj,i Âj,i)
2
PK
j=1(Aj,i Āj,i)2
, for i = 1, 2, . . . , N (5.33)
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 72
In other words, the coefficient of determination is employed to indicate the
goodness of fit of the LSTM model. In this context, R2 could be employed as
an accuracy metric where a value of about 1 indicates that the data predictions
perfectly fitted the proposed model, whereas an R2 = 0 indicates that the
proposed model did not improve the predictions [90].
J. CNIR Time-series Calculation
Predicted by the deep learning network, the rain attenuation time-series are trained
and denoted by Â0
t+1,i = F(A0(m)
t,i ) for each GW location, i = 1, 2, . . . , N . As
previously stated, the normalization must be undone so the time-series can be
expressed on the original scale. Thus, rain attenuation time-series can be given by
Ât+1,i 2 R
K . By using Eq. (2.3), the predicted sequence of CNR time-series for each
feeder uplink can be expressed by
Γ̂t+1,i = 10 log10( csi) Ât+1,i, for
(
i = 1, 2, . . . , N
t = 1, 2, . . . , K
(5.34)
where Γ̂t+1,i 2 R
K and each element, Γ̂t+1,i, is given in decibels, dB. Finally,
substituted the predicted CNR into Eq. (2.6), the predicted sequence of CNIR
time-series is calculated for each feeder uplink and rewritten as
Ξ̂t+1,i =
Γ̂t+1,i · ⇣T
Γ̂t+1,i + ⇣T
, for
(
i = 1, 2, . . . , N
t = 1, 2, . . . , K
(5.35)
where Ξ̂t+1,i 2 R
K is the CNIR prediction for the next time step, t + 1. Figure 5.10
depicts a block diagram of the proposed method where each GW processes its both
rain-attenuation and CNIR predictions to send them to the Network Control Center
(NCC), At NCC, each predicted time-series component of both rain attenuation and
CNIR can be stored into matrices, Â, Ξ̂ 2 R
K⇥N , respectively. Thus, the NCC
analyzes each feeder-uplink status in order to determine which of them are threatened
in advance by rain impairments.
5.3 Experimentation
In order to prove the deep learning network performance, two experiments were
executed by employing a high-level programming language. For this purpose, 24
locations were chosen as possible GW sites, therefore, there were obtained 24 rain
attenuation datasets with 40 sequences for each one of them. In this context, the
possible GW-sites are dispersed in Mexico, Central America, and The Caribbean,
as these cities can access easier to high-quality services and infrastructures so the
implementation and deployment can be faster and more efficient than in minor cities.
Furthermore, the number of examples, m, emulates a historical database of 40 years.
Table A.2 (Appendix A.2) shows in detail the main features of each GW site, in
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 73
NCC
Matrices:
Â, Ξ̂ 2 R
K⇥N
Data processing in situ, GW1
Feature
Scaling
Deep Learning
Network
F
CNIR
Calculation
A
(m)
t,1
time
step, t
A0(m)
t,1
time
step, t
Â0
t+1,1
time
step,
t+ 1
Ξ̂t+1,1
time
step,
t+ 1
Data processing in situ, GW2
Feature
Scaling
Deep Learning
Network
F
CNIR
Calculation
A
(m)
t,2
time
step, t
A0(m)
t,2
time
step, t
Â0
t+1,2
time
step,
t+ 1
Ξ̂t+1,2
time
step,
t+ 1
...
Data processing in situ, GWN
Feature
Scaling
Deep Learning
Network
F
CNIR
Calculation
A
(m)
t,N
time
step, t
A0(m)
t,N
time
step, t
Â0
t+1,N
time
step,
t+ 1
Ξ̂t+1,N
time
step,
t+ 1
Figure 5.10: Block diagram of the proposed method based on the deep learning
network.
addition to its calculated carrier-to-noise ratio under clear-sky conditions according
to the discussed method in Chapters 2 and 3, csi , for i = 1, 2, . . . , N [dB]. In this
context, a lot of time was invested by computer coding to carry out the experiments
successfully. Provided in the following sub-sections, each experiment details the
employed settings and procedure.
5.3.1 Experiment 1: Training and validation subsets
partitioning, 70/30
To begin with, the rain attenuation datasets for the set [N ] = {1, 2, . . . , 24} were
calculated by using the model presented in sub-section 2.4. This model employed
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 74
geographic coordinates (latitude, longitude, and height) of each GW site, satellite
orbital position at 92 West, the frequency band of the feeder uplink at 50 GHz, and
Ts = 1 min. It is important to highlight that the rain attenuation time-series were
sampled every Ts = 1 min at 1 year period, therefore, the number of samples of each
rain-attenuation time-series is 525600, that is, each sequence has a dimensionality of
K = 525600.
Each time-series was partitioned into training and validation subsets, where the
training subset was 70% and the validation subset was 30%. Hereinafter, each
rain attenuation time-series was normalized by feature scaling to be trained and
validated in the deep learning network. Table 5.1 shows the hyperparameters that
are incorporated into the deep learning network configuration. To be precise, the
hyperparameters adjustment is an important role in the learning process. For this
purpose, different values were assigned to each hyperparameter such as the number
of hidden layers, dropout rate, number of LSTM units, learning rate, etc. In this
case, the deep learning network employed a single LSTM layer, S = {LU}. After
several trials, where about 400 models were tried out with different hyperparameter
combinations, the best model was found according to the best hyperparameters. In
summary, this deep learning network configuration was able to train and validate
960 rain attenuation time-series regarding 24 GW sites with 40 examples by the
site-specific.
Table 5.1: Hyperparameters for Experiment 1
Parameter Arg./Value Parameter Arg./Value
LSTM units, U 100 Dense Layer: (D) output 1
Activation Function (LU ) tanh Activation Function (D) selu
L2 R 1⇥ 10 4 1 0.9
Dropout 0.2 2 0.999
Loss Function MAE ✏ 1⇥ 10 8
Optimizer ADAM Metrics RMSE, MAE, R2
Learning rate 5⇥ 10 3 Validation Yes
Epochs 250 Batch Size 1024
Early Stopping Yes Patient 5
5.3.2 Experiment 2: Training and validation subsets
partitioning, 90/10
This procedure was similar to Experiment 1, but the rain attenuation time-series
were divided into 90% for the training subsets and 10% for the validation subsets.
Unlike in Experiment 1, the number of hidden units increased in the LSTM layer.
Also, the LSTM network was modified by adding more layers. Therefore, the LSTM
multi-layer network contributed to the proposed model to be deeper than Experiment
1. Thus, the deeper network might need fewer epochs to train the model, but the
computational cost might be higher.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 75
Table 5.2 indicates the obtained hyperparameters for Experiment 2. Similarly,
the hyperparameters tuning was performed according to the Experiment 1 process.
However, it was necessary to test about 600 models to find the best hyperparameters
that are part of the model settings for Experiment 2. It is important to mention that
the deep learning network is comprised of 5 LSTM layers, where each layer of the
LSTM multi-layer network, S = {LU1,LU2,LU3,LU4,LU5}, such that U1 = 200, U2 =
75, U3 = 50, U4 = 25, U5 = 5 has its own LSTM hidden units, respectively.
Table 5.2: Hyperparameters for Experiment 2
Parameter Arg./Value Parameter Arg./Value
LSTM units, U1 200 LSTM units, U2 75
Activation Function (LU1) tanh Activation Function (LU2) tanh
LSTM units, U3 50 LSTM units, U4 25
Activation Function (LU3) tanh Activation Function (LU4) tanh
LSTM units, U5 5 Dense Layer: (D) output 1
Activation Function (LU5) tanh Activation Function (D) selu
L2 R 1⇥ 10 6 1 0.9
Dropout 0.2 2 0.999
Loss Function MAE ✏ 1⇥ 10 8
Optimizer ADAM Metrics RMSE, MAE, R2
Learning rate 1⇥ 10 3 Validation Yes
Epochs 250 Batch 1024
Early Stopping Yes Patient 5
5.4 Results
Both experiments were performed by computer resources to train and validate the
proposed models, where Python programming language was employed, in addition to
using tensor processor units (TPU) and graphics processor units (GPU) to accelerate
the training and validation processes, reducing the computing time and improving the
general performance. In order to evaluate both experiments, three statistical metrics
were used for this purpose: root mean squared error (RMSE), mean absolute error
(MAE), and the coefficient of determination (R2).
Figure 5.11 and Figure 5.12 depict RMSE scores for both training and validation
subsets, respectively. In both cases, the best results were obtained by Experiment 2.
However, some RMSE scores for training subsets that were evaluated in each site by
Experiment 1 such as San Jose, Tegucigalpa, Guadalajara, La Habana, and Belmopan
were lower than RMSE scores from Experiment 2, without having a remarkable
difference for all reported cases. Furthermore, for validation experiments, some RMSE
scores from Experiment 1 such as Tijuana, Torreon, La Paz B.C., Tuxtla Gtz., San
Pedro Sula, Belmopan, Kingston, and Oaxaca were also lower than RMSE scores
from Experiment 2. Anyway, the average RMSE score obtained from Experiment 2
was lower than Experiment 1.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 76
Panama
San Jose
Tegucigalpa
Mexico City
Monterrey
Guadalajara
Tijuana
La Habana
Sto. Domingo
San Salvador
San Juan
Torreon
La Paz, B.C.
Veracruz
Cancun
Queretaro
Tuxtla, Gtz.
Cd. Juarez
San Pedro Sula
Belmopan
Kingston
Merida
Pto. Cabezas
Oaxaca
Average
RMSE scores
0.00 0.50 1.00 1.50 2.00 2.50 3.00 3.50
1.2497
0.5645
2.1529
1.6796
1.4175
1.6505
1.4767
0.4130
0.9962
0.5574
1.4978
1.4395
0.5575
0.3757
2.2306
1.9227
1.5236
1.4686
0.4728
0.8090
0.9335
0.8344
1.0816
1.3889
2.5494
1.4128
0.6186
2.2341
1.8961
1.4263
1.6471
1.5923
0.8004
1.0823
0.6657
1.9313
1.8645
0.6558
0.3986
2.2711
2.2610
1.8496
1.4448
0.5073
0.7815
1.8806
1.0982
0.9481
1.0899
2.9620
Experiment 1 Experiment 2
Figure 5.11: RMSE scores for training
subsets.
Panama
San Jose
Tegucigalpa
Mexico City
Monterrey
Guadalajara
Tijuana
La Habana
Sto. Domingo
San Salvador
San Juan
Torreon
La Paz, B.C.
Veracruz
Cancun
Queretaro
Tuxtla, Gtz.
Cd. Juarez
San Pedro Sula
Belmopan
Kingston
Merida
Pto. Cabezas
Oaxaca
Average
RMSE scores
0.00 0.50 1.00 1.50 2.00 2.50 3.00
1.1161
0.5084
1.5361
0.6879
1.7276
1.4397
1.4027
0.2835
1.3592
0.4522
1.2996
1.1320
0.4628
0.4678
2.0507
1.6287
1.8867
1.4921
0.4668
0.7949
0.5452
0.6607
1.1814
1.6696
1.6492
1.2780
0.4671
1.7611
0.8933
1.4451
1.3434
1.3723
0.7270
1.1784
0.6492
1.6838
1.6038
0.4368
0.4123
2.2320
1.9099
2.1635
1.7415
0.3821
0.9179
1.9136
0.7154
1.2707
1.7785
1.6739
Experiment 1 Experiment 2
Figure 5.12: RMSE scores for validation
subsets.
Figure 5.13 and Figure 5.14 show MAE scores for training and validation subsets,
respectively. For both training and validation subsets, the MAE scores were closely
similar to each other. Here, the average MAE score from Experiment 2 was lower than
Experiment 1 for both training and validation evaluations. Nevertheless, some MAE
scores from Experiment 1 were lower than Experiment 2, but these scores did not
influence strongly in the final average score. In many cases, RMSE scores are larger
than MAE scores, but if the errors have the same magnitude, then RMSE=MAE.
Figure 5.15 and Figure 5.16 illustrate the training and validation rain attenuation
subsets of each site, respectively. In order to visualize properly, 50 samples were picked
out by each training and validation figure. That is, each range was randomly chosen
by illustrating rain attenuation samples. Furthermore, each figure, either training or
validation subset, depicted both Experiments and the Observed data. In this sense,
results from Experiment 2 were adequately fitted better than Experiment 1. For
instance, although Tijuana training and validation subsets, Figure 5.15g and Figure
5.16g respectively, were practically the same for both experiments, other evaluations
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 77
Panama
San Jose
Tegucigalpa
Mexico City
Monterrey
Guadalajara
Tijuana
La Habana
Sto. Domingo
San Salvador
San Juan
Torreon
La Paz, B.C.
Veracruz
Cancun
Queretaro
Tuxtla, Gtz.
Cd. Juarez
San Pedro Sula
Belmopan
Kingston
Merida
Pto. Cabezas
Oaxaca
Average
MAE scores
0.00 0.40 0.80 1.20 1.60 2.00
0.2737
0.0742
0.4900
0.2450
0.5665
0.5031
0.2653
0.0242
0.1728
0.0740
0.6175
0.2043
0.0429
0.0945
0.8975
0.2852
0.4857
0.1622
0.0924
0.2127
0.1111
0.1141
0.1396
0.4334
0.2615
0.5384
0.1768
0.7351
0.1654
0.3561
0.2078
0.7701
0.6915
0.4645
0.2688
1.4042
0.9200
0.1918
0.0330
0.2894
0.4480
1.3543
0.7583
0.0827
0.2096
1.6538
0.4870
0.1823
0.2891
0.7828
Experiment 1 Experiment 2
Figure 5.13: MAE scores for training
subsets.
Panama
San Jose
Tegucigalpa
Mexico City
Monterrey
Guadalajara
Tijuana
La Habana
Sto. Domingo
San Salvador
San Juan
Torreon
La Paz, B.C.
Veracruz
Cancun
Queretaro
Tuxtla, Gtz.
Cd. Juarez
San Pedro Sula
Belmopan
Kingston
Merida
Pto. Cabezas
Oaxaca
Average
MAE scores
0.00 0.40 0.80 1.20 1.60 2.00
0.2777
0.0777
0.4528
0.1977
0.5786
0.5131
0.2702
0.0189
0.2155
0.0639
0.6010
0.2026
0.0424
0.1047
0.8935
0.3230
0.5557
0.1875
0.0943
0.2100
0.0753
0.1320
0.1737
0.4506
0.2295
0.5355
0.1639
0.7318
0.1255
0.3462
0.1876
0.7517
0.6835
0.4798
0.2744
1.3750
0.9172
0.1825
0.0362
0.2831
0.4503
1.4161
0.7709
0.0752
0.2149
1.6585
0.4774
0.2058
0.3196
0.7259
Experiment 1 Experiment 2
Figure 5.14: MAE scores for validation
subsets.
such as Monterrey (Figure 5.15e and Figure 5.16e) or Torreon (Figure 5.15l and
Figure 5.16l) had shown a remarkable difference. Unfortunately, some training subsets
experienced troubles during the training process, therefore, the hyperparameters such
as dropout or LSTM units were modified in some evaluations such as: Torreon, La
Paz, B.C., Queretaro, Cd. Juarez, and Pto. Cabezas. This was necessary to avoid
under-fitting and over-fitting on predicted subsets.
Figure 5.17 and Figure 5.18 indicate the R2 values of both training and validation
subsets respectively, where Experiment 2 predicted the outcomes slightly better than
Experiment 1. R2 indicated how well the regression predictions approximated the
real data points.
According to results from Experiment 2, Figure 5.19 shows the validation metrics
for each evaluated rain-attenuation. Experiment 2 was chosen because the overall
model performance obtained from it was better than Experiment 1. It was possible
to note that in some cases because of the many outliers, the predicted values did
not fit well on the regression line. For instance, in a few cases such as La Paz, B.C.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 78
(a) Panama, range: k = 19110, . . . , 19160 (b) San Jose, range: k = 25540, . . . , 25590
(c) Tegucigalpa, range: k = 14075, . . . , 14125 (d) Mexico City, range: k = 10520, . . . , 10570
(e) Monterrey, range: k = 26090, . . . , 26140 (f) Guadalajara, range: k = 13100, . . . , 13150
(g) Tijuana, range: k = 259980, . . . , 260030 (h) La Habana, range: k = 26170, . . . , 26220
(i) Sto. Domingo, range: k = 50465, . . . , 50515 (j) San Salvador, range: k = 32255, . . . , 32305
(k) San Juan, range: k = 18050, . . . , 18100 (l) Torreon, range: k = 41965, . . . , 42015, dropout=0.5
Figure 5.15: The training rain-attenuation subsets obtained from deep learning models.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 79
(m) La Paz, B.C., range: k = 52215, . . . , 52265, dropout=
0.5
(n) Veracruz, range: k = 14183, . . . , 14233
(o) Cancun, range: k = 37100, . . . , 37150, dropout= 0.5 (p) Queretaro, range: k = 15215, . . . , 15265, dropout= 0.5
(q) Tuxtla, Gtz., range: k = 38320, . . . , 38370
(r) Cd. Juarez, range: k = 74145, . . . , 74195, LU1 = 400,
dropout=0.5
(s) San Pedro Sula, range: k = 10000, . . . , 10050 (t) Belmopan, range: k = 12165, . . . , 12215
(u) Kingston, range: k = 17350, . . . , 17400 (v) Merida, range: k = 27598, . . . , 27648, dropout= 0.3
(w) Pto. Cabezas, range: k = 19180, . . . , 19230, dropout=
0.5
(x) Oaxaca, range: k = 11680, . . . , 11730
Figure 5.15: The training rain-attenuation subsets obtained from deep learning models
(cont.).
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 80
(a) Panama, range: k = 475460, . . . , 475510 (b) San Jose, range: k = 477390, . . . , 477440
(c) Tegucigalpa, range: k = 482140, . . . , 482190 (d) Mexico City, range: k = 477490, . . . , 477540
(e) Monterrey, range: k = 474550, . . . , 474600 (f) Guadalajara, range: k = 479480, . . . , 479530
(g) Tijuana, range: k = 510090, . . . , 510140 (h) La Habana, range: k = 505318, . . . , 505368
(i) Sto. Domingo, range: k = 504330, . . . , 504380 (j) San Salvador, range: k = 503138, . . . , 503188
(k) San Juan, range: k = 519440, . . . , 519490 (l) Torreon, range: k = 496275, . . . , 496325, dropout=0.5
Figure 5.16: The validation rain-attenuation subsets obtained from deep learning models.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 81
(m) La Paz, B.C., range: k = 518740, . . . , 518790,
dropout= 0.5
(n) Veracruz, range: k = 493940, . . . , 493990
(o) Cancun, range: k = 495150, . . . , 495200, dropout= 0.5
(p) Queretaro, range: k = 501165, . . . , 501215, dropout=
0.5
(q) Tuxtla, Gtz., range: k = 483430, . . . , 483480
(r) Cd. Juarez, range: k = 475245, . . . , 475295, LU1 =
400, dropout=0.5
(s) San Pedro Sula, range: k = 483840, . . . , 483890 (t) Belmopan, range: k = 488025, . . . , 488075
(u) Kingston, range: k = 486550, . . . , 486600 (v) Merida, range: k = 499855, . . . , 499905, dropout= 0.3
(w) Pto. Cabezas, range: k = 476680, . . . , 476730,
dropout= 0.5
(x) Oaxaca, range: k = 505970, . . . , 506020
Figure 5.16: The validation rain-attenuation subsets obtained from deep learning models
(cont.).
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 82
Panama
San Jose
Tegucigalpa
Mexico City
Monterrey
Guadalajara
Tijuana
La Habana
Sto. Domingo
San Salvador
San Juan
Torreon
La Paz, B.C.
Veracruz
Cancun
Queretaro
Tuxtla, Gtz.
Cd. Juarez
San Pedro Sula
Belmopan
Kingston
Merida
Pto. Cabezas
Oaxaca
Average
R-Squared
0.00 0.22 0.44 0.66 0.88 1.10
0.9020
0.9013
0.9216
0.9414
0.8828
0.9090
0.8947
0.8851
0.9032
0.9047
0.8856
0.9186
0.8588
0.8242
0.9156
0.9168
0.9045
0.8998
0.9088
0.9062
0.9101
0.9251
0.9054
0.9036
0.9214
0.8626
0.8935
0.9238
0.9389
0.8896
0.9166
0.8873
0.6403
0.8920
0.8602
0.8331
0.8711
0.8396
0.8187
0.9128
0.8923
0.8316
0.8748
0.9088
0.8942
0.5694
0.8948
0.8952
0.9116
0.9117
Experiment 1 Experiment 2
Figure 5.17: R2 for training subsets.
Panama
San Jose
Tegucigalpa
Mexico City
Monterrey
Guadalajara
Tijuana
La Habana
Sto. Domingo
San Salvador
San Juan
Torreon
La Paz, B.C.
Veracruz
Cancun
Queretaro
Tuxtla, Gtz.
Cd. Juarez
San Pedro Sula
Belmopan
Kingston
Merida
Pto. Cabezas
Oaxaca
Average
R-Squared
0.00 0.22 0.44 0.66 0.88 1.10
0.8783
0.8810
0.8892
0.8435
0.9068
0.8979
0.8988
0.7265
0.9224
0.8653
0.8877
0.8959
0.8045
0.8156
0.9027
0.9033
0.9160
0.8900
0.8867
0.9180
0.8080
0.8864
0.9038
0.9223
0.9076
0.8399
0.8922
0.9021
0.9023
0.8908
0.9128
0.8833
0.3332
0.8963
0.8543
0.7579
0.8534
0.7324
0.7762
0.9062
0.8846
0.8819
0.9064
0.8969
0.9160
0.6495
0.7884
0.9264
0.9221
0.8909
Experiment 1 Experiment 2
Figure 5.18: R2 for validation subsets.
(Figure 5.19m), or Cd. Juarez (Figure 5.19r), the squared residuals with respect to
the linear regression were higher than the best models such as Tuxtla, Gtz. (Figure
5.19q), or San Jose (Figure 5.19b).
By using Eq. (5.34) and Eq. (5.35), it was possible to predict the CNIR for
the next time step, t + 1. For this purpose, the co-channel interference value, ⇣co =
36.09 dB, and the adjacent interference value, ⇣adj = 38.17 dB, were obtained from
subsection 4.5.3. Figure 5.20 depicts the model with higher R-Squared, R2 = 0.9224,
which demonstrates that the rain attenuation, Figure 5.20a, directly influences on the
CNIR, Figure 5.20b, and spectral efficiency, Figure 5.20c. On the other hand, Figure
5.21 indicates the model with lower R-Squared, R2 = 0.7265. In both cases, only 50
samples were employed to visualize properly the effects of rain attenuation. Also, the
spectral efficiency was determined from CNIR values by the DVB-S2X table [6].
These two approaches can be also analyzed from a climatological perspective to
compare with the obtained prediction models. For instance, Tuxtla, Gtz. is located
in the rainforest area of Mexico regarding the Köppen-Geiger climate type map [91].
On the other hand, Cd. Juarez is located in the cold desert zone of Mexico according
to the Köppen-Geiger climate type map of North America [91]. This makes sense,
as the mean annual precipitation at Tuxtla, Gtz. is about 1900 mm, whereas at Cd.
Juarez is less than 200 mm [92]. For this reason, the Tuxtla prediction model was not
difficult to train because there were more precipitation data than at Cd. Juarez. That
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 83
(a) Panama (b) San Jose
(c) Tegucigalpa (d) Mexico City
(e) Monterrey (f) Guadalajara
(g) Tijuana (h) La Habana
(i) St. Domingo (j) San Salvador
(k) San Juan (l) Torreon, dropout=0.5
Figure 5.19: Model performance results: measured vs. predicted rain-attenuation values
for validation.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 84
(m) La Paz, dropout= 0.5 (n) Veracruz
(o) Cancun, dropout= 0.5 (p) Queretaro, dropout= 0.5
(q) Tuxtla, Gtz. (r) Cd. Juarez, LU1 = 400, dropout=0.5
(s) San Pedro Sula (t) Belmopan
(u) Kingston (v) Merida, dropout= 0.3
(w) Pto. Cabezas, dropout= 0.5 (x) Oaxaca
Figure 5.19: Model performance results: measured vs. predicted rain-attenuation values
for validation (cont.).
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 85
(a) Rain attenuation window
(b) CNIR window
(c) Spectral efficiency window
Figure 5.20: The higher R-Squared, R2 =
0.9224 at Tuxtla, Gtz.
(a) Rain attenuation window
(b) CNIR window
(c) Spectral efficiency window
Figure 5.21: The lower R-Squared, R2 =
0.7265 at Cd. Juarez.
is, the lack of data is a big problem because it can be difficult to train and validate the
model, obtaining low accuracy, in addition to not finding reasonable results. Finally,
Figure 5.22 indicates the performance of the deep learning network configuration by
showing both training and validation loss functions for each rain-attenuation time
series. Both loss function graphics were the results of Experiment 2.
5.5 Discussions
The outcomes of Experiment 2 were better than Experiment 1, improving scores of
metrics such as RMSE and MSE. Therefore, the trained and validated model obtained
from Experiment 2 was the basis for understanding and forecasting accurately the
rain-attenuation and their strong influence on feeder links. However, the most
important metric was the coefficient of determination or R2, which was decisive in
order to pick out the best Experiment in this study. R2 is a number between 0 and 1
that estimates the changes in the dependent variable by changes in the independent
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 86
(a) Panama (b) San Jose
(c) Tegucigalpa (d) Mexico City
(e) Monterrey (f) Guadalajara
(g) Tijuana (h) La Habana
(i) St. Domingo (j) San Salvador
(k) San Juan (l) Torreon, dropout=0.5
Figure 5.22: Training and validation loss functions for each rain-attenuation time series.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 87
(m) La Paz, dropout= 0.5 (n) Veracruz
(o) Cancun, dropout= 0.5 (p) Queretaro, dropout= 0.5
(q) Tuxtla, Gtz. (r) Cd. Juarez, LU1 = 400, dropout=0.5
(s) San Pedro Sula (t) Belmopan
(u) Kingston (v) Merida, dropout= 0.3
(w) Pto. Cabezas, dropout= 0.5 (x) Oaxaca
Figure 5.22: Training and validation loss functions for each rain-attenuation time series
(cont.).
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 88
variable. In fact, a high R2, close to 1, is necessary for precise predictions, therefore,
the found R2 from validation subsets were between 0.7265 and 0.9224. Potentially,
these values demonstrate a high accuracy of deep learning networks obtained from
Experiment 2 unlike deep learning networks from Experiment 1, as seen in Figure
5.18.
In order to understand graphically R2, data points of deep learning networks
with high R2 are lying near to the regression line, whereas data points of deep
learning networks with low R2 are dispersed, as demonstrated in Figure 5.19. After
all, determining an adequate R2 is a matter of judgment. Hence, the found R2s
accomplish the requirements for all rain-attenuation models.
In this sense, the predictive model of rain attenuation with high R2, Figure 5.20, is
compared with the predictive model with low R2, Figure 5.21. For instance, the first
model predicted an aggressive rain-attenuation in the range from k = 483430–483480.
In this case, CNIR predictions varied from -50.0–30.0 dB, approximately, impacting
strongly on the spectral efficiency. On the contrary, a non-aggressive rain-attenuation
was predicted by the second model in the range from k = 475245–475295. In this case,
CNIR predictions are more stable due to the lower attenuation of the rain, therefore,
the spectral efficiency is almost constant in that range, as seen in Figure 5.21c.
It is important to mention that the upper limit of the DVB-S2X standard, in
power terms, is 19.57 dB, which corresponds to the best modulation and codification,
256 APSK 3/4, and spectral efficiency of 5.9009 b/s/Hz [6]. That is, the CNIR has
a margin of about 10.00 dB until the signal is degraded by rain attenuation. In
other words, the spectral efficiency varies according to the CNIR level, as shown in
Figure 5.20b and Figure 5.20c. In this case, the rain attenuation greater than 10.00
dB affects strongly the CNIR level, degrading the spectral efficiency until there is no
signal, i.e., if the CNIR < 2.85 dB, then spectral efficiency = 0.00 b/s/Hz.
Furthermore, it was figured out that the validation loss, L(Θ)val, is lower than
training loss, L(Θ)train, during the training and validation processes of the proposed
models, as shown in Figure 5.22. Nonetheless, this is not fundamentally an uncertain
phenomenon in essence. Indeed, the structure of both training and validation subsets
can have a smaller validation loss than the training loss. This usually happens for
five main reasons,
• When the training data is harder to train or learn patterns on it, whereas it is
possible to have an easier validation subset.
• Dropout is not enabled during the validation process, therefore, this can lead
to greater outcomes on the validation subset.
• By default, regularization is solely applied during the training process but not
during the validation or testing process. Nevertheless, when the regularization
is present during the validation process, then both loss functions can be more
similar.
• On average, the training loss is performed during each epoch whereas the
validation loss is performed after each epoch. That is, the training loss can
be out of phase a half epoch regarding the validation loss.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 89
Accuracy Model [%]
0% 25% 50% 75% 100%
Random Forest, XGBoost 77.00%
Statistical Method 77.55%
Statistical Method 81.15%
LSTM 87.99%
LSTM 87.83%
LSTM 72.65%
LSTM 92.24%
Proposed method (High R-Squared)
Proposed method (Low R-Squared)
Proposed method (Average 24 sites)
Poornima - Pushpalatha method
ARIMA method
Holt-Winters method
Aguasca-Colomo et. al method
Figure 5.23: Comparison between predictive models.
• It is possible that in the early phase of the training process the validation loss
is smaller than the training loss.
In this thesis, dropout and regularization methods were added to the LSTM
networks, there were even some different configurations to train hard LSTM networks,
for instance, Torreon (dropout = 0.5), La Paz, B.C. (dropout= 0.5), Cancun (dropout
= 0.5), Queretaro (dropout = 0.5), Cd. Juarez (LU1 = 400 units and dropout
= 0.5), Merida (dropout = 0.3), and Pto. Cabezas (dropout = 0.5). Thus, the
first reason is essentially the main factor to diminish the validation loss in contrast
to the training loss. Hence, data are prepared carefully, as the outcomes depend
significantly on the quality of data, i.e, pre-processing data, cleaning data, and
partitioning training/validation subsets.
Finally, Figure 5.23 illustrates the comparison between predictive models/methods
according to the accuracy metric. For instance, the best model found in this
thesis, Experiment 2: 90/10, has better accuracy than the Poornima-Pushpalatha
method [93], i.e., 92.24% vs. 87.99%. However, on average, the proposed method
is slightly lower than the Poornima-Pushpalatha method, 87.83% vs. 87.99%. It
is important to mention that the proposed method in this thesis evaluated 24
locations whereas the Poornima-Pushpalatha method evaluated a single location at
Hyderabad, India. Other proposals employed statistical methodologies, such as the
autoregressive integrated moving average (ARIMA) method, 81.15% of accuracy, [94],
and Holt-Winters method, 77.55% of accuracy, [95], but they could not overcome the
accuracy of the proposed LSTM method. At last but not least, the Aguasca-Colomo
et. al method proposed a random forest model based on classification with an accuracy
of 77.00% [96]. As a result, Machine Learning methods based on regression techniques
have better performance and accuracy than statistical methods in order to predict
rain events.
To sum up, this proposal guarantees an accurate rain-attenuation prediction at
the evaluated sites, generalizing efficiently the models. Therefore, this method can
efficiently contribute to managing the resources of both the space segment and the
ground segment of a high throughput satellite system.
CHAPTER 5. METHOD OF RAIN ATTENUATION PREDICTION . . . 90
5.6 Contributions of the Research
The proposed method and found results in this chapter were submitted to Neural
Processing Letters journal to be reviewed and published later. Also, in this context,
an initial study based on neural networks was presented in a specialized conference:
"XXXIII Simposium Nacional de la Unión Científica Internacional de Radio, URSI
2018," at Granada, Spain.
• Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez Rodriguez-Osorio,
and Jose Maria Matias. Optimization of the Ground Segment for an UHTS
System through Neural Networks. In XXXIII Simp. Nac. la Unión Científica
Int. Radio (URSI 2018), pages 1–5, Granada, Spain, 2018
Chapter 6
Ground Segment Optimization by
Using Smart Strategies of Switching
Between Gateway Stations
6.1 Introduction
The ground segment (GS) is an important part of satellite communications, which
contains the gateways stations (GW) and user terminals (UT-VSAT) necessary to
communicate to/from the high throughput satellite HTS. When a user requires the
satellite service, then the operator adds a VSAT to the network, as long as there
is availability. Meanwhile, the number of nominal gateways NGWs depends on the
satellite capacity, i.e., each NGW manages a portion of the total capacity, as can be
seen in Section 3.4. In this context, it is simple to determine the number of GWs, but
it does not consider when the rain affects the feeder links. That is, if the rain impairs
one or more GWs, the satellite capacity diminishes, impacting directly on the quality
of service (QoS) and leaving several VSATs out-of-service.
In a sense, by using the Adaptive Coding and Modulation (ACM) from the
DVB-S2X standard, the system availability might not drop dramatically [6]. Also,
the Uplink Power Control (ULPC) can manage a few dBs to compensate for channel
degradation [74–76]. Despite the fact that both techniques can mitigate channel
degradation, they can not combat the aggressive effects of the heavy rain over feeder
links.
For this reason, some researchers recommend that the NGWs must be separated
by a distance D to avoid the correlated rain to each other, i.e. so that the rain affects
only one NGW at a time. Each NGW has an extra resource to support when an
affected feeder link is in an outage due to the heavy rain. This technique is known as
Smart Gateway Diversity (SGD), where the affected NGW is capable of re-routing
traffic to an alternate NGW with extra resources for temporary allocation, improving
the availability considerably [26,28].
Other researchers have proposed to add redundant gateway stations as a backup
(PGWs), i.e., when the rain attenuation impacts on NGW’s feeder link, an available
91
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 92
PGW gets into the network rather than the affected NGW. Temporarily, the PGW
manages all traffic of the affected NGW and establishes a link between the satellite
and PGW, waiting for the NGW is available again. This configuration is also known
as the N̄ + P̄ scheme, whose NGWs are denoted by N̄ , and PGWs are described by
P̄ [22, 23,26,97].
However, this scheme could be an inelegant solution and very little efficient
without evaluating the adequate number of PGWs necessary for keeping up the
system availability. Here, some researchers have simplified the analysis by proposing
a 1 + 1 scheme and using stationary distributions or Markov Chains to calculate the
number of PGWs properly [21, 25, 98, 99]. Naturally, this leads to a better switching
strategy between NGWs and PGWs but without achieving an adequate optimization
of the ground segment. Other authors have developed interesting methods in order
to optimize the ground segment in a particular way [100,101].
Despite the smart gateway diversity is focused on increasing the system availability
and improving the switching strategies, it is not based on reducing the number of
PGWs optimally. For this purpose, a predictive method, as proposed in Chapter
5, provides predictions of rain attenuation by time intervals, identifying how many
times the NGWs could be with a link outage due to the rain and which PGWs
could temporarily replace to each one of them. Compared with other methods, this
predictive method aids to a large degree to optimize the number of PGWs. This
contribution must guarantee an availability higher than 99.9%, foreseeing accurately
when a feeder link is going to be affected by the rain. Also, this method contributes
to the switching strategy to be of lower complexity than previous methods.
6.2 Proposed Model
To begin with, the ITU-R 1815 recommended that the distance between two GWs
must be 80 km to avoid the correlated rain, dropping the spatial correlation
coefficient ⇢a to < 0.1, as can be seen in Section2.3. By using the coordinates of
Appendix A.2, it was possible to determine the distance matrix, D, between the GW
locations. Furthermore, GW locations were the same as those employed by analysis
of Chapter 5. Appendix A.3 presents the distance matrix between pairs of locations.
In all cases, the distances were greater than 80 km between each other, ensuring the
uncorrelated rain for all locations.
6.2.1 Sizing of Nominal Gateways, NGWs
In the high throughput satellite (HTS) systems, the total capacity is strongly related
to the number of nominal gateways (NGW) on the ground segment (GS). Initially,
the data traffic is symmetrically assumed between forward and return links. That is,
50% of the inbound’s throughput and 50% in the outbound, [1 : 1]. In this context,
the upload speed is not strained by the symmetrical connection, whose data is flowed
in a proper way [18].
Each NGW manages its feeder link where the feeder uplink is the focus of this
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 93
study. For this case, the feeder uplink does not use the frequency reuse, although it
uses double polarization. Based on Eq. (3.1) and Eq. (3.2), the capacity, in b/s, of
each feeder uplink is expressed as
Cpi =
Np · Savail
Nsb
Seffi , for i = 1, 2, . . . , N̄ , (6.1)
where Np is the number of polarizations, Nsb is the number of frequency sub-bands,
Savail is the available spectrum, and Seff is the spectral efficiency from the DVB-S2X
standard by each feeder uplink [6,18], as was discussed in Chapter 3. At this point, the
inbound uplink beams are received and demodulated by the satellite communication
subsystem. Naturally, each NGW transmits its corresponding feeder uplink beam.
Thus, the total capacity is directly related to the number of NGWs on the GS, which
is denoted by
CpT =
N̄
X
i=1
Cpi (6.2)
In this study, the inbound CpT is computed by four configurations of NGWs, where
N̄ = 4, 8, 12, and 16.
6.2.2 Smart Method for Forecasting Rain Attenuation and
CNIR at each GW
This novel method is based on Machine Learning algorithms by the implementation
of LSTM networks. To begin with, the rain attenuation time-series for each possible
GW location was obtained by ITU Recommendation, ITU-R 1853 [38]. Thus, each
rain attenuation time-series was trained and validated by LSTM networks, as can be
seen in Chapter 5 in detail. As a result, each predicted rain attenuation time-series,
at time step t + 1, can be denoted by Ât+1,i = F(A
(m)
t,i ) for i = 1, 2, . . . , N , where
F is the deep learning model, m is the number of examples of the rain attenuation
time-series by each GW-site. Figure 5.10 illustrates the block diagram of the smart
method employed to predict rain attenuation time-series in this proposal.
Finally, each prediction of rain attenuation time-series can be stored in matrix
 2 R
K⇥N . Using Eq. (5.34) and Eq. (5.35), it was possible to compute the
predicted CNIR, Ξ̂ 2 R
K⇥N , at time step t + 1. Both matrices can be processed
and employed by NCC to supervise and monitor the status of each feeder uplink in
advance.
6.2.3 Method to Decide the Best NGWs
In order to evaluate the status of each feeder uplink, a CNIR threshold, ⇠th, leads
to understanding how well the feeder uplink signal works between the NGW and
satellite, i.e., when the CNIR above the threshold, in addition to knowing if the rain
affects each feeder uplink, i.e., when the CNIR below the threshold. However, an
adequate CNIR threshold could result in a challenge in this study. For this purpose,
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 94
the predicted CNIR matrix, Ξ̂, is evaluated by the CNIR threshold. By each GW-site,
it is important to mention that each CNIR prediction is implicit in each column i of
the matrix. Therefore, if Ξ̂t+1,i ⇠th for i = 1, 2, . . . , N , and t = 1, 2, . . . , K, then
the feeder uplink is active at that time step, otherwise, the feeder uplink is inactive.
Based on the previous statement, it is possible to find the availability of each feeder
uplink, which is denoted by
⌫i =
1
K
K
X
k=1
⇥
(Ξ̂k,i ⇠th ! x = 1) ^ (Ξ̂k,i < ⇠th ! x = 0)
⇤
, for i = 1, 2, . . . , N, (6.3)
where x is a temporary variable. As a result, GWs can be sorted from the highest
to the lowest according to individual availability, ⌫i. Thus, each i column of the
matrix Ξ̂ 2 R
K⇥N is also sorted by individual availability, ⌫i. This method allows
us to define the best candidates in order to set up the ground network system with
NGWs. Therefore, it is essential to find system availability regarding its network
configuration. In this context, the system availability, [102], is given by
⌫ = ⌫1 · ⌫2·, . . . , ·⌫N̄ =
N̄
Y
i=1
⌫i, (6.4)
where N̄ corresponds to the number of NGWs, comprising the ground network
configuration. However, it is highly probable that the system availability is lower than
99.9%, dropping even more when increasing the number of NGWs. Indeed, the ITU
Recommendation, ITU-R S.1557, indicates that for gateway stations at geographically
dispersed hubs must require the end-to-end system availability of at least 99.9% [16].
For this reason, it is necessary to add redundancy to the ground network, increasing
the system availability equal to or greater than 99.9%.
6.2.4 A Strategy to Allocate the Best PGWs
It is necessary to mention that the sorted matrix Ξ̂ contains the possible NGWs and
PGWs for this study. The first N̄ columns correspond to the number of NGWs
of the ground network, whereas the next columns correspond to the number of
possible PGWs to be added to the ground network by this strategy. Therefore, in
general terms, the columns for redundant gateways are denoted by Ξ̂t+1,N̄+j, for
j = 1, 2, . . . , P̄ , where P̄ is the number of PGWs.
To be precise, this procedure consists of adding a redundant gateway, PGW, when
an NGW feeder uplink is an outage due to the rain. For this purpose, a switching
strategy is implemented between NGW and PGW by a straightforward mechanism.
That is, when an NGW is going to experience an outage in its feeder uplink at
time step t + 1, i.e., Ξ̂t+1,i < ⇠th, for i = 1, 2, . . . , N̄ , then the affected NGW will
commute all its data traffic to an available PGW. Also, this mechanism is able to
verify if the assigned PGW feeder uplink is active, i.e., Ξ̂t+1,N̄+1 ⇠th. Otherwise,
the mechanism goes to the next candidate, as long as Ξ̂t+1,N̄+2 ⇠th, and so on
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 95
until reaching the condition. However, the picked PGW could still be working with
another affected NGW. Therefore, the mechanism is also capable of looking for the
next PGW candidate. When the initial NGW is active again, then the mechanism
commutes all data traffic from the current PGW towards the initial NGW. To sum up,
this switching mechanism can be called the 1 + P̄ configuration, where each NGW is
backed up by redundant gateways P̄ to increase the feeder uplink availability. Figure
6.1 depicts the flowchart of the 1+ P̄ switching strategy for each NGW that explains
the described process in this proposal in detail. In other words, the multiple 1 + P̄
mechanism, running N̄ times, provides a scheme N̄ + P̄ as a result.
The number of feeder uplink outages can be individually accounted for by the
expression (1 ⌫i) ·K, for i = 1, 2, . . . , N̄ , where K is the total number of samples
at a time period. Based on the above formulation, the total number of feeder uplink
outages in the system is given by
outages = K ·
N̄
X
i=1
(1 ⌫i) (6.5)
In order to decrease the network complexity, the system unavailability only with
NGWs must be equal or lower than 5% in a year, i.e. (1 ⌫) ·100 5%, reducing the
number of switchings between NGWs and PGWs. As a result, the value of the CNIR
threshold, ⇠th, is set up based on the system’s unavailability percentage. Finally, used
the 1 + P̄ strategy, the feeder-uplink availability expression can be denoted by
⌫Bi
=
1
K
K
X
k=1
n
(Ξ̂k,i ⇠th ! x = 1)^
^
⇥
(Ξ̂k,N̄+1 ⇠th ! x = 1) ^ (Ξ̂k,N̄+2 ⇠th ! x = 1) ^ . . .
^(Ξ̂k,N̄+P̄ ⇠th ! x = 1) ^ (Ξ̂k,N̄+P̄ < ⇠th ! x = 0)
⇤
o
, for i = 1, 2, . . . , N̄
(6.6)
Indeed, this expression determines how many redundant PGW gateways, denoted
by P̄ , can provide instant support to an affected NGW at an annual period, improving
the feeder uplink availability as much as possible. Finally, the total system availability
is computed by the Eq. (6.4), so that ⌫B = ⌫B1 · ⌫B2 · . . . · ⌫BN̄
. Naturally, this leads
to determine the necessary number of P̄ to aid in the ground network to maintain it
above 99.9%, finding an optimized N̄ + P̄ scheme.
6.3 The 1 + P̄ Scheme Analyzed by a Markov Chain
The 1+P̄ scheme was also analyzed using a stationary distribution by a Markov Chain.
Therefore, it was necessary to find the probability of rain for each GW location. For
this purpose, the ITU Recommendation ITU-R P. 837 models the characteristics of
precipitation for propagation, whose aim is to calculate the rainfall rate exceeded
for the desired average annual probability of exceedance and a given location of the
Earth [29].
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 96
Start (k, i = 1)
Collect all CNIRs
Ξ̂ 2 R
K⇥N
Sort Ξ̂ according to ⌫
⌫1 ⌫2 . . . ⌫N
NGW scheme
N̄ = 4, 8, 12 or 16
Ξ̂k,i ⇠th
j = 1
Ξ̂k,N̄+j < ⇠th
_
Pj is busy
j = j+1
N̄ + j > N
Switching
Unreachable
i = i+1 i > N̄
k = k + 1
i = 1
k > K
Stop
Switching between
N̄i and P̄j
No
Yes
Yes Yes
No
Yes
No
No
Yes
No
Figure 6.1: Flowchart of the multiple 1 + P̄ switching strategy.
6.3.1 Method to Calculate the Probability of Rain, P0
To begin with, the ITU Recommendation ITU-R P.837 includes digital maps with
information about the monthly mean total rainfall data, MTii, and the annual rainfall
rate data exceeded for 0.01% of an average year, R0.01. In both cases, these maps
are georeferenced by grids of geographical coordinates. For the MTii map, each
coordinate of latitude and longitude is separated by steps with a resolution of 0.25 ,
whereas a resolution of 0.125 for the R0.01 map.
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 97
In this context, the ITU-R P.837 method enables three input parameters, which
are the desired annual probability of exceedance, pd [%], the latitude of the desired
location Lati [degrees], and the longitude of the desired location Loni [degrees]. Table
A.2 shows the latitude, Lati, and longitude, Loni, for each i = 1, 2, . . . , N . Also, the
method generates two output parameters, which are the rainfall rate exceeded for the
desired probability of exceedance, Rp [mm/h], and the annual probability of rain, P0i
[%]. The ITU-R P.837 method [29] is discussed and developed in the following.
• Step 1: Table 6.1 indicates the representations of months to employ in this
method. Let ii the month number and let Nii the number of days in each
month.
Table 6.1: Month Numbers and Number of Days
Month Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
ii 01 02 03 04 05 06 07 08 09 10 11 12
Nii 31 28.25 31 30 31 30 31 31 30 31 30 31
• Step 2: Determine the monthly mean surfaces temperatures, Tii [K], at the i-th
GW location (Lati, Loni), where ii = {01, 02, 03, . . . , 12} and i = 1, 2, . . . , N .
In this case, the monthly mean surface temperatures are obtained from the
digital maps included in ITU Recommendation ITU-R P.1510 [34].
• Step 3: Likewise, determine the monthly mean total rainfall, MTii, at
the i-th GW location (Lati, Loni), from the digital maps included in this
ITU Recommendation. However, to determine the MTii at the specific site
(Lati, Loni), it is necessary to use bilinear interpolation, which gives by the
ITU Recommendation ITU-R P.1144 in more detail [33].
• Step 4: Convert the temperature Tii[K] to tii[
C], for each month number, ii.
• Step 5: Calculate rii [mm/h] for each month number as follows
rii = 0.5874 · exp (0.0833 · tii) for tii 0 C
rii = 0.5874 for tii < 0 C
(6.7)
• Step 6a: Calculate the monthly probability of rain [%] for each month number,
ii, as follows
P0ii = 100 ·
MTii
24 ·Nii · rii
(6.8)
• Step 6b: if P0ii > 70, then set P0ii = 70 and rii =
100
70
·
MTii
24 ·Nii
• Step 7: For each i-th GW location, calculate the annual probability of rain,
P0i = P (R > 0) [%] as follows
P0i =
P12
ii=1 Nii · P0ii
365.25
, for i = 1, 2, . . . , N (6.9)
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 98
• Step 8: In this case, it is not necessary to find Rp for this study. For further
details, consult the ITU-R P.837 [29].
It is important to note that the probabilities of rain P0i and P rain
i are the same but
using different notations. The first one is well-known by the ITU-R P.837, whereas
the ITU-R P.618 uses the second one often [28,29].
6.3.2 The Switching Strategy, 1 + P̄ , from the perspective of
a Markov Chain
Calculated the annual probability of rain, P0i(Lati, Loni, pd), for each i-th GW-site
by Table A.2 coordinates, the GWs can be sorted from the best to the worst regarding
no-rain probabilities by each site, so that 1 P01 1 P02 . . . 1 P0N . Likewise,
the best first N̄ sites are assigned to be part of the NGWs scheme, whereas the rest
of the sites are becoming part of the PGWs as those are added to the ground network
later. Although this analysis is using the Markov Chain, the switching mechanism is
similar to the scheme 1 + P̄ , as was discussed in subsection 6.2.4 in more detail.
Thus, each GW has three states, the active state, non-active state, and the
switching state. The active state is related to the no-rain probability at each GW-site,
whereas the non-active state occurs as one of the NGWs of the ground network
experiences an outage due to the rain, i.e., there is a total outage probability of
the system, Pout. At this point, the switching state works as a link between the
affected NGW and the picked PGW. It is important to note that there is always a
strong emphasis on the first PGW. Nonetheless, if the mentioned PGW is busy or
in a feeder-uplink outage, the transition probability goes to the next switching state
towards an available PGW, and so on.
As a result, the first three states correspond to the analyzed NGW, whereas the
next states, in threes, are for each PGW. Figure 6.2 shows the transition probabilities
from one state to another by a Markov chain graph. Also, as the Markov chain is
a discrete-time process, it can be represented by a state transition matrix P, for
i = 1, 2, . . . , N̄ and j = 1, 2, . . . , P̄ , as can be seen in Eq. (6.10).
In general terms, the stationary distribution is the fraction of time that the system
spends in each state as the number of samples tends to the infinity. As a result, if
there are 3(1+P̄ ) states, then the stationary distribution is a vector of length 3(1+P̄ ).
Indeed, the stationary distribution is usually referred to as π and is given by
π = s0 ·P
n, n!1 (6.11)
where n is the number of iterations and s0 is the initial state vector s0 =
[1, 0, 0, . . . , 0] 2 [0, 1]3(1+P̄ ). In this context, a Markov chain is stationary with
stationary distribution π if π · P = π, and
P3(1+P̄ )
i=1 πi = 1, where π =
[⇡1, ⇡2, . . . , ⇡3(1+P̄ )]. Therefore, the switching probability is calculated as
Psw = ⇡3 + ⇡6 + . . .+ ⇡3(1+j), for j = 0, 1, 2 . . . , P̄ (6.12)
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 99
Figure 6.2: The 1 + P̄ strategy represented by a Markov Chain graph.
Finally, the switching rate is calculated by Rsw = Psw/Tsw, where Tsw is the interval
between switching instants [21,98]. This switching rate is important to determine the
effectiveness of the switching strategy in addition to finding the optimal number of
PGWs necessary to maintain the availability of the ground network above 99.9%.
6.4 Results
To begin with, the smart method to predict both rain attenuation time-series and
CNIR time-series at each GW was trained and validated in Chapter 5 by Experiment
2. This method found a model with an average accuracy of 87.83%, generating
high-reliability results in comparison with other models.
Each ground network configuration has a different capacity, depending on the
number of NGWs. By using Eq. (6.1), it was possible to find the feeder uplink
capacity at each GW location. For this purpose, the V band provided 4 GHz of
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 100
P
=
s
1
s
2
s
3
s
4
s
5
s
6
s
7
s
8
s
9
·
·
·
s
3
j
+
1
s
3
j
+
2
s
3
j
+
3
s
1
1
P
0
i
P
o
u
t
P
0
i
P
o
u
t
0
0
0
0
0
0
·
·
·
0
0
0
s
2
1
P
0
i
P
o
u
t
P
0
i
P
o
u
t
0
0
0
0
0
0
·
·
·
0
0
0
s
3
0
0
0
1
P
0
N̄
+
1
P
o
u
t
P
0
N̄
+
1
P
o
u
t
0
0
0
·
·
·
0
0
0
s
4
0
0
0
1
P
0
N̄
+
1
P
o
u
t
P
0
N̄
+
1
P
o
u
t
0
0
0
·
·
·
0
0
0
s
5
0
0
0
1
P
0
N̄
+
1
P
o
u
t
P
0
N̄
+
1
P
o
u
t
0
0
0
·
·
·
0
0
0
s
6
1
P
0
i
P
o
u
t
0
0
0
0
P
0
i
P
o
u
t
0
0
·
·
·
0
0
0
s
7
0
0
0
0
0
0
1
P
0
N̄
+
2
P
o
u
t
P
0
N̄
+
2
P
o
u
t
·
·
·
0
0
0
s
8
0
0
0
0
0
0
1
P
0
N̄
+
2
P
o
u
t
P
0
N̄
+
2
P
o
u
t
·
·
·
0
0
0
s
9
1
P
0
i
P
o
u
t
0
0
0
0
0
0
0
·
·
·
P
0
i
P
o
u
t
0
0
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
. . .
s
3
j
+
1
0
0
0
0
0
0
0
0
0
·
·
·
1
P
0
N̄
+
j
P
o
u
t
P
0
N̄
+
j
P
o
u
t
s
3
j
+
2
0
0
0
0
0
0
0
0
0
·
·
·
1
P
0
N̄
+
j
P
o
u
t
P
0
N̄
+
j
P
o
u
t
s
3
j
+
3
1
P
0
i
P
o
u
t
P
0
i
P
o
u
t
0
0
0
0
0
0
·
·
·
0
0
0
(6
.1
0)
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 101
available spectrum, Savail. Furthermore, the feeder uplink models employed the 1F2P
scheme so that they did not use frequency reuse, Nsb = 1, however, they employed
double polarization, Np = 2. In order to determine the maximum capacity, each
feeder uplink was defined to work under clear sky conditions. Therefore, the best
MODCOD, 256 APSK 3/4, based on DVB-S2X, has a spectral efficiency of 5.6199
b/s/Hz with a roll-off of 5%, as can be seen in Chapter 3. As a result, all feeder
uplinks obtained an individual capacity of 44.96 Gb/s.
Consequently, the inbound capacity was calculated by Eq. (6.2) and using schemes
N̄ = 4, 8, 12, and 16. To be precise, Table 6.2 indicates the inbound capacity of an
HTS system for each N̄ scheme with a traffic ratio of [1 : 1].
Table 6.2: The Inbound and Total Capacity of the Ground Network Segment.
Scheme
N̄
Inbound
Capacity
[Gb/s]
Total
Capacity
[Gb/s]
4 179.84 ⇡ 360.00
8 359.68 ⇡ 720.00
12 539.52 ⇡ 1000.00
16 719.36 ⇡ 1400.00
The matrix Ξ̂ 2 R
K⇥N contains the predicted CNIR time-series of each GW
obtained from the deep learning model. To be specific, 24 GW locations (Table A.2)
were trained and validated by the machine learning model, as can be seen in Chapter
5. Thus, the availability, ⌫i, of each feeder uplink was calculated by Eq. (6.3). For
this purpose, the CNIR threshold varied from 0–29 dB to verify the feeder uplink
behavior at each step, in addition to sorting the obtained availability from the best
to the worst for each N̄ configuration scheme. The best N̄s, according to availability,
were part of each ground network configuration, finding the system availability, ⌫, for
each N̄ configuration scheme by using Eq. (6.4).
Figure 6.3 depicts the unavailability percentage, 1 ⌫ [%], vs. the CNIR threshold
range, ⇠th, for all configuration schemes, N̄ = 4, 8, 12, 16. In this case, ⇠th = 14 dB was
a good measure as for all N̄ schemes, the system unavailability percentage was lower
than 5% per year. For 1 ⌫ > 5%, the system unavailability increased considerably,
so that the number of outages of the system could be tough to manage.
The capacity of the feeder uplink is also affected by the rain. For this reason, the
CNIR threshold must be capable of balancing as the system available as the capacity
of the feeder uplink. Figure 6.4 shows how the threshold ⇠th = 14 dB indicates the
limit of the system availability, ⌫, not to be lower than 0.95, and the capacity of
feeder uplinks, on average, must not be lower than 44.70 Gb/s.
Here the desired availability, ⌫i of each feeder uplink for the GW set, [N ] =
{1, 2, . . . , 24}, was calculated using the threshold ⇠th = 14 dB by Eq. (6.3). Then, the
GW locations were sorted from the highest to the lowest according to the availability.
Table 6.3 shows the sorted availabilities where the best N̄ GWs correspond to the
number of GWs according to the ground network configuration, N̄ = 4, 8, 12, 16.
As a result, none of the four configuration schemes reached the minimum system
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 102
Figure 6.3: Unavailability percentage per year, for all N̄ configuration schemes.
Figure 6.4: System availability vs. the feeder uplink capacity, on average, by ranging the
CNIR threshold, ⇠th.
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 103
availability, established in 99.90%. Therefore, all schemes must add redundancy to
outperform the required minimum availability.
Table 6.3: Gateway Availabilities.
Locations
[N ] = {1, 2, . . . , 24}
Availability
⌫i
N̄ = 4 N̄ = 8 N̄ = 12 N̄ = 16
Torreon 0.9994 N̄1 N̄1 N̄1 N̄1
Cd. Juarez 0.9992 N̄2 N̄2 N̄2 N̄2
La Paz, B.C. 0.9992 N̄3 N̄3 N̄3 N̄3
Tijuana 0.9990 N̄4 N̄4 N̄4 N̄4
Oaxaca 0.9986 P̄N̄+1 N̄5 N̄5 N̄5
Queretaro 0.9985 P̄N̄+2 N̄6 N̄6 N̄6
Mexico City 0.9982 P̄N̄+3 N̄7 N̄7 N̄7
Guadalajara 0.9977 P̄N̄+4 N̄8 N̄8 N̄8
Monterrey 0.9968 P̄N̄+5 P̄N̄+1 N̄9 N̄9
Tegucigalpa 0.9963 P̄N̄+6 P̄N̄+2 N̄10 N̄10
Merida 0.9961 P̄N̄+7 P̄N̄+3 N̄11 N̄11
Tuxtla, Gtz. 0.9953 P̄N̄+8 P̄N̄+4 N̄12 N̄12
La Habana 0.9949 P̄N̄+9 P̄N̄+5 P̄N̄+1 N̄13
Kingston 0.9942 P̄N̄+10 P̄N̄+6 P̄N̄+2 N̄14
San Jose 0.9941 P̄N̄+11 P̄N̄+7 P̄N̄+3 N̄15
Cancun 0.9938 P̄N̄+12 P̄N̄+8 P̄N̄+4 N̄16
Veracruz 0.9934 P̄N̄+13 P̄N̄+9 P̄N̄+5 P̄N̄+1
San Pedro Sula 0.9932 P̄N̄+14 P̄N̄+10 P̄N̄+6 P̄N̄+2
Belmopan 0.9925 P̄N̄+15 P̄N̄+11 P̄N̄+7 P̄N̄+3
San Salvador 0.9917 P̄N̄+16 P̄N̄+12 P̄N̄+8 P̄N̄+4
St. Domingo 0.9911 P̄N̄+17 P̄N̄+13 P̄N̄+9 P̄N̄+5
San Juan 0.9902 P̄N̄+18 P̄N̄+14 P̄N̄+10 P̄N̄+6
Panama 0.9898 P̄N̄+19 P̄N̄+15 P̄N̄+11 P̄N̄+7
Pto. Cabezas 0.9889 P̄N̄+20 P̄N̄+16 P̄N̄+12 P̄N̄+8
System Availability, ⌫ [%] 99.68% 98.98% 97.46% 95.23%
In this context, the proposed strategy was carried out in this analysis by the
switching mechanism, 1+P̄ . Table 6.4 indicates the number of feeder uplink outages of
each NGW, backing up by one PGW, at a year-round. Table 6.5 shows the combined
availability between NGW and PGW, proving the increase of availability of each
feeder uplink. As a result, only one redundant GW was necessary to overcome the
required system availability, picking out Oaxaca as the backup GW. Therefore, the
optimized scheme was 4 + 1.
For the scheme 8+ P̄ , two PGWs (Monterrey and Tegucigalpa) were necessary to
back up the ground network about 100%. Table 6.6 displays the number of outages
of each NGW, where P̄1 (Monterrey) backs up 99% of feeder uplink outages from all
NGWs. Table 6.7 shows the system availability with and without redundancy. In
this case, only one PGW (Monterrey) was needful to overcome the required system
availability (99.9%), obtaining a 99.99%. Therefore, the 8+ 1 scheme met the aim of
improving the required system availability.
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 104
Table 6.4: The Number of Outages of Each NGW for the 4 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, 3, 4}
Oaxaca
P̄1
Torreon 293
Cd. Juarez 412
La Paz, B.C. 444
Tijuana 549
Total Outages 1698
Table 6.5: The System Availability for the 4 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, 3, 4}
4 + 0 4 + 1
Torreon 0.9994 1.0000
Cd. Juarez 0.9992 1.0000
La Paz, B.C. 0.9992 1.0000
Tijuana 0.9990 1.0000
⌫B [%] 99.68% 100.0%
Table 6.6: The Number of Outages of Each NGW for the 8 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 8}
Monterrey
P̄1
Tegucigalpa
P̄2
Torreon 292 1
Cd. Juarez 412 0
La Paz, B.C. 444 0
Tijuana 549 0
Oaxaca 730 0
Queretaro 793 20
Mexico City 907 24
Guadalajara 1190 6
Total Outages 5317 51
Table 6.7: The System Availability for the 8 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 8}
8 + 0 8 + 1 8 + 2
Torreon 0.9994 0.9999 1.0000
Cd. Juarez 0.9992 1.0000 1.0000
La Paz, B.C. 0.9992 1.0000 1.0000
Tijuana 0.9990 1.0000 1.000
Oaxaca 0.9986 1.0000 1.0000
Queretaro 0.9985 0.9999 1.0000
Mexico City 0.9982 0.9999 1.0000
Guadalajara 0.9977 0.9999 1.0000
⌫B [%] 98.98% 99.99% 100.0%
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 105
Likewise, Table 6.8 indicates the feeder uplink outages for the scheme 12 + P̄ ,
where two PGWs (La Habana and Kingston) support all outages. Naturally, the P̄1
(La Habana) backed up 98.9%of the feeder uplink outages for all NGWs. Table 6.9
exhibits the system availability for 12 + P̄ schemes. The switching strategy gave two
PGWs as a result, but only one PGW (La Habana) was useful to improve the system
availability of the ground network. For this reason, the 12 + 1 scheme overcame the
required system availability, providing a 99.97%.
Table 6.8: The Number of Outages of Each NGW for the 12 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 12}
La Habana
P̄1
Kingston
P̄2
Torreon 292 1
Cd. Juarez 412 0
La Paz, B.C. 444 0
Tijuana 547 2
Oaxaca 730 0
Queretaro 810 3
Mexico City 898 33
Guadalajara 1189 7
Monterrey 1668 30
Tegucigalpa 1892 30
Merida 2052 7
Tuxtla, Gtz. 2419 40
Total Outages 13353 153
Table 6.9: The System Availability for the 12 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 12}
12 + 0 12 + 1 12 + 2
Torreon 0.9994 0.9999 1.0000
Cd. Juarez 0.9992 1.0000 1.0000
La Paz, B.C. 0.9992 1.0000 1.0000
Tijuana 0.9990 0.9999 1.0000
Oaxaca 0.9986 1.0000 1.0000
Queretaro 0.9985 0.9999 1.0000
Mexico City 0.9982 0.9999 1.0000
Guadalajara 0.9977 0.9999 1.0000
Monterrey 0.9968 0.9999 1.0000
Tegucigalpa 0.9963 0.9999 1.0000
Merida 0.9961 0.9999 1.0000
Tuxtla, Gtz. 0.9953 0.9999 1.0000
⌫B [%] 97.46% 99.97% 100.0%
In the last case, the results of the evaluated 16 + P̄ scheme were published
by Tables 6.10 and 6.11. In the same context, Table 6.10 shows the number of
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 106
feeder uplink outages in each NGW, whereas three PGWs (Veracruz, San Pedro Sula,
and Belmopan) give support to them. The first P̄1 (Veracruz) provided support
to the ground network in 97.72% of the time. Table 6.11 indicates the system
availability of each 16+P̄ configuration scheme. To be precise, the scheme needed two
PGWs (Veracruz and San Pedro Sula) to outperform the required system availability.
Consequently, the optimal scheme resulted in 16 + 2, providing a system availability
of 99.99%.
Table 6.10: The Number of Outages of Each NGW for the 16 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 16}
Veracruz
P̄1
San Pedro Sula
P̄2
Belmopan
P̄3
Torreon 293 0 0
Cd. Juarez 412 0 0
La Paz, B.C. 444 0 0
Tijuana 548 1 0
Oaxaca 705 25 0
Queretaro 808 5 0
Mexico City 905 23 3
Guadalajara 1173 23 0
Monterrey 1664 34 0
Tegucigalpa 1882 40 0
Merida 2044 15 0
Tuxtla, Gtz. 2409 50 0
La Habana 2653 39 1
Kingston 2973 95 0
San Jose 3025 82 1
Cancun 3124 148 0
Total Outages 25062 580 5
Finally, for better visualization, these results can be presented in an unavailability
graph. Figure 6.5 depicts the different N̄ + P̄ schemes, where the minimum reference
is 1⇥ 10 3 (99.9%). Hence, it is possible to perceive the best schemes that overcome
the reference 1⇥ 10 3, which are: 4 + 1, 8 + 1, 12 + 1, and 16 + 2. However, it
is impossible to have an ideal system availability, that is, an availability of 100.0%.
Although it is theoretically possible to obtain 100.0% of availability, a more realistic
scenario could be as the best unavailability reaches a value of 1⇥ 10 7.
For the Markov Chain method, it was necessary to obtain the probability of rain
of each GW-site, P0i , by the ITU Recommendation ITU-R P.837, as can be seen in
the subsection 6.3.1. Then, the operation (no-rain) probabilities were sorted from the
result of the expression 1 P0i , i.e., from the highest to the lowest. Table 6.12 shows
the sorted operation probability of each GW-site. Here, the best N̄ GWs became part
of the ground network according to the N̄ schemes, 4, 8, 12, and 16.
By applying the switching strategy of Section 6.3, the stationary distribution, π,
of each 1+P̄ scheme was obtained from the state transition matrix, Eq. 6.10. For this
purpose, the system outage probability, Pout, was set in 1⇥ 10 4. Table 6.13 presents
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 107
Table 6.11: The System Availability for the 16 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 16}
16 + 0 16 + 1 16 + 2 16 + 3
Torreon 0.9994 1.0000 1.0000 1.0000
Cd. Juarez 0.9992 1.0000 1.0000 1.0000
La Paz, B.C. 0.9992 1.0000 1.0000 1.0000
Tijuana 0.9990 0.9999 1.0000 1.0000
Oaxaca 0.9986 0.9999 1.0000 1.0000
Queretaro 0.9985 0.9999 1.0000 1.0000
Mexico City 0.9982 0.9998 0.9999 1.0000
Guadalajara 0.9977 0.9999 1.0000 1.0000
Monterrey 0.9968 0.9999 1.0000 1.0000
Tegucigalpa 0.9963 0.9999 1.0000 1.0000
Merida 0.9961 0.9999 1.0000 1.0000
Tuxtla, Gtz. 0.9953 0.9999 1.0000 1.0000
La Habana 0.9949 0.9998 0.9999 1.0000
Kingston 0.9942 0.9999 1.0000 1.0000
San Jose 0.9941 0.9998 0.9999 1.0000
Cancun 0.9938 0.9999 1.0000 1.0000
⌫B [%] 95.23% 99.89% 99.99% 100.0%
Figure 6.5: Performance of the N̄ + P̄ schemes, as a function of system unavailability and
the number of P̄ .
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 108
Table 6.12: The Operation Probability of Gateways.
Locations
[N ] = {1, 2, . . . , 24}
Operation
Probability
1 P0i
N̄ = 4 N̄ = 8 N̄ = 12 N̄ = 16
La Paz, B.C. 0.9962 N̄1 N̄1 N̄1 N̄1
Torreon 0.9945 N̄2 N̄2 N̄2 N̄2
Cd. Juarez 0.9920 N̄3 N̄3 N̄3 N̄3
Merida 0.9840 N̄4 N̄4 N̄4 N̄4
Monterrey 0.9835 P̄N̄+1 N̄5 N̄5 N̄5
Tijuana 0.9826 P̄N̄+2 N̄6 N̄6 N̄6
Queretaro 0.9797 P̄N̄+3 N̄7 N̄7 N̄7
Cancun 0.9752 P̄N̄+4 N̄8 N̄8 N̄8
Oaxaca 0.9738 P̄N̄+5 P̄N̄+1 N̄9 N̄9
La Habana 0.9729 P̄N̄+6 P̄N̄+2 N̄10 N̄10
Tegucigalpa 0.9728 P̄N̄+7 P̄N̄+3 N̄11 N̄11
Tuxtla, Gtz. 0.9727 P̄N̄+8 P̄N̄+4 N̄12 N̄12
Guadalajara 0.9723 P̄N̄+9 P̄N̄+5 P̄N̄+1 N̄13
Kingston 0.9714 P̄N̄+10 P̄N̄+6 P̄N̄+2 N̄14
Veracruz 0.9709 P̄N̄+11 P̄N̄+7 P̄N̄+3 N̄15
St. Domingo 0.9671 P̄N̄+12 P̄N̄+8 P̄N̄+4 N̄16
San Juan 0.9671 P̄N̄+13 P̄N̄+9 P̄N̄+5 P̄N̄+1
Mexico City 0.9645 P̄N̄+14 P̄N̄+10 P̄N̄+6 P̄N̄+2
San Salvador 0.9638 P̄N̄+15 P̄N̄+11 P̄N̄+7 P̄N̄+3
San Pedro Sula 0.9614 P̄N̄+16 P̄N̄+12 P̄N̄+8 P̄N̄+4
Panama 0.9592 P̄N̄+17 P̄N̄+13 P̄N̄+9 P̄N̄+5
Belmopan 0.9568 P̄N̄+18 P̄N̄+14 P̄N̄+10 P̄N̄+6
Pto. Cabezas 0.9473 P̄N̄+19 P̄N̄+15 P̄N̄+11 P̄N̄+7
San Jose 0.9356 P̄N̄+20 P̄N̄+16 P̄N̄+12 P̄N̄+8
System Operation Probability,
[%]
96.70% 89.28% 80.03% 70.97%
the stationary distribution of each NGW for the 4+P̄ scheme. Although each π has 63
states, only the first 12 states were necessary for this analysis as the next probabilities
were practically negligible. Table 6.14 displays the system operation probabilities for
each 4 + P̄ scheme. In this case, the scheme 4 + 2 overcame the required system
availability of 99.9% in terms of probability. The two picked redundant GWs were
Monterrey and Tijuana. It is important to note that the inactive states were not
considered to determine the operation probability.
Table 6.15 and Table 6.16 indicate the stationary distribution and the system
operation probabilities for the 8+P̄ Scheme, respectively. In this case, each stationary
distribution had 51 states, but only the first 12 states were useful to determine the
number of PGWs to be added to the ground network. As a result, the scheme 8 + 3
was the most suitable scheme (99.92%) to work above the reference (99.9%), in terms
of probability. The three picked redundant GWs were Oaxaca, La Habana, and
Tegucigalpa.
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 109
Table 6.13: The Stationary Distribution of Each NGW for the 4 + P̄ Scheme.
States N̄1 N̄2 N̄3 N̄4
s1 0.8114 0.7473 0.6705 0.4996
s2 8.1449⇥ 10 5 7.5146⇥ 10 5 6.7586⇥ 10 5 5.0776⇥ 10 5
s3 0.0030 0.0041 0.0053 0.0080
s4 0.1818 0.2432 0.3164 0.4768
s5 1.8485⇥ 10 5 2.4726⇥ 10 5 3.2171⇥ 10 5 4.8480⇥ 10 5
s6 0.0030 0.0041 0.0053 0.0080
s7 6.5697⇥ 10 4 1.2762⇥ 10 3 2.4081⇥ 10 3 7.3383⇥ 10 3
s8 6.5704⇥ 10 8 1.2764⇥ 10 7 2.4084⇥ 10 7 7.3390⇥ 10 7
s9 1.1356⇥ 10 5 2.2061⇥ 10 5 4.1626⇥ 10 5 1.2685⇥ 10 4
s10 2.1016⇥ 10 6 5.9290⇥ 10 6 1.6224⇥ 10 5 9.9978⇥ 10 5
s11 2.1018⇥ 10 10 5.9296⇥ 10 10 1.6226⇥ 10 9 9.9988⇥ 10 9
s12 4.2484⇥ 10 8 1.1986⇥ 10 7 3.2797⇥ 10 7 2.0211⇥ 10 6
Table 6.14: The System Operation Probabilities for the 4 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, 3, 4}
4 + 0 4 + 1 4 + 2 4 + 3
La Paz, B.C. 0.9962 0.9992 0.9999 0.9999
Torreon 0.9945 0.9986 0.9999 0.9999
Cd. Juarez 0.9920 0.9974 0.9999 0.9999
Merida 0.9840 0.9923 0.9998 0.9999
Operation
Probability, [%]
96.70% 98.76% 99.95% 99.96%
Table 6.17 and Table 6.18 provide the results of the stationary distribution and
the system operation probabilities for the 12 + P̄ Scheme, respectively. The total
number of states of each stationary distribution was 39. However, the first 15
states of each distribution were required to evaluate the operation probability of
the system. Consequently, the scheme 12 + 3 obtained an operation probability of
99.90%, achieving the minimum requirement in terms of probability. The three picked
redundant GWs were Guadalajara, Kingston, and Veracruz.
Finally, Table 6.19 and Table 6.20 give the outcomes obtained from the stationary
distribution and the system operation probabilities for the 16+P̄ Scheme, respectively.
Each stationary distribution vector had a length of 27 elements or states, where
the only first 15 states were necessary to obtain the operation probability of the
system. To sum up, the scheme 16 + 4 provided an operation probability of 99.91%,
which outperformed the reference of 99.9%, in terms of probability. The four picked
redundant GWs were San Juan, Mexico City, San Salvador, and San Pedro Sula.
It is important to note that all evaluated state-transition-matrices were ergodic
and irreducibles. In other words, a Markov chain is ergodic if it is both irreducible and
aperiodic. This condition is equivalent to the state transition matrix being a primitive
non-negative matrix. In this context, ergodicity means the long-term proportion of
time spent by the chain in state si, corresponding to the steady-state probability.
C
H
A
P
T
E
R
6
.
G
R
O
U
N
D
S
E
G
M
E
N
T
O
P
T
IM
IZ
A
T
IO
N
B
Y
U
S
IN
G
...
110
Table 6.15: The Stationary Distribution of Each NGW for the 8 + P̄ Scheme.
States N̄1 N̄2 N̄3 N̄4 N̄5 N̄6 N̄7 N̄8
s1 0.8714 0.8234 0.7623 0.6114 0.6041 0.5916 0.5526 0.5020
s2 8.7480⇥ 10 5 8.2793⇥ 10 5 7.6842⇥ 10 5 6.2140⇥ 10 5 6.1426⇥ 10 5 6.0208⇥ 10 5 5.6407⇥ 10 5 5.1473⇥ 10 5
s3 0.0033 0.0045 0.0060 0.0097 0.0099 0.0102 0.0112 0.0124
s4 0.1215 0.1667 0.2238 0.3630 0.3696 0.3810 0.4162 0.4613
s5 1.2474⇥ 10 5 1.7116⇥ 10 5 2.2982⇥ 10 5 3.7277⇥ 10 5 3.7961⇥ 10 5 3.9127⇥ 10 5 4.2740⇥ 10 5 4.7369⇥ 10 5
s6 0.0033 0.0045 0.0060 0.0097 0.0099 0.0102 0.0112 0.0124
s7 0.0005 0.0009 0.0018 0.0057 0.0060 0.0065 0.0084 0.0113
s8 4.5125⇥ 10 8 8.9921⇥ 10 8 1.7509⇥ 10 7 5.7431⇥ 10 7 6.0280⇥ 10 7 6.5392⇥ 10 7 8.3537⇥ 10 7 1.1296⇥ 10 6
s9 1.2197⇥ 10 5 2.4306⇥ 10 5 4.7327⇥ 10 5 1.5524⇥ 10 4 1.6294⇥ 10 4 1.7675⇥ 10 4 2.2580⇥ 10 4 3.0532⇥ 10 4
s10 1.6812⇥ 10 6 4.8653⇥ 10 6 1.3739⇥ 10 5 9.1128⇥ 10 5 9.8584⇥ 10 5 1.1256⇥ 10 4 1.6816⇥ 10 4 2.7742⇥ 10 4
s11 1.6813⇥ 10 10 4.8658⇥ 10 10 1.3740⇥ 10 9 9.1137⇥ 10 9 9.8594⇥ 10 9 1.1257⇥ 10 8 1.6818⇥ 10 8 2.7744⇥ 10 8
s12 4.5630⇥ 10 8 1.3205⇥ 10 7 3.7289⇥ 10 7 2.4734⇥ 10 6 2.6757⇥ 10 6 3.0550⇥ 10 6 4.5641⇥ 10 6 7.5296⇥ 10 6
Table 6.16: The System Operation Probabilities for the 8 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 8}
8 + 0 8 + 1 8 + 2 8 + 3
La Paz, B.C. 0.9962 0.9994 0.9999 0.9999
Torreon 0.9945 0.9990 0.9999 0.9999
Cd. Juarez 0.9920 0.9981 0.9999 0.9999
Merida 0.9840 0.9939 0.9998 0.9999
Monterrey 0.9835 0.9936 0.9998 0.9999
Tijuana 0.9826 0.9931 0.9998 0.9999
Queretaro 0.9797 0.9911 0.9997 0.9999
Cancun 0.9752 0.9880 0.9996 0.9999
Operation
Probability, [%]
89.28% 95.70% 99.84% 99.92%
C
H
A
P
T
E
R
6
.
G
R
O
U
N
D
S
E
G
M
E
N
T
O
P
T
IM
IZ
A
T
IO
N
B
Y
U
S
IN
G
...
111
Table 6.17: The Stationary Distribution of Each NGW for the 12 + P̄ Scheme.
States N̄1 N̄2 N̄3 N̄4 N̄5 N̄6 N̄7 N̄8 N̄9 N̄10 N̄11 N̄12
s1 0.8773 0.8310 0.7718 0.6240 0.6168 0.6044 0.5657 0.5153 0.5004 0.4918 0.4907 0.4902
s2 8.8067⇥ 10 5 8.3559⇥ 10 5 7.7801⇥ 10 5 6.3418⇥ 10 5 6.2714⇥ 10 5 6.1511⇥ 10 5 5.7747⇥ 10 5 5.2838⇥ 10 5 5.1388⇥ 10 5 5.0550⇥ 10 5 5.0449⇥ 10 5 5.0396⇥ 10 5
s3 0.0033 0.0045 0.0061 0.0099 0.0101 0.0104 0.0114 0.0127 0.0131 0.0133 0.0133 0.0133
s4 0.1156 0.1590 0.2142 0.3502 0.3568 0.3680 0.4028 0.4476 0.4607 0.4682 0.4691 0.4696
s5 1.1889⇥ 10 5 1.6355⇥ 10 5 2.2029⇥ 10 5 3.6018⇥ 10 5 3.6693⇥ 10 5 3.7845⇥ 10 5 4.1426⇥ 10 5 4.6035⇥ 10 5 4.7380⇥ 10 5 4.8155⇥ 10 5 4.8248⇥ 10 5 4.8297⇥ 10 5
s6 0.0033 0.0045 0.0061 0.0099 0.0101 0.0104 0.0114 0.0127 0.0131 0.0133 0.0133 0.0133
s7 0.0004 0.0009 0.0017 0.0056 0.0058 0.0063 0.0081 0.0110 0.0120 0.0126 0.0127 0.0127
s8 4.3083⇥ 10 8 8.6069⇥ 10 8 1.6813⇥ 10 7 5.5587⇥ 10 7 5.8367⇥ 10 7 6.3359⇥ 10 7 8.1107⇥ 10 7 1.0997⇥ 10 6 1.1995⇥ 10 6 1.2608⇥ 10 6 1.2683⇥ 10 6 1.2723⇥ 10 6
s9 1.2279⇥ 10 5 2.4530⇥ 10 5 4.7917⇥ 10 5 1.5843⇥ 10 4 1.6635⇥ 10 4 1.8058⇥ 10 4 2.3116⇥ 10 4 3.1342⇥ 10 4 3.4188⇥ 10 4 3.5934⇥ 10 4 3.6149⇥ 10 4 3.6263⇥ 10 4
s10 1.5841⇥ 10 6 4.5960⇥ 10 6 1.3020⇥ 10 5 8.7049⇥ 10 5 9.4207⇥ 10 5 1.0763⇥ 10 4 1.6113⇥ 10 4 2.6654⇥ 10 4 3.0815⇥ 10 4 3.3496⇥ 10 4 3.3831⇥ 10 4 3.4011⇥ 10 4
s11 1.5843⇥ 10 10 4.5964⇥ 10 10 1.3021⇥ 10 9 8.7057⇥ 10 9 9.4217⇥ 10 9 1.0764⇥ 10 8 1.6115⇥ 10 8 2.6657⇥ 10 8 3.0818⇥ 10 8 3.3499⇥ 10 8 3.3835⇥ 10 8 3.4014⇥ 10 8
s12 4.5936⇥ 10 8 1.3327⇥ 10 7 3.7754⇥ 10 7 2.5242⇥ 10 6 2.7318⇥ 10 6 3.1211⇥ 10 6 4.6725⇥ 10 6 7.7292⇥ 10 6 8.9357⇥ 10 6 9.7130⇥ 10 6 9.8104⇥ 10 6 9.8624⇥ 10 6
s13 5.2411⇥ 10 9 2.2083⇥ 10 8 9.0723⇥ 10 8 1.2266⇥ 10 6 1.3682⇥ 10 6 1.6453⇥ 10 6 2.8805⇥ 10 6 5.8133⇥ 10 6 7.1231⇥ 10 6 8.0072⇥ 10 6 8.1201⇥ 10 6 8.1806⇥ 10 6
s14 5.2416⇥ 10 13 2.2086⇥ 10 12 9.0732⇥ 10 12 1.2267⇥ 10 10 1.3684⇥ 10 10 1.6454⇥ 10 10 2.8808⇥ 10 10 5.8139⇥ 10 10 7.1238⇥ 10 10 8.0080⇥ 10 10 8.1209⇥ 10 10 8.1814⇥ 10 10
s15 1.7185⇥ 10 10 7.2408⇥ 10 10 2.9746⇥ 10 9 4.0218⇥ 10 8 4.4862⇥ 10 8 5.3946⇥ 10 8 9.4446⇥ 10 8 1.9061⇥ 10 7 2.3355⇥ 10 7 2.6254⇥ 10 7 2.6624⇥ 10 7 2.6823⇥ 10 7
Table 6.18: The System Operation Probabilities for the 12 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 12}
12 + 0 12 + 1 12 + 2 12 + 3 12 + 4
La Paz, B.C. 0.9962 0.9995 0.9999 0.9999 0.9999
Torreon 0.9945 0.9990 0.9999 0.9999 0.9999
Cd. Juarez 0.9920 0.9982 0.9999 0.9999 0.9999
Merida 0.9840 0.9941 0.9998 0.9999 0.9999
Monterrey 0.9835 0.9938 0.9998 0.9999 0.9999
Tijuana 0.9826 0.9933 0.9998 0.9999 0.9999
Queretaro 0.9797 0.9914 0.9997 0.9999 0.9999
Cancun 0.9752 0.9883 0.9996 0.9999 0.9999
Oaxaca 0.9738 0.9872 0.9996 0.9999 0.9999
La Habana 0.9729 0.9866 0.9995 0.9999 0.9999
Tegucigalpa 0.9728 0.9865 0.9995 0.9999 0.9999
Tuxtla, Gtz. 0.9727 0.9865 0.9995 0.9999 0.9999
Operation
Probability, [%]
80.03% 90.82% 99.66% 99.90% 99.91%
C
H
A
P
T
E
R
6
.
G
R
O
U
N
D
S
E
G
M
E
N
T
O
P
T
IM
IZ
A
T
IO
N
B
Y
U
S
IN
G
...
112
Table 6.19: The Stationary Distribution of Each NGW for the 16 + P̄ Scheme.
States N̄1 N̄2 N̄3 N̄4 N̄5 N̄6 N̄7 N̄8 N̄9 N̄10 N̄11 N̄12 N̄13 N̄14 N̄15 N̄16
s1 0.8942 0.8532 0.7999 0.6624 0.6555 0.6437 0.6064 0.5570 0.5423 0.5337 0.5327 0.5322 0.5283 0.5202 0.5158 0.4842
s2 8.9761⇥ 10 5 8.5790⇥ 10 5 8.0635⇥ 10 5 6.7323⇥ 10 5 6.6656⇥ 10 5 6.5511⇥ 10 5 6.1898⇥ 10 5 5.7118⇥ 10 5 5.5691⇥ 10 5 5.4863⇥ 10 5 5.4763⇥ 10 5 5.4710⇥ 10 5 5.4340⇥ 10 5 5.3550⇥ 10 5 5.3123⇥ 10 5 5.0063⇥ 10 5
s3 0.0033 0.0046 0.0063 0.0106 0.0108 0.0111 0.0123 0.0137 0.0142 0.0144 0.0145 0.0145 0.0146 0.0148 0.0150 0.0159
s4 0.0987 0.1367 0.1859 0.3114 0.3176 0.3282 0.3616 0.4052 0.4181 0.4256 0.4265 0.4269 0.4303 0.4373 0.4412 0.4683
s5 1.0203⇥ 10 5 1.4138⇥ 10 5 1.9224⇥ 10 5 3.2194⇥ 10 5 3.2837⇥ 10 5 3.3937⇥ 10 5 3.7387⇥ 10 5 4.1901⇥ 10 5 4.3234⇥ 10 5 4.4005⇥ 10 5 4.4098⇥ 10 5 4.4147⇥ 10 5 4.4490⇥ 10 5 4.5222⇥ 10 5 4.5616⇥ 10 5 4.8417⇥ 10 5
s6 0.0033 0.0046 0.0063 0.0106 0.0108 0.0111 0.0123 0.0137 0.0142 0.0144 0.0145 0.0145 0.0146 0.0148 0.0150 0.0159
s7 0.0004 0.0007 0.0014 0.0048 0.0050 0.0054 0.0070 0.0096 0.0105 0.0110 0.0111 0.0111 0.0114 0.0119 0.0123 0.0147
s8 3.5367⇥ 10 8 7.1172⇥ 10 8 1.4034⇥ 10 7 4.7527⇥ 10 7 4.9964⇥ 10 7 5.4348⇥ 10 7 7.0021⇥ 10 7 9.5743⇥ 10 7 1.0470⇥ 10 6 1.1021⇥ 10 6 1.1089⇥ 10 6 1.1125⇥ 10 6 1.1380⇥ 10 6 1.1942⇥ 10 6 1.2255⇥ 10 6 1.4708⇥ 10 6
s9 1.2515⇥ 10 5 2.5186⇥ 10 5 4.9662⇥ 10 5 1.6818⇥ 10 4 1.7681⇥ 10 4 1.9232⇥ 10 4 2.4778⇥ 10 4 3.3880⇥ 10 4 3.7051⇥ 10 4 3.9000⇥ 10 4 3.9240⇥ 10 4 3.9367⇥ 10 4 4.0272⇥ 10 4 4.2259⇥ 10 4 4.3366⇥ 10 4 5.2046⇥ 10 4
s10 1.2968⇥ 10 6 3.7901⇥ 10 6 1.0838⇥ 10 5 7.4224⇥ 10 5 8.0424⇥ 10 5 9.2073⇥ 10 5 1.3873⇥ 10 4 2.3143⇥ 10 4 2.6823⇥ 10 4 2.9199⇥ 10 4 2.9497⇥ 10 4 2.9656⇥ 10 4 3.0795⇥ 10 4 3.3361⇥ 10 4 3.4829⇥ 10 4 4.7264⇥ 10 4
s11 1.2970⇥ 10 10 3.7905⇥ 10 10 1.0839⇥ 10 9 7.4231⇥ 10 9 8.0432⇥ 10 9 9.2082⇥ 10 9 1.3874⇥ 10 8 2.3145⇥ 10 8 2.6826⇥ 10 8 2.9202⇥ 10 8 2.9500⇥ 10 8 2.9659⇥ 10 8 3.0798⇥ 10 8 3.3365⇥ 10 8 3.4832⇥ 10 8 4.7268⇥ 10 8
s12 4.6819⇥ 10 8 1.3683⇥ 10 7 3.9129⇥ 10 7 2.6797⇥ 10 6 2.9035⇥ 10 6 3.3241⇥ 10 6 5.0084⇥ 10 6 8.3553⇥ 10 6 9.6840⇥ 10 6 1.0542⇥ 10 5 1.0649⇥ 10 5 1.0707⇥ 10 5 1.1118⇥ 10 5 1.2044⇥ 10 5 1.2574⇥ 10 5 1.7063⇥ 10 5
s13 4.5534⇥ 10 9 1.9327⇥ 10 8 8.0148⇥ 10 8 1.1100⇥ 10 6 1.2396⇥ 10 6 1.4936⇥ 10 6 2.6318⇥ 10 6 5.3567⇥ 10 6 6.5801⇥ 10 6 7.4077⇥ 10 6 7.5134⇥ 10 6 7.5701⇥ 10 6 7.9793⇥ 10 6 8.9241⇥ 10 6 9.4781⇥ 10 6 1.4543⇥ 10 5
s14 4.5539⇥ 10 13 1.9329⇥ 10 12 8.0157⇥ 10 12 1.1101⇥ 10 10 1.2397⇥ 10 10 1.4938⇥ 10 10 2.6321⇥ 10 10 5.3572⇥ 10 10 6.5808⇥ 10 10 7.4085⇥ 10 10 7.5142⇥ 10 10 7.5709⇥ 10 10 7.9801⇥ 10 10 8.9250⇥ 10 10 9.4791⇥ 10 10 1.4545⇥ 10 9
s15 1.7515⇥ 10 10 7.4341⇥ 10 10 3.0830⇥ 10 9 4.2695⇥ 10 8 4.7682⇥ 10 8 5.7453⇥ 10 8 1.0124⇥ 10 7 2.0605⇥ 10 7 2.5311⇥ 10 7 2.8494⇥ 10 7 2.8901⇥ 10 7 2.9119⇥ 10 7 3.0693⇥ 10 7 3.4327⇥ 10 7 3.6458⇥ 10 7 5.5942⇥ 10 7
Table 6.20: The System Operation Probabilities for the 16 + P̄ Scheme.
Locations
[N̄ ] = {1, 2, . . . , 16}
16 + 0 16 + 1 16 + 2 16 + 3 16 + 4
La Paz, B.C. 0.9962 0.9995 0.9999 0.9999 0.9999
Torreon 0.9945 0.9992 0.9999 0.9999 0.9999
Cd. Juarez 0.9920 0.9984 0.9999 0.9999 0.9999
Merida 0.9840 0.9949 0.9998 0.9999 0.9999
Monterrey 0.9835 0.9946 0.9998 0.9999 0.9999
Tijuana 0.9826 0.9942 0.9998 0.9999 0.9999
Queretaro 0.9797 0.9925 0.9998 0.9999 0.9999
Cancun 0.9752 0.9897 0.9997 0.9999 0.9999
Oaxaca 0.9738 0.9888 0.9996 0.9999 0.9999
La Habana 0.9729 0.9882 0.9996 0.9999 0.9999
Tegucigalpa 0.9728 0.9881 0.9996 0.9999 0.9999
Tuxtla, Gtz. 0.9727 0.9881 0.9996 0.9999 0.9999
Guadalajara 0.9723 0.9878 0.9996 0.9999 0.9999
Kingston 0.9714 0.9872 0.9995 0.9999 0.9999
Veracruz 0.9709 0.9868 0.9995 0.9999 0.9999
St. Domingo 0.9671 0.9842 0.9994 0.9999 0.9999
Operation
Probability, [%]
70.97% 87.06% 99.50% 99.85% 99.91%
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 113
Meanwhile, the Markov chain is irreducible when the state si of the Markov chain is
accessible from another state sj by a finite sequence of the transition of si to sj with
positive probability [83].
In summary, Figure 6.6 depicts the obtained schemes N̄ + P̄ in terms of system
outage probability, Pout. Here, it is possible to appreciate the best schemes that
outperform the reference 1⇥ 10 3, which are: 4 + 2, 8 + 3, 12 + 3, and 16 + 4.
Figure 6.6: Performance of the N̄ + P̄ schemes, as a function of system outage probability,
Pout, and the number of P̄ .
6.5 Discussions
There are remarkable differences between the availabilities and probabilities of each
GW. For instance, the sorted list of Table 6.3 is different in comparison with the
sorted list of Table 6.12. Naturally, the difference lies in the methods to define the
availability (deterministic approach) and probability (probabilistic approach) at each
GW. Nevertheless, this difference is not a big problem as the analysis of each feeder
uplink was rigorous, following the steps aforementioned of both methods. Therefore,
both sorted lists are theoretically reliable according to their mathematical nature.
In both smart predictive and Markov chain 1 + P̄ strategy methods, the first
PGW backs up the most of feeder uplink outages of NGWs. For this reason, the
mechanism has a direct approach when an NGW feeder uplink is in an outage in
terms of assignation, backing up immediately the affected NGW with the first PGW.
On the other hand, if the first PGW has its feeder uplink in an outage or busy with
another NGW, then the mechanism picks out the next available PGW. Motivated by
this, the method encourages to avoid correlated rain in two or more GWs at the same
time by a separation between them based on the correlation coefficient ⇢a obtained
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 114
from ITU-R P.1815. As a result, all GWs are separated by a distance > 80 km,
reducing the risk of simultaneous rain in two or more sites at the same time, and
increasing the chances of the first PGW is available in most of the time. Moreover,
it was necessary to assume spatially independent and identically distributed (i.i.d)
feeder uplinks.
The main aim of this study is to define the most suitable ground network to
manage a satellite system of 1 Tb/s and beyond. According to Table 6.2, 12 NGWs
are needed to manage a satellite system of about 1 Tb/s. In a sense, for capacities
of 1 Tb/s and beyond, we have coined the term Ultra High Throughput Satellite
Systems (UHTS). To be specific, the ground networks of 4 and 8 NGWs are for HTS
systems, whereas the ground networks of 12 and 16 NGWs are for UHTS systems.
In this context, these 4 configurations of NGWs were tested by both switching
strategies to determine their effectiveness, as deterministically (smart predictive
method) as probabilistically (Markov chain method). Figure 6.7 shows the results
of both proposed methods in this study to find N̄ + P̄ schemes, in addition to
Gharanjik’s method [25]. The results obtained from the smart predictive method
were highly superior compared with the Markov Chain method and the Gharanjik’s
method. That is, the smart predictive method gave the following results: 4+1, 8+1,
12 + 1, and 16 + 2, whereas the Markov Chain method: 4 + 2, 8 + 3, 12 + 3, and
16 + 4. Gharanjik’s method provided linear results, that is, N̄/P̄ = 4, so that the
scheme results are the following: 4 + 1, 8 + 2, 12 + 3, and 16 + 4.
Figure 6.7: N̄ + P̄ configuration schemes obtained from the three aforementioned methods.
The efficiency of the smart predictive method is, on average, 41.67% and 58.34%
better than Gharanjik’s model and Markov chain model, respectively. Therefore, the
1 + P̄ switching strategy based on the smart method for forecasting rain attenuation
and CNIR at each GW is highly efficient and reliable, optimizing the ground segment
network as much as possible. Indeed, for a 1 Tb/s UHTS, the optimized ground
CHAPTER 6. GROUND SEGMENT OPTIMIZATION BY USING . . . 115
network is 12 + 1, rather than 12 + 3 obtained from Gharanjik’s method, reducing
the network complexity and generating economic savings on the ground segment.
Despite the ground network could be not optimized by the Markov chain method,
each stationary distribution obtained from this method gave us a panorama on the
switching probability between the affected NGW and the first PGW. No doubt, this
parameter could help to get more information about the ground network behavior
when there is a switching between an NGW and PGW under different weather
conditions. In other words, the switching probability could be determinant to define
the most suitable technology to implement on the ground network, that is, the network
infrastructure to communicate GWs to each other.
6.6 Contributions of the Research
The proposed methods and found results in this chapter will be submitted to a IEEE
Transactions on Reliability Journal to be reviewed and published later. Furthermore,
the previous study was exposed to a specialized conference: "2020 IEEE 11th Latin
American Symposium on Circuits and Systems, LASCAS 2020," at San Jose, Costa
Rica.
• Andres Cornejo, Salvador Landeros-Ayala, Jose M. Matias, and Ramon
Martinez. Applying Learning Methods to Optimize the Ground Segment for
HTS Systems. 2020 IEEE 11th Lat. Am. Symp. Circuits Syst. LASCAS 2020,
pages 1–4, 2020
Chapter 7
Conclusions, Future Work, and
Contributions
7.1 Conclusions
In this thesis, we designed, allocated, and optimized the ground segment by
using novel methods and procedures with excellent results. The obtained feeder
link was designed in the Q/V band, whose performance was determined by
numerical evaluations. Further, the designed satellite antenna was based on an
offset-parabolic-reflector antenna to evaluate the interferences that could influence
the satellite links, directly or indirectly. As a result, we found excellent levels of the
carrier-to-noise and interference (CNIR) of each analyzed feeder uplink in clear-sky
conditions, which reached out to the uplink expectations for our system. However,
for a more realistic scenario, each feeder uplink was evaluated by applying a dynamic
rain attenuation based on the ITU Recommendation P.1853, defining the upper and
lower limits of feeder uplinks operation under the rain conditions.
For this purpose, we have successfully developed a method to generate, train,
and validate rain-attenuation prediction models based on LSTM networks. Two
Experiments were carried out to determine the best models, where Experiment 2
was found to be the most complete and accurate. In this sense, Machine Learning
methods can be beneficial for solving a specific problem whether the adaptation is
correct. In this case, deep learning networks based on LSTM regression models were
adequately adapted to our study, thereby, they were able to predict rain attenuation
events more accurately. In order to validate our proposal, this method was compared
with other predictive methods, where our proposal obtained a higher accuracy and a
better performance.
Furthermore, real rain-attenuation data were difficult to obtain due to the lack
of historical database, however, ITU Recommendation (ITU-R P.1853) was able to
compute artificial rain-attenuation, providing a plethora of data. Indeed, a lot of
data is required to train LSTM networks due to they are notably ubiquitous. Here,
it was possible to split data into both training and validation subsets because each
rain attenuation time series had hundreds of thousands of samples.
116
CHAPTER 7. CONCLUSIONS, FUTURE WORK, AND CONTRIBUTIONS 117
The processing of data was essential for obtaining good results as the answer lies in
the data. For this reason, subsets were carefully processed in terms of normalization,
partitioning, training, and validation. Thus, the model accuracy, on average, is about
87.83%, where the best model reaches out to the accuracy of 92.24%. All evaluated
GW-sites indicate good performance, giving solid predictions of rain attenuation.
In this context, the found results are successful to determine in advance how
the feeder uplink will be affected by weather impairments. This certainly leads
to better decision-making under rain-attenuation events, thereby, mechanisms such
as ULPC, traffic switching between GWs with extra resources, and satellite-link
switching towards backup gateway stations, can be improved by CNIR predictions
obtained from this proposal in order to assure high availability on satellite links.
To be precise, we focus on the optimization of the ground network by the switching
strategies of traffic between GWs. In this context, two switching strategy mechanisms
were proposed by a smart-predictive-method based on the deep learning model and
a Markov chain method. As a result, the proposed smart-predictive method reaches
a superior efficiency, on average, of 41.67% vs Gharanjik’s method, and 58.34% vs
the proposed Markov chain method, optimizing the ground network with excellent
performance. For instance, for a UHTS system of 1 Tb/s, Gharanjik’s method
provided a ground network scheme of 12 + 3 to operate above the required network
availability of 99.9%. Meanwhile, the proposed smart-predictive method gave, as a
result, an optimized scheme of 12 + 1, which was 66.67% more efficient than the
previous configuration scheme. In other words, the success of the smart-predictive
method lies in the ability to detect, in advance, a feeder uplink in outage due to the
rain, anticipating the switching between the affected NGW and an available PGW.
To sum up, this prediction method leads to optimizing the ground segment by the
smart gateway diversity, reducing the number of gateway stations, that is, the network
complexity.
7.2 Future Work
As a result of this thesis, some interesting topics have been branched for future
research work in both Machine Learning and Satellite Communication fields, which
comprise,
• Although the artificial data were useful to develop this proposal, rain
attenuation from a historical database could help to tune up and to improve
models. Further studies are necessary to determine much more accurate models
by employing real and artificial data together.
• By using Cloud Computing, the prediction method could be accessed from any
GW station, to be specific, the prediction method could be implemented by
software as a service (SaaS). No doubt that the prediction method as a cloud
service will guarantee to manage data in real-time, in addition to providing
flexibility and security.
CHAPTER 7. CONCLUSIONS, FUTURE WORK, AND CONTRIBUTIONS 118
• Finally, the stationary distributions obtained from the Markov chain method
provide worth information (the switching rate) to define the best infrastructure
capable of transferring data traffic between GWs, reducing latency and delays.
In this context, the best option is the network infrastructure based on optical
fiber, but it is not the cheapest choice. Hence, other options could be considered
to simplify the terrestrial network and reduce costs by the analysis of the
switching probability and rate, e.g., microwave links.
7.3 Contributions and Production
The research results of each topic of this thesis have been published or have been
submitted for publications in journals and conference proceedings, as indicated in the
following. It is worth to mention that the main ideas, the problem formulations, the
analysis, simulations, and numerical assessments for all publications are contributions
of the author of this thesis.
Journals
• [103] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez, and Jose M
Matias. Analysis to Quantify and Optimize Spot Beams for a High Throughput
Satellite in Ka and Q/V Bands. IEEE Lat. Am. Trans., 17(02):219–227, feb
2019
• [104] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martínez
Rodríguez-Osorio, and Jose M Matias. A method for designing an
offset-parabolic-reflector antenna for a ultra-high throughput satellite. In
Rev. Tecnol. e Inf., volume 16, pages 84–87, Lima, Peru, 2018. Universidad
Inca Garcilaso de la Vega
• [105] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez-Rodriguez,
and Jose M Matias. Interference Evaluations in Frequency Reuse by Using
Offset- Parabolic-Reflector Antennas for a UHTS System. Int. J. Eng. Sci.,
7(8):34–45, 2018
• Andres Cornejo, Salvador Landeros-Ayala, Jose M. Matias, Flor Ortiz-Gomez,
Ramon Martinez, and Miguel Salas-Natera. Method of Rain Attenuation
Prediction Based On Long-Short Term Memory Network. In Review., 1(1):1–35,
2020
• Andres Cornejo, Salvador Landeros-Ayala, and Jose Maria Matias.
Optimization of the Ground Network Segment by Implementing Machine
Learning. To be Submitted, 1(1):30, 2020
Conferences
• [106] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez
Rodriguez-Osorio, and Jose Maria Matias. Optimization of the Ground
Segment for an UHTS System through Neural Networks. In XXXIII Simp.
CHAPTER 7. CONCLUSIONS, FUTURE WORK, AND CONTRIBUTIONS 119
Nac. la Unión Científica Int. Radio (URSI 2018), pages 1–5, Granada, Spain,
2018
• [107] Andres Cornejo, Salvador Landeros-Ayala, Jose M. Matias, and Ramon
Martinez. Applying Learning Methods to Optimize the Ground Segment for
HTS Systems. 2020 IEEE 11th Lat. Am. Symp. Circuits Syst. LASCAS 2020,
pages 1–4, 2020
• [108] Andres Cornejo, Flor Ortiz-Gomez, Jose M. Matias, Ramon Martinez,
and Miguel Salas-Natera. Quantum Cryptography for Satellite Communication
Systems. In Congr. Nac. Act. Espac. 2020-Agencia Espac. Mex., Mexico City,
2020
Finally, all code scripts employed in this thesis will remain in the Github repository
permanently, https://github.com/KunaCornejo. All code scripts were developed by
the author of this thesis using Python® and Matlab®, except for the ITU-R P.1853
script, which was written in Matlab by ONERA®, French Aerospace Lab.
Appendix A
List of Geographic Coordinates for
Gateway (GW) Locations
A.1 List of Geographic Coordinates for Gateway
(GW) Locations, Chapter 3
A.2 List of Geographic Coordinates for Gateway
(GW) Locations, Chapter 5
The clear sky carrier-to-noise ratio ( cs) for each feeder uplink was calculated by using
the geographic coordinates regarding its location, geostationary satellite position at
92.0 West, roll-off of 5%, and the frequency band at 50 GHz (V-band). Furthermore,
the transmitter power of each GW is about 16.0 W, and the antenna diameter of 5.0
m, as can be seen in Chapter 3.
A.3 Distance Matrix Between Pairs of Locations,
Chapter 6
The distance between a pair of gateway locations was calculated by their geographic
coordinates. For this purpose, the geographic coordinates for each GW location were
provided by Table A.2. Finally, each computed value was stored in the distance
matrix D.
120
APPENDIX A. LIST OF GEOGRAPHIC COORDINATES FOR GW . . . 121
Table A.1: List of Geographic Coordinates for GW Locations, Chapter 3
Location Country
Latitude
[deg]
Longitude
[deg]
Height
(a.m.s.l.)
[m]
Buenos Aires Argentina -34.611758 -58.430232 25.0
Santiago Chile -33.458205 -70.666972 567.0
Montevideo Uruguay -34.81721 -56.188972 43.0
Asuncion Paraguay -25.298939 -57.636595 43.0
La Paz Bolivia -16.503502 -68.124903 3625.0
Sao Paulo Brazil -23.572845 -46.623028 760.0
Lima Peru -12.052542 -77.043969 154.0
Guayaquil Ecuador -2.185092 -79.908501 4.0
Quito Ecuador -0.178182 -78.482904 2700.0
Rio de Janeiro Brazil -22.904847 -43.219689 11.0
Bogota Colombia 4.696189 -74.077603 2640.0
Caracas Venezuela 10.474799 -66.905474 900.0
Panama Panama 9.010642 -79.509318 2.0
San Jose Costa Rica 9.924622 -84.094463 1300.0
Tegucigalpa Guatemala 14.069427 -87.201883 1000.0
Mexico City Mexico 19.336688 -99.165822 2250.0
Monterrey Mexico 25.707672 -100.334175 530.0
Guadalajara Mexico 20.65853 -103.347001 1556.0
Cordoba Argentina -31.408778 -64.196749 390.0
Iquique Chile -20.232089 -70.134301 1.0
Manaus Brazil -3.11014 -60.000065 92.0
Belo Horizonte Brazil -19.925298 -43.940535 760.0
Tijuana Mexico 32.505354 -116.971508 31.0
Arequipa Peru -16.405163 -71.541072 2335.0
Cuzco Peru -13.527847 -71.964484 3399.0
Mar del Plata Argentina -38.030257 -57.604656 38.0
Rosario Argentina -32.953004 -60.687002 25.0
Porto Alegre Brazil -30.079685 -51.231157 10.0
Valdivia Chile -39.838558 -73.23594 14.0
Cali Colombia 3.4074 -76.533906 1018.0
Cartagena Colombia 10.401403 -75.51452 2.0
Pto. Concordia Colombia 1.252249 -70.232856 201.0
Maracaibo Venezuela 10.6808 -71.635416 15.0
Barinas Venezuela 8.603336 -70.229713 200.0
La Habana Cuba 23.104927 -82.286028 59.0
St. Domingo Dominican. Rep. 18.48892 -69.935325 14.0
Chiclayo Peru -6.784283 -79.858654 27.0
San Salvador El Salvador 13.697313 -89.219673 658.0
San Juan Puerto Rico 18.401754 -66.036714 7.9
Brasilia Brazil -15.811467 -47.874582 1172.0
Ushuaia Argentina -54.817684 -68.329799 58.0
Recife Brazil -8.043475 -34.955255 58.0
Pto. Baquerizo Ecuador -0.907203 -89.604353 2.0
APPENDIX A. LIST OF GEOGRAPHIC COORDINATES FOR GW . . . 122
Table A.2: List of Geographic Coordinates for GW Locations, Chapter 5
i Location Country
Latitude
[deg]
Longitude
[deg]
Height
(a.m.s.l.)
[m]
cs
[dB]
1 Panama Panama 9.010642 -79.509318 2.0 32.44
2 San Jose Costa Rica 9.924622 -84.094463 1300.0 32.46
3 Tegucigalpa Guatemala 14.069427 -87.201883 1000.0 32.45
4 Mexico City Mexico 19.336688 -99.165822 2250.0 32.39
5 Monterrey Mexico 25.707672 -100.334175 530.0 32.31
6 Guadalajara Mexico 20.65853 -103.347001 1566.0 32.36
7 Tijuana Mexico 32.505354 -116.971508 31.0 32.09
8 La Habana Cuba 23.104927 -82.286028 59.0 32.34
9 St. Domingo Dominican Rep. 18.48892 -69.935325 14.0 32.29
10 San Salvador El Salvador 13.697313 -89.219673 658.0 32.45
11 San Juan Puerto Rico 18.401754 -66.036714 7.9 32.24
12 Torreon Mexico 25.560636 -103.374418 1120.0 32.30
13 La Paz, B.C. Mexico 24.129867 -110.292943 27.0 32.27
14 Veracruz Mexico 19.170373 -96.181209 10.0 32.40
15 Cancun Mexico 21.12228 -86.840371 11.0 32.38
16 Queretaro Mexico 20.614965 -100.411033 2017.0 32.37
17 Tuxtla, Gtz. Mexico 16.705306 -93.173516 522.0 32.43
18 Cd. Juarez Mexico 31.761652 -106.526185 1137.0 32.19
19 San Pedro Sula Honduras 15.517251 -87.992356 83.0 32.43
20 Belmopan Belize 17.265146 -88.768523 76.0 32.42
21 Kingston Jamaica 18.013036 -76.811471 9.0 32.36
22 Merida Mexico 20.970434 -89.608201 10.0 32.39
23 Pto. Cabezas Nicaragua 14.047875 -83.407334 7.0 32.43
24 Oaxaca Mexico 17.113126 -96.74277 1555.0 32.42
APPENDIX A. LIST OF GEOGRAPHIC COORDINATES FOR GW . . . 123
D
=
P
a
n
a
m
a
S
a
n
J
o
s
e
T
e
g
u
c
ig
a
lp
a
M
e
x
ic
o
C
it
y
M
o
n
t
e
r
r
e
y
G
u
a
d
a
la
ja
r
a
T
ij
u
a
n
a
L
a
H
a
b
a
n
a
S
t
.
D
o
m
in
g
o
S
a
n
S
a
lv
a
d
o
r
S
a
n
J
u
a
n
T
o
r
r
e
o
n
P
a
n
a
m
a
5
1
3
.0
6
1
0
0
9
.0
8
2
4
0
6
.6
7
2
8
7
7
.8
5
2
8
6
5
.1
4
4
6
4
9
.5
6
1
5
9
4
.8
8
1
4
7
5
.4
9
1
1
7
9
.5
9
1
7
8
9
.5
7
3
1
2
1
.0
0
S
a
n
J
o
s
e
5
1
3
.0
6
5
7
1
.4
8
1
9
2
7
.8
3
2
4
5
1
.3
5
2
3
8
1
.3
1
4
2
0
1
.7
8
1
4
7
8
.1
4
1
7
9
7
.3
6
6
9
7
.8
7
2
1
6
0
.7
4
2
6
7
4
.7
8
T
e
g
u
c
ig
a
lp
a
1
0
0
9
.0
8
5
7
1
.4
8
1
4
0
1
.6
7
1
8
8
4
.2
2
1
8
6
1
.9
3
3
6
4
5
.4
0
1
1
3
0
.0
9
1
9
0
6
.2
8
2
2
1
.7
1
2
3
0
8
.7
3
2
1
1
6
.4
6
M
e
x
ic
o
C
it
y
2
4
0
6
.6
7
1
9
2
7
.8
3
1
4
0
1
.6
7
7
1
8
.5
0
4
6
0
.9
3
2
2
9
9
.1
8
1
7
9
7
.8
9
3
0
7
2
.6
4
1
2
3
1
.3
1
3
4
8
2
.0
9
8
1
5
.9
3
M
o
n
t
e
r
r
e
y
2
8
7
7
.8
5
2
4
5
1
.3
5
1
8
8
4
.2
2
7
1
8
.5
0
6
4
0
.2
8
1
7
8
1
.4
8
1
8
4
8
.7
7
3
2
2
4
.9
4
1
7
6
9
.1
7
3
6
1
5
.8
0
3
0
5
.2
2
G
u
a
d
a
la
ja
r
a
2
8
6
5
.1
4
2
3
8
1
.3
1
1
8
6
1
.9
3
4
6
0
.9
3
6
4
0
.2
8
1
8
8
6
.1
8
2
1
8
8
.1
7
3
5
0
2
.8
1
1
6
8
7
.3
2
3
9
0
9
.7
2
5
4
5
.1
0
T
ij
u
a
n
a
4
6
4
9
.5
6
4
2
0
1
.7
8
3
6
4
5
.4
0
2
2
9
9
.1
8
1
7
8
1
.4
8
1
8
8
6
.1
8
3
5
5
0
.1
5
4
9
2
7
.5
6
3
5
0
6
.1
1
5
2
9
6
.0
2
1
5
2
8
.9
8
L
a
H
a
b
a
n
a
1
5
9
4
.8
8
1
4
7
8
.1
4
1
1
3
0
.0
9
1
7
9
7
.8
9
1
8
4
8
.7
7
2
1
8
8
.1
7
3
5
5
0
.1
5
1
3
8
1
.9
0
1
2
7
5
.8
5
1
7
6
7
.3
6
2
1
5
1
.6
8
S
t
.
D
o
m
in
g
o
1
4
7
5
.4
9
1
7
9
7
.3
6
1
9
0
6
.2
8
3
0
7
2
.6
4
3
2
2
4
.9
4
3
5
0
2
.8
1
4
9
2
7
.5
6
1
3
8
1
.9
0
2
1
2
6
.6
1
4
1
1
.3
4
3
5
2
5
.3
5
S
a
n
S
a
lv
a
d
o
r
1
1
7
9
.5
9
6
9
7
.8
7
2
2
1
.7
1
1
2
3
1
.3
1
1
7
6
9
.1
7
1
6
8
7
.3
2
3
5
0
6
.1
1
1
2
7
5
.8
5
2
1
2
6
.6
1
2
5
2
9
.8
5
1
9
8
1
.3
4
S
a
n
J
u
a
n
1
7
8
9
.5
7
2
1
6
0
.7
4
2
3
0
8
.7
3
3
4
8
2
.0
9
3
6
1
5
.8
0
3
9
0
9
.7
2
5
2
9
6
.0
2
1
7
6
7
.3
6
4
1
1
.3
4
2
5
2
9
.8
5
3
9
1
8
.0
1
T
o
r
r
e
o
n
3
1
2
1
.0
0
2
6
7
4
.7
8
2
1
1
6
.4
6
8
1
5
.9
3
3
0
5
.2
2
5
4
5
.1
0
1
5
2
8
.9
8
2
1
5
1
.6
8
3
5
2
5
.3
5
1
9
8
1
.3
4
3
9
1
8
.0
1
L
a
P
a
z
,
B
.C
.
3
6
7
2
.7
3
3
1
9
2
.2
8
2
6
6
6
.2
5
1
2
6
6
.2
2
1
0
1
9
.2
1
8
1
1
.5
5
1
1
3
7
.1
9
2
8
5
0
.9
5
4
2
1
3
.3
0
2
4
9
7
.0
5
4
6
1
1
.5
6
7
1
5
.9
0
V
e
r
a
c
r
u
z
2
1
2
0
.8
0
1
6
5
6
.6
4
1
1
1
1
.8
3
3
1
3
.8
5
8
4
2
.7
8
7
6
7
.1
3
2
5
4
6
.6
2
1
5
0
5
.2
3
2
7
6
0
.6
8
9
5
9
.7
1
3
1
7
0
.5
8
1
0
2
5
.2
2
C
a
n
c
u
n
1
5
5
9
.0
0
1
2
7
9
.2
7
7
8
5
.1
8
1
3
0
0
.8
6
1
4
6
7
.2
4
1
7
1
4
.8
1
3
2
3
2
.3
7
5
1
8
.3
4
1
7
9
1
.7
0
8
6
3
.2
9
2
1
9
6
.2
9
1
7
5
7
.1
8
Q
u
e
r
e
t
a
r
o
2
5
8
6
.3
1
2
1
1
2
.8
5
1
5
7
8
.5
6
1
9
2
.7
1
5
6
6
.3
4
3
0
5
.5
5
2
1
0
7
.1
7
1
8
8
9
.5
8
3
1
9
7
.5
2
1
4
1
5
.2
5
3
6
0
4
.7
1
6
2
7
.8
7
T
u
x
t
la
,
G
t
z
.
1
7
0
9
.3
7
1
2
3
7
.8
1
7
0
4
.0
3
6
9
7
.8
3
1
2
4
5
.5
2
1
1
5
7
.8
9
2
9
6
6
.5
5
1
3
4
1
.5
7
2
4
6
9
.3
4
5
4
0
.2
0
2
8
8
0
.5
2
1
4
4
4
.0
7
C
d
.
J
u
a
r
e
z
3
7
6
3
.3
2
3
3
5
0
.4
9
2
7
8
0
.9
4
1
5
6
5
.3
4
9
0
3
.8
4
1
2
7
4
.4
8
9
8
6
.6
1
2
5
7
1
.2
8
3
9
4
4
.6
9
2
6
7
2
.3
6
4
3
1
0
.3
1
7
5
4
.8
7
S
a
n
P
e
d
r
o
S
u
la
1
1
7
1
.3
1
7
5
1
.8
7
1
8
2
.0
4
1
2
5
8
.7
6
1
7
1
0
.8
2
1
7
1
9
.6
0
3
4
7
8
.3
7
1
0
3
4
.2
5
1
9
4
7
.3
7
2
4
1
.6
4
2
3
5
5
.5
8
1
9
4
9
.7
0
B
e
lm
o
p
a
n
1
3
5
8
.5
3
9
5
9
.6
9
3
9
2
.9
3
1
1
2
1
.3
4
1
5
1
9
.4
9
1
5
7
8
.1
2
3
2
9
4
.9
0
9
3
7
.4
3
1
9
9
6
.7
9
3
9
9
.6
6
2
4
0
7
.9
8
1
7
6
9
.0
3
K
in
g
s
t
o
n
1
0
4
2
.5
5
1
1
9
3
.8
4
1
1
9
3
.4
7
2
3
5
7
.8
0
2
5
6
9
.0
7
2
7
9
6
.5
8
4
3
1
6
.7
4
8
0
3
.1
9
7
2
8
.0
1
1
4
1
0
.7
8
1
1
3
8
.7
7
2
8
6
2
.1
0
M
e
r
id
a
1
7
1
4
.5
8
1
3
6
2
.5
2
8
0
8
.6
0
1
0
1
3
.9
1
1
2
1
4
.5
3
1
4
2
7
.9
7
2
9
9
2
.6
4
7
9
1
.0
1
2
0
7
6
.1
4
8
0
9
.7
8
2
4
8
2
.0
4
1
4
9
4
.9
8
P
t
o
.
C
a
b
e
z
a
s
7
0
2
.8
2
4
6
4
.5
3
4
0
9
.3
0
1
7
7
7
.3
4
2
1
9
0
.0
3
2
2
3
7
.9
6
3
9
6
8
.9
2
1
0
1
3
.9
9
1
5
1
9
.7
9
6
2
8
.6
4
1
9
1
5
.6
6
2
4
4
4
.7
6
O
a
x
a
c
a
2
0
7
0
.7
9
1
5
8
2
.9
0
1
0
7
6
.2
3
3
5
5
.8
3
1
0
2
5
.2
6
7
9
8
.7
1
2
6
5
5
.0
0
1
6
4
8
.7
0
2
8
3
9
.7
1
8
9
1
.2
7
3
2
5
1
.0
4
1
1
6
3
.0
9
APPENDIX A. LIST OF GEOGRAPHIC COORDINATES FOR GW . . . 124
D
=
L
a
P
a
z
,
B
.C
.
V
e
r
a
c
r
u
z
C
a
n
c
u
n
Q
u
e
r
e
t
a
r
o
T
u
x
t
la
,
G
t
z
.
C
d
.
J
u
a
r
e
z
S
a
n
P
e
d
r
o
S
u
la
B
e
lm
o
p
a
n
K
in
g
s
t
o
n
M
e
r
id
a
P
t
o
.
C
a
b
e
z
a
s
O
a
x
a
c
a
P
a
n
a
m
a
3
6
7
2
.7
3
2
1
2
0
.8
0
1
5
5
9
.0
0
2
5
8
6
.3
1
1
7
0
9
.3
7
3
7
6
3
.3
2
1
1
7
1
.3
1
1
3
5
8
.5
3
1
0
4
2
.5
5
1
7
1
4
.5
8
7
0
2
.8
2
2
0
7
0
.7
9
S
a
n
J
o
s
e
3
1
9
2
.2
8
1
6
5
6
.6
4
1
2
7
9
.2
7
2
1
1
2
.8
5
1
2
3
7
.8
1
3
3
5
0
.4
9
7
5
1
.8
7
9
5
9
.6
9
1
1
9
3
.8
4
1
3
6
2
.5
2
4
6
4
.5
3
1
5
8
2
.9
0
T
e
g
u
c
ig
a
lp
a
2
6
6
6
.2
5
1
1
1
1
.8
3
7
8
5
.1
8
1
5
7
8
.5
6
7
0
4
.0
3
2
7
8
0
.9
4
1
8
2
.0
4
3
9
2
.9
3
1
1
9
3
.4
7
8
0
8
.6
0
4
0
9
.3
0
1
0
7
6
.2
3
M
e
x
ic
o
C
it
y
1
2
6
6
.2
2
3
1
3
.8
5
1
3
0
0
.8
6
1
9
2
.7
1
6
9
7
.8
3
1
5
6
5
.3
4
1
2
5
8
.7
6
1
1
2
1
.3
4
2
3
5
7
.8
0
1
0
1
3
.9
1
1
7
7
7
.3
4
3
5
5
.8
3
M
o
n
t
e
r
r
e
y
1
0
1
9
.2
1
8
4
2
.7
8
1
4
6
7
.2
4
5
6
6
.3
4
1
2
4
5
.5
2
9
0
3
.8
4
1
7
1
0
.8
2
1
5
1
9
.4
9
2
5
6
9
.0
7
1
2
1
4
.5
3
2
1
9
0
.0
3
1
0
2
5
.2
6
G
u
a
d
a
la
ja
r
a
8
1
1
.5
5
7
6
7
.1
3
1
7
1
4
.8
1
3
0
5
.5
5
1
1
5
7
.8
9
1
2
7
4
.4
8
1
7
1
9
.6
0
1
5
7
8
.1
2
2
7
9
6
.5
8
1
4
2
7
.9
7
2
2
3
7
.9
6
7
9
8
.7
1
T
ij
u
a
n
a
1
1
3
7
.1
9
2
5
4
6
.6
2
3
2
3
2
.3
7
2
1
0
7
.1
7
2
9
6
6
.5
5
9
8
6
.6
1
3
4
7
8
.3
7
3
2
9
4
.9
0
4
3
1
6
.7
4
2
9
9
2
.6
4
3
9
6
8
.9
2
2
6
5
5
.0
0
L
a
H
a
b
a
n
a
2
8
5
0
.9
5
1
5
0
5
.2
3
5
1
8
.3
4
1
8
8
9
.5
8
1
3
4
1
.5
7
2
5
7
1
.2
8
1
0
3
4
.2
5
9
3
7
.4
3
8
0
3
.1
9
7
9
1
.0
1
1
0
1
3
.9
9
1
6
4
8
.7
0
S
t
.
D
o
m
in
g
o
4
2
1
3
.3
0
2
7
6
0
.6
8
1
7
9
1
.7
0
3
1
9
7
.5
2
2
4
6
9
.3
4
3
9
4
4
.6
9
1
9
4
7
.3
7
1
9
9
6
.7
9
7
2
8
.0
1
2
0
7
6
.1
4
1
5
1
9
.7
9
2
8
3
9
.7
1
S
a
n
S
a
lv
a
d
o
r
2
4
9
7
.0
5
9
5
9
.7
1
8
6
3
.2
9
1
4
1
5
.2
5
5
4
0
.2
0
2
6
7
2
.3
6
2
4
1
.6
4
3
9
9
.6
6
1
4
1
0
.7
8
8
0
9
.7
8
6
2
8
.6
4
8
9
1
.2
7
S
a
n
J
u
a
n
4
6
1
1
.5
6
3
1
7
0
.5
8
2
1
9
6
.2
9
3
6
0
4
.7
1
2
8
8
0
.5
2
4
3
1
0
.3
1
2
3
5
5
.5
8
2
4
0
7
.9
8
1
1
3
8
.7
7
2
4
8
2
.0
4
1
9
1
5
.6
6
3
2
5
1
.0
4
T
o
r
r
e
o
n
7
1
5
.9
0
1
0
2
5
.2
2
1
7
5
7
.1
8
6
2
7
.8
7
1
4
4
4
.0
7
7
5
4
.8
7
1
9
4
9
.7
0
1
7
6
9
.0
3
2
8
6
2
.1
0
1
4
9
4
.9
8
2
4
4
4
.7
6
1
1
6
3
.0
9
L
a
P
a
z
,
B
.C
.
1
5
5
8
.1
4
2
4
2
7
.3
1
1
0
8
8
.2
9
1
9
6
3
.4
2
9
2
5
.5
6
2
5
1
7
.2
7
2
3
6
2
.0
9
3
5
3
1
.1
6
2
1
5
0
.9
5
3
0
3
2
.1
9
1
6
1
0
.1
7
V
e
r
a
c
r
u
z
1
5
5
8
.1
4
9
9
8
.7
8
4
7
0
.5
1
4
1
9
.9
4
1
7
4
0
.9
7
9
5
9
.1
7
8
1
1
.0
0
2
0
4
4
.4
3
7
1
5
.0
0
1
4
7
4
.7
6
2
3
6
.3
3
C
a
n
c
u
n
2
4
2
7
.3
1
9
9
8
.7
8
1
4
1
0
.7
0
8
2
7
.4
6
2
2
8
3
.4
7
6
3
4
.9
9
4
7
4
.2
7
1
1
0
5
.8
6
2
8
7
.7
3
8
6
6
.6
0
1
1
3
1
.4
7
Q
u
e
r
e
t
a
r
o
1
0
8
8
.2
9
4
7
0
.5
1
1
4
1
0
.7
0
8
7
7
.4
8
1
3
8
0
.7
5
1
4
2
9
.2
1
1
2
7
9
.4
6
2
4
9
1
.0
9
1
1
2
3
.4
7
1
9
4
5
.2
7
5
4
8
.2
1
T
u
x
t
la
,
G
t
z
.
1
9
6
3
.4
2
4
1
9
.9
4
8
2
7
.4
6
8
7
7
.4
8
2
1
4
8
.8
5
5
6
9
.0
1
4
7
2
.5
5
1
7
4
2
.0
1
6
0
4
.6
6
1
0
8
7
.7
8
3
8
2
.4
2
C
d
.
J
u
a
r
e
z
9
2
5
.5
6
1
7
4
0
.9
7
2
2
8
3
.4
7
1
3
8
0
.7
5
2
1
4
8
.8
5
2
6
0
4
.6
4
2
4
0
6
.9
7
3
3
4
8
.5
9
2
0
6
4
.4
7
3
0
6
6
.9
8
1
9
0
3
.9
3
S
a
n
P
e
d
r
o
S
u
la
2
5
1
7
.2
7
9
5
9
.1
7
6
3
4
.9
9
1
4
2
9
.2
1
5
6
9
.0
1
2
6
0
4
.6
4
2
1
1
.2
6
1
2
2
2
.0
7
6
2
9
.9
0
5
1
9
.3
0
9
5
0
.4
2
B
e
lm
o
p
a
n
2
3
6
2
.0
9
8
1
1
.0
0
4
7
4
.2
7
1
2
7
9
.4
6
4
7
2
.5
5
2
4
0
6
.9
7
2
1
1
.2
6
1
2
6
9
.5
6
4
2
1
.3
4
6
7
6
.2
8
8
4
7
.2
0
K
in
g
s
t
o
n
3
5
3
1
.1
6
2
0
4
4
.4
3
1
1
0
5
.8
6
2
4
9
1
.0
9
1
7
4
2
.0
1
3
3
4
8
.5
9
1
2
2
2
.0
7
1
2
6
9
.5
6
1
3
8
0
.6
0
8
3
1
.2
7
2
1
1
4
.3
1
M
e
r
id
a
2
1
5
0
.9
5
7
1
5
.0
0
2
8
7
.7
3
1
1
2
3
.4
7
6
0
4
.6
6
2
0
6
4
.4
7
6
2
9
.9
0
4
2
1
.3
4
1
3
8
0
.6
0
1
0
1
2
.0
2
8
6
3
.7
0
P
t
o
.
C
a
b
e
z
a
s
3
0
3
2
.1
9
1
4
7
4
.7
6
8
6
6
.6
0
1
9
4
5
.2
7
1
0
8
7
.7
8
3
0
6
6
.9
8
5
1
9
.3
0
6
7
6
.2
8
8
3
1
.2
7
1
0
1
2
.0
2
1
4
6
8
.0
2
O
a
x
a
c
a
1
6
1
0
.1
7
2
3
6
.3
3
1
1
3
1
.4
7
5
4
8
.2
1
3
8
2
.4
2
1
9
0
3
.9
3
9
5
0
.4
2
8
4
7
.2
0
2
1
1
4
.3
1
8
6
3
.7
0
1
4
6
8
.0
2
Bibliography
[1] Gerard Maral and Michel Bousquet. Satellite Communications Systems. John
Wiley & Sons, Ltd, Chichester, UK, 5th edition, dec 2009.
[2] Rodolfo Neri and Salvador Landeros. Comunicaciones por satélite. Editorial
Veracruzana, 1st edition, 2015.
[3] Ahmad Gharanjik. Transmission Optimization for High Throughput Satellite
Systems. PhD thesis, Royal Institue of Technology, 2016.
[4] ITU. Radio Regulations. Technical report, International Telecommunication
Union, ITU, Geneve, Switzerland, 2016.
[5] Darren Chang and Olivier de Weck. Basic Capacity Calculation Methods and
Benchmarking for MF-TDMA and MF-CDMA Communication Satellites. In
21st Int. Commun. Satell. Syst. Conf. Exhib., pages 1–9, Reston, Virigina, apr
2003. American Institute of Aeronautics and Astronautics.
[6] EN ETSI. Digital Video Broadcasting (DVB); Second generation framing
structure, channel coding and modulation systems for Broadcasting, Interactive
Services, News Gathering and other broadband satellite applications;
Part 2: DVB-S2 Extensions (DVB-S2X). Technical report, European
Telecommunications Standards Institute (ETSI), France, 2014.
[7] EN ETSI. Digital Video Broadcasting (DVB); Second Generation DVB
Interactive Satellite System (DVB-RCS2); Part 1: Overview and System
Level specification. Technical report, European Telecommunications Standards
Institute (ETSI), France, 2014.
[8] Barry Evans and Paul Thompson. Key issues and technologies for a Terabit/s
satellite. In 28th AIAA Int. Commun. Satell. Syst. Conf., pages 1–11, Anaheim,
CA, aug 2010. American Institute of Aeronautics and Astronautics.
[9] Sudhakar Rao, Minh Tang, and Chih Chien Hsu. Multiple beam antenna
technology for satellite communications payloads. Appl. Comput. Electromagn.
Soc. J., 21(3):353–363, 2006.
[10] Rajeev Gopal, David Whitefield, and Steve Arnold. Technology readiness of
future generation networks leveraging regenerative satellite mesh architecture -
125
BIBLIOGRAPHY 126
A spaceway perspective. Proc. - IEEE Mil. Commun. Conf. MILCOM, pages
1–7, 2007.
[11] Hughes Inc. Hughes High-Throughput Satellite Constellation.
https://www.hughes.com/technologies/hughes-high-throughput-satellite-constellation,
2019.
[12] Echostar Corporation. Brochure: Connecting The World.
http://www.echostarsatelliteservices.com, 2019.
[13] Hector Fenech, Alessia Tomatis, Sonya Amos, Viphakone Soumpholphakdy,
and Jose Luis Serrano-Merino. Eutelsat HTS Systems. Int. J. Satell. Commun.
Netw., 34(January):1–6, 2011.
[14] Paul Thompson, Barry Evans, Michel Bousquet, and Laurent Castenet.
Concepts and Technologies for a Terabit/s Satellite Supporting future
broadband services via satellite. Third Int. Conf. Adv. Satell. Sp. Commun.
Concepts, SPACOMM, pages 12–19, 2011.
[15] Isabelle Dahman, Philippe Arbogast, Nicolas Jeannin, and Bouchra Benammar.
Rain attenuation prediction model for satellite communications based on the
Météo-France ensemble prediction system PEARP. Nat. Hazards Earth Syst.
Sci., 18(12):3327–3341, 2018.
[16] ITU-R Recomendation S.1557. Operational requirements and characteristics
of fixed-satellite service systems operating in the 50/40 GHz bands for use
in sharing studies between the fixed-satellite service and the fixed service.
Technical report, International Telecommunication Union, Geneve, Switzerland,
2002.
[17] Peter Balling. Analytical high-efficiency spot-beam model for high throughput
satellites. Radioengineering, 21(4):1078–1084, 2012.
[18] Argyrios Kyrgiazos, Barry Evans, Paul Thompson, P. Takis Mathiopoulos,
and Stylianos Papaharalabos. A terabit/second satellite system for European
broadband access: a feasibility study. Int. J. Satell. Commun. Netw.,
32(2):63–92, mar 2014.
[19] Richard Swinford and Bertrand Grau. High Throughput Satellites: Delivering
future capacity needs. Technical report, Arthur D. Little, 2015.
[20] Russell Ju Fu Fang. Spatial diversity satellite communications system with
error control. US Patent 4,099,121, 1978.
[21] Ahmad Gharanjik, Bhavani Shankar Mysore Rama Rao, Pantelis Daniel
Arapoglou, and Björn Ottersten. Gateway switching in Q/V band satellite
feeder links. IEEE Commun. Lett., 17(7):1384–1387, 2013.
BIBLIOGRAPHY 127
[22] Muhammad Muhammad, Giovanni Giambene, and Tomaso De Cola. Channel
prediction and network coding for smart gateway diversity in terabit satellite
networks. 2014 IEEE Glob. Commun. Conf. GLOBECOM 2014, pages
3549–3554, 2014.
[23] Giovanni Giambene, Doanh Kim Luong, Van Anh Le, and Tomaso De Cola.
Gateway handover implications on transport layer performance in terabit
satellite networks. 2014 7th Adv. Satell. Multimed. Syst. Conf. 13th Signal
Process. Sp. Commun. Work. ASMS/SPSC 2014, 2014-Janua:9–16, 2014.
[24] Andra Pastrav, Tudor Palade, and Emanuel Puschita. A proposed mechanism
for traffic management during gateway outage periods in high throughput
satellite systems. 2015 38th Int. Conf. Telecommun. Signal Process. TSP 2015,
pages 200–204, 2015.
[25] Ahmad Gharanjik, Bhavani Shankar, Pantelis-daniel Arapoglou, and Bjorn
Ottersten. Multiple Gateway Transmit Diversity in Q/V Band Feeder Links.
IEEE Trans. Commun., 63(3):916–926, mar 2015.
[26] Zoltán Katona, Federico Clazzer, Kevin Shortt, Simon Watts, Hans Peter
Lexow, and Ratna Winduratna. Performance, cost analysis, and ground
segment design of ultra high throughput multi-spot beam satellite networks
applying different capacity enhancing techniques. Int. J. Satell. Commun.
Netw., 34(4):547–573, jul 2016.
[27] ITU-R Recommendation P.1815-1. Differential rain attenuation. Technical
Report October, International Telecommunication Union, Geneve, Switzerland,
2009.
[28] ITU-R Recommendation P.618-13. Propagation data and prediction
methods required for the design of Earth-space telecommunication systems.
Technical Report December, International Telecommunication Union, Geneve,
Switzerland, 2017.
[29] ITU-R Recomendation P.837-7. Characteristics of precipitation for propagation
modelling. Technical report, International Telecommunication Union, Geneve,
Switzerland, 2017.
[30] ITU-R Recomendation P.838-3. Specific attenuation model for rain. Technical
report, International Telecommunication Union, Geneve, Switzerland, 2005.
[31] ITU-R Recomendation P.839-4. Rain Height Model for Prediction Methods.
Technical report, International Telecommunication Union, Geneve, Switzerland,
2013.
[32] ITU-R Recomendation P.841-5. Conversion of annual statistics to worst-month
statistics. Technical report, International Telecommunication Union, Geneve,
Switzerland, 2016.
BIBLIOGRAPHY 128
[33] ITU-R Recomendation P.1144-9. Guide to the application of the propagation
methods of Radiocommunication Study Group 3. Technical report,
International Telecommunication Union, Geneve, Switzerland, 2017.
[34] ITU-R Recomendation P.1510-1. Mean surface temperature. Technical report,
International Telecommunication Union, Geneve, Switzerland, 2017.
[35] ITU-R Recomendation P.1511-1. Topography for Earth-space propagation
modelling. Technical report, International Telecommunication Union, Geneve,
Switzerland, 2015.
[36] S. H. Lin. Statistical Behavior of Rain Attenuation. Bell Syst. Tech. J.,
52(4):557–581, 1973.
[37] Torleiv Maseng and Petter M. Bakken. A Stochastic Dynamic Model of Rain
Attenuation. IEEE Trans. Commun., 29(5):660–669, 1981.
[38] ITU-R Recommendation P.1853-1. Tropospheric Attenuation Time Series.
Technical Report February, International Telecommunication Union, Geneve,
Switzerland, 2012.
[39] ITU-R Recommendation P.1057-6. Probability distributions relevant to
radiowave propagation modelling. Technical Report August, International
Telecommunication Union, Geneve, Switzerland, 2019.
[40] Edwin Fernando Rojas, Laura Poveda, and Nicolás Grimblatt. Estado de la
Banda Ancha en América Latina y el Caribe. Technical report, United Nations,
Santiago, Chile, 2016.
[41] ITU and UNESCO. The State of Broadband 2016: Broadband catalyzing
sustainable development. Technical Report September, International
Telecommunication Union, ITU, Geneve, Switzerland, 2016.
[42] SATNEWS. SatMagazine: Special SmallSat Symposium. SATNEWS,
17(February):39–44, feb 2019.
[43] CISCO. Cisco Visual Networking Index: Forecast Methodology, 2016–2021.
White Pap., .(June):1–17, 2017.
[44] CISCO. The Zettabyte Era: Trends and Analysis. White Pap., .(May):1–29,
2015.
[45] 5G AMERICAS. Spectrum Recommendations for the U.S. Technical report,
5G Americas, 2018.
[46] DTVE. Industry Survey 2018. Technical Report February, Digital TV Europe,
London, 2018.
[47] BMI Research. Latin America Investment Opportunities In Telecoms : Risk /
Reward Analysis. Technical report, BMI Research, London, 2018.
BIBLIOGRAPHY 129
[48] VIA SATELLITE. Space 2.0: Taking AI Far Out. Technical Report December,
Via Satellite, Rockville, MD - US, 2019.
[49] Andres Cornejo and Salvador Landeros. Simulation Model and Comparison for
Satellite Links in Ku and Ka Bands for Standards-based on DVB-S2 and NS3.
IEEE Lat. Am. Trans., 14(4):1689–1697, apr 2016.
[50] Louis Ippolito. Satellite Communications Systems Engineering. John Wiley &
Sons, Ltd, Chichester, UK, sep 2008.
[51] ITU-R Recommendation P.676-12. Attenuation by atmospheric gases and
related effects. Technical report, International Telecommunication Union,
Geneve, Switzerland, 2019.
[52] ITU-R Recommendation P.834-9. Effects of tropospheric refraction on
radiowave propagation. Technical report, International Telecommunication
Union, Geneve, Switzerland, 2017.
[53] ITU-R Recommendation P.840-8. Attenuation due to clouds and fog. Technical
report, International Telecommunication Union, Geneve, Switzerland, 2019.
[54] R. Gallager. Low-density parity-check codes. IEEE Trans. Inf. Theory,
8(1):21–28, jan 1962.
[55] Sarah J Johnson. Introducing Low Density Parity Check ( LDPC ) Codes.
Technical Report October, The University of Newcastle, Australia, 2015.
[56] EN ETSI. Digital Video Broadcasting (DVB); modulation systems for
Broadcasting , other broadband satellite applications (DVB-S2). Technical
report, European Telecommunications Standards Institute (ETSI), France,
2009.
[57] Claude Berrou and Alain Glavieux. Near optimum error correcting coding and
decoding: Turbo-codes. IEEE Trans. Commun., 44(9):1261–1271, 1996.
[58] Dirk P. Kroese, Tim Brereton, Thomas Taimre, and Zdravko I. Botev. Why the
Monte Carlo method is so important today. Wiley Interdiscip. Rev. Comput.
Stat., 6(6):386–392, 2014.
[59] S.K. Rao. Design and analysis of multiple-beam reflector antennas. IEEE
Antennas Propag. Mag., 41(4):53–59, 1999.
[60] W. Stutzman and M. Terada. Design of offset-parabolic-reflector antennas for
low cross-pol and low sidelobes. IEEE Antennas Propag. Mag., 35(6):46–49,
dec 1993.
[61] Thomas A. Milligan. Modern Antenna Design. John Wiley & Sons, Inc.,
Hoboken, NJ, USA, jun 2005.
BIBLIOGRAPHY 130
[62] Dimitri Serrano-Velarde, Emmanuel Lance, Hector Fenech, and Georges E.
Rodriguez-Guisantes. Novel dimensioning method for high-throughput
satellites: Forward link. IEEE Trans. Aerosp. Electron. Syst., 50(3):2146–2163,
2014.
[63] Warren L. Stutzman and Gary A. Thiele. Antenna Theory and Design. John
Wiley & Sons., Blacksburg, Virginia, US, 3rd edition, 1998.
[64] Erich Lutz, Markus Werner, and Axel Jahn. Satellite Systems for Personal and
Broadband Communications. Springer Berlin Heidelberg, Berlin, Heidelberg,
2000.
[65] D. Tipper. Fundamentals of Cellular Networks. University of Pittsburgh,
available at: http://www.pitt.edu/˜dtipper/2720/2720_Slides4.pdf, 2018.
[66] ITU-R Recommendation S.1328-4. Satellite system characteristics to be
considered in frequency sharing analyses within the fixed-satellite service
ANNEX 1. Technical report, International Telecommunication Union, Geneve,
Switzerland, 2002.
[67] TICRA. GRASP: Analysis and Design of Reflector Antennas. Trial Version.
Available at: http://www.ticra.com/software/grasp. Copenhagen, Denmark,
2020.
[68] Pantelis-Daniel Arapoglou, Konstantinos P. Liolis, and Athanasios D.
Panagopoulos. Railway satellite channel at Ku band and above: Composite
dynamic modeling for the design of fade mitigation techniques. Int. J. Satell.
Commun. Netw., 30(1):1–17, jan 2012.
[69] Georgios A. Karagiannis, Athanasios D. Panagopoulos, and John D.
Kanellopoulos. Multidimensional rain attenuation stochastic dynamic
modeling: Application to earth-space diversity systems. IEEE Trans. Antennas
Propag., 60(11):5400–5411, 2012.
[70] O. Koudelka, M. Schmidt, and J. Ebert. Design of a 40/50 GHz satellite ground
station for fade mitigation experiments. Acta Astronaut., 86:68–76, 2013.
[71] Tommaso Rossi, Mauro De Sanctis, and Fabio Maggio. Evaluation of outage
probability for satellite systems exploiting smart gateway configurations. IEEE
Commun. Lett., 21(7):1541–1544, 2017.
[72] Tomaso De Cola and Maurizio Mongelli. Adaptive Time Window Linear
Regression for Outage Prediction in Q/V Band Satellite Systems. IEEE Wirel.
Commun. Lett., 7(5):808–811, 2018.
[73] Nicolas Jeannin, Laurent Castanet, Isabelle Dahman, Vivien Pourret, and
Béatrice Pouponneau. Smart gateways switching control algorithms based on
tropospheric propagation forecasts. Int. J. Satell. Commun. Netw., 37(1):43–55,
2019.
BIBLIOGRAPHY 131
[74] D.G. Sweeney and C.W. Bostian. Implementing adaptive power control as a
30/20-GHz fade countermeasure. IEEE Trans. Antennas Propag., 47(1):40–46,
1999.
[75] K. Kastamonitis, B. Gremont, and M. Filip. Short-term prediction of rain
attenuation based on fade slope. Electron. Lett., 39(8):687, 2003.
[76] Y. Vasavada, Y. Liu, J. Sanchez, P. Kapoor, Y. Antia, J. Rozmaryn, and
J. Border. An Uplink Power Control driven compensation of wideband,
semi-static, rain or interference induced link impairments affecting satellite
feeder links operating at Ka band and above. In MILCOM 2016 - 2016 IEEE
Mil. Commun. Conf., pages 373–378. IEEE, nov 2016.
[77] Sepp Hochreiter and Jürgen Schmidhuber. Long Short-Term Memory. Neural
Comput., 9(8):1735–1780, nov 1997.
[78] Steven Elsworth and Stefan Güttel. Time Series Forecasting Using LSTM
Networks: A Symbolic Approach. ArXiv, 1(2003.05672):1–12, 2020.
[79] Alex Sherstinsky. Fundamentals of Recurrent Neural Network (RNN) and
Long Short-Term Memory (LSTM) network. Phys. D Nonlinear Phenom.,
404(March):1–43, 2020.
[80] The Mathworks. Introducing Machine Learning. Technical Report January
2016, The MathWorks, Inc., Natick, Massachusetts, USA, 2016.
[81] Keras. Dense layer. https://keras.io/api/layers/core_layers/dense/, 2020.
[82] Martin T. Hagan, Howard B. Demuth, Mark H. Beale, and Orlando De Jesus.
Neural Network Design. Oklahoma State University, Stillwater, OK, United
States, Stillwater, OK, 2nd editio edition, sep 2014.
[83] Simon Haykin. Neural Networks and Learning Machines. McMaster University,
Hamilton, Ontario, Canada, 3rd edition, 2009.
[84] Kevin P Murphy. Machine Learning: A Probabilistic Perspective. The MIT
Press, Cambridge, Massachusetts, 1st edition, 2012.
[85] Keras. Layer activation functions. https://keras.io/api/layers/activations/,
2020.
[86] Günter Klambauer, Thomas Unterthiner, Andreas Mayr, and Sepp Hochreiter.
Self-normalizing neural networks. Adv. Neural Inf. Process. Syst.,
2017-Decem(12):972–981, 2017.
[87] Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic
Optimization. In 3rd Int. Conf. Learn. Represent., pages 1–15, San Diego,
CA, USA, dec 2015.
BIBLIOGRAPHY 132
[88] Gábor Petneházi. Recurrent Neural Networks for Time Series Forecasting.
Handb. Nat. Comput., 1-4:461–477, dec 2018.
[89] Rob J. Hyndman and Anne B. Koehler. Another look at measures of forecast
accuracy. Int. J. Forecast., 22(4):679–688, oct 2006.
[90] John O. Rawlings, Sastry G. Pantula, and David A. Dickey. Applied Regression
Analysis. Springer Texts in Statistics. Springer-Verlag, New York, may 1998.
[91] M. C. Peel, B. L. Finlayson, and T. A. McMahon. Updated world map of the
Köppen-Geiger climate classification. Hydrol. Earth Syst. Sci., 11(5):1633–1644,
2007.
[92] Roger G. Barry. Atmosphere, Weather and Climate. Routledge, London, 1st
editio edition, oct 2009.
[93] S. Poornima and M. Pushpalatha. Prediction of rainfall using intensified LSTM
based recurrent Neural Network with Weighted Linear Units. Atmosphere
(Basel)., 10(11):1–18, 2019.
[94] Dhaval P. Patel, Mitul M. Patel, and Devendra R. Patel. Implementation of
ARIMA model to predict Rain Attenuation for Ku-band 12 Ghz Frequency.
IOSR J. Electron. Commun. Eng., 9(1):83–87, 2014.
[95] Yan Jun Puah, Yuk Feng Huang, Kuan Chin Chua, and Teang Shui Lee. River
catchment rainfall series analysis using additive holt–Winters method. J. Earth
Syst. Sci., 125(2):269–283, 2016.
[96] Ricardo Aguasca-Colomo, Dagoberto Castellanos-Nieves, and Máximo Méndez.
Comparative analysis of rainfall prediction models using machine learning in
islands with complex orography: Tenerife Island. Appl. Sci., 9(22):1–17, 2019.
[97] Riccardo De Gaudenzi, Emiliano Re, and Piero Angeletti. Smart Gateways
Concepts for High-Capacity Multi-beam Networks. In 30th AIAA Int. Commun.
Satell. Syst. Conf., pages 1–18, Reston, Virigina, sep 2012. American Institute
of Aeronautics and Astronautics.
[98] Hong Chuan Yang and Mohamed Slim Alouini. Markov chains and
performance comparison of switched diversity systems. IEEE Trans. Commun.,
52(7):1113–1125, 2004.
[99] Jiahao Yang, Wenkai Zhang, Mingchuan Yang, and Yanyong Su. A
New Gateway Switching Strategy in Q/V Band High Throughput Satellite
Communication Systems Feeder Links. In Weixiao Meng, editor, 10th EAI Int.
Conf. WiSATS 2019, Part II, pages 506–514, Harbin, China, 2019. Springer.
[100] Zhuoming Li, Huiyun Xia, Yu Zhang, Junqing Qi, Shaohua Wu, and Shushi
Gu. Multi-objective network optimization combining topology and routing
algorithms in multi-layered satellite networks. Sci. China Inf. Sci., 61(8):1–3,
2018.
BIBLIOGRAPHY 133
[101] Roberto Nebuloni, Carlo Riva, Lorenzo Luini, Tommaso Rossi, Mauro De
Sanctis, Marina Ruggieri, Giuseppe Codispoti, and Giorgia Parca. Optimization
of Q/V-band Smart Gateway Switching in the framework of Q/V-Lift Project.
IEEE Aerosp. Conf. Proc., 2019-March:1–8, 2019.
[102] Rohani Hoda and Roosta Azad Kamali. Calculating Total System Availability.
Technical report, KLM-Air France, 2014.
[103] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez, and Jose M
Matias. Analysis to Quantify and Optimize Spot Beams for a High Throughput
Satellite in Ka and Q/V Bands. IEEE Lat. Am. Trans., 17(02):219–227, feb
2019.
[104] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martínez Rodríguez-Osorio,
and Jose M Matias. A method for designing an offset-parabolic-reflector antenna
for a ultra-high throughput satellite. In Rev. Tecnol. e Inf., volume 16, pages
84–87, Lima, Peru, 2018. Universidad Inca Garcilaso de la Vega.
[105] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez-Rodriguez, and
Jose M Matias. Interference Evaluations in Frequency Reuse by Using Offset-
Parabolic-Reflector Antennas for a UHTS System. Int. J. Eng. Sci., 7(8):34–45,
2018.
[106] Andres Cornejo, Salvador Landeros-Ayala, Ramon Martinez Rodriguez-Osorio,
and Jose Maria Matias. Optimization of the Ground Segment for an UHTS
System through Neural Networks. In XXXIII Simp. Nac. la Unión Científica
Int. Radio (URSI 2018), pages 1–5, Granada, Spain, 2018.
[107] Andres Cornejo, Salvador Landeros-Ayala, Jose M. Matias, and Ramon
Martinez. Applying Learning Methods to Optimize the Ground Segment for
HTS Systems. 2020 IEEE 11th Lat. Am. Symp. Circuits Syst. LASCAS 2020,
pages 1–4, 2020.
[108] Andres Cornejo, Flor Ortiz-Gomez, Jose M. Matias, Ramon Martinez, and
Miguel Salas-Natera. Quantum Cryptography for Satellite Communication
Systems. In Congr. Nac. Act. Espac. 2020-Agencia Espac. Mex., Mexico City,
2020.
STATEMENT
The presented information in this thesis was obtained from several sources that
were considered trustworthy and punctually consigned to the references. The
nature of used information is strictly for academic research and disclosure, non-
profit, or another characteristic. It has also been done the best effort to
acknowledge data properly, opinions, and presented content in this work, so any
mistakes or omissions can be totally involuntary.
Mexico City, November 2020
IVAN ANDRES CORNEJO GAIBOR