Addis Ababa University

Addis Ababa Institute of Technology

School of Electrical and Computer Engineering

Spectrum Occupancy Prediction Using Deep Learning

Algorithms

By: Addisu Melkie Tafere

A Thesis Submitted to the Department of Electrical and Computer Engineering

for the Partial Fulfillment of the Degree of Master of Science in Computer
Engineering

Addis Ababa, Ethiopia

July 2024
 Addis Ababa University
Addis Ababa Institute of Technology

School of Electrical and Computer Engineering

Spectrum Occupancy Prediction Using Deep Learning

Algorithms

By: Addisu Melkie Tafere

A Thesis Submitted to the Department of Electrical and Computer Engineering

for the Partial Fulfillment of the Degree of Master of Science in Computer
Engineering

Advisor: Dr. Eng. Getachew Alemu

Addis Ababa, Ethiopia

July 2024
 Board of Examiners

Addis Ababa University

Addis Ababa Institute of Technology
School of Electrical and Computer Engineering

By: Addisu Melkie Tafere

This is to certify that the thesis prepared by Addisu Melkie, titled Spectrum Occupancy
Prediction Using Deep Learning Algorithms and Submitted in partial fulfillment of the
requirements for the Degree of Master of Science in Computer Engineering compiles with the
regulations of the University and meets the accepted standards to originality and quality.

Approved by the Board of Examining Committee:

Name Signature

Dean, School of Electrical and

Computer Engineering: Dr. Bisrat Derebssa ___________

Advisor: Dr.Eng. Getachew Alemu ____________

Internal Examiner: _____________________ _____________

External Examiner: _____________________ ____________

Addis Ababa, Ethiopia

July 2024

i|Page
 Declaration

I hereby declare that this MSc thesis is my original work and has not been presented for a degree
in any other university, and all sources of material used for this thesis have been duly
acknowledged.

Name: Addisu Melkie

Signature:_________________
Date:_________________

This MSc thesis has been submitted for examination with my approval as a university
advisor.
Dr.-Eng. Getachew Alemu
Signature:_________________
Date:_________________

ii | P a g e
 DEDICATION

I would like to dedicate this thesis to my beloved family, especially to Yegna Enat and Yegna
Abat, for their unreserved support and encouragement throughout my life.

iii | P a g e
 Acknowledgments
First, I would like to thank Almighty God and his mother, Saint Virgin Mary with the herald angel
San Gabriel, for helping me reach this milestone after so many ups and downs.

Very special gratitude goes out to all down to my advisor Dr.-Eng Getachew Alemu for the
encouragement and support by giving me good comments and suggestions. I will not forget to
remember his encouragement and their timely support and guidance until the completion of the
study.

I would also like to thank Dr. Bisrat Derebssa, Dr. Fistum Assaminew, and their colleagues for
their valuable support and all the difficulties they solved.

I am also thanking Dr. Ephrem T Bekel, Dr. Beneyam B Haile, and all Zeregaw Internet and Cloud
Services Provider staff for their help in the data collection by devoting their time from helping in
the equipment configuration and the material support.

I am also so grateful to thank my friends who were with me during the regional data collection
time, especially Aron Woledu, Seifu Girma, Rediet Million, and Samson Takle for their great
collaborations and encouragement.

I would like to express my heartfelt gratitude to my friends who were with me throughout my
study. In particular, I want to extend a special thanks to Ashenafi Workie, who stood by me during
many challenging times.

Finally, I would like to thank the Ethiopia Communication Authority and Ethio Telecom for their
valuable support during the data collection time.

Thank you for all your encouragement, God will repay you for all the good things you have done
for me.

iv | P a g e
 Table of Contents

Board of Examiners ......................................................................................................................... i

Declaration ...................................................................................................................................... ii

Acknowledgments ......................................................................................................................... iv

List of Figures ................................................................................................................................ ix

List of Tables ................................................................................................................................. xi

List of Abbreviations .................................................................................................................... xii

List of Acronyms ......................................................................................................................... xiv

Abstract ......................................................................................................................................... xv

Chapter One .................................................................................................................................... 1

1. Introduction .......................................................................................................................... 1

1.1. Background of the Study ...................................................................................................... 1

1.2. Motivation of the Study........................................................................................................ 3

1.3. Statement of the Problem ..................................................................................................... 4

1.4. Objective of the Study .......................................................................................................... 5

1.4.1. General Objective ........................................................................................................ 5

1.4.2. Specific Objectives ...................................................................................................... 6

1.5. Scope and Limitations of the Study ..................................................................................... 6

1.5.1. Scope of the Study ....................................................................................................... 6

1.5.2. Limitations of the Study .............................................................................................. 6

1.6. Significance of the Study ..................................................................................................... 6

1.7. Contribution of the Study ..................................................................................................... 7

1.8. Organization of the Study .................................................................................................... 7

Chapter Two ................................................................................................................................... 9

2. Theoretical Background and Related Works ....................................................................... 9

v|Page
 2.1. Cognitive Radio.................................................................................................................... 9

2.1.1. Spectrum Sensing Techniques ................................................................................... 10

2.1.2. Spectrum Access Techniques .................................................................................... 12

2.1.3. Applications of Cognitive Radio ............................................................................... 13

2.1.4. Challenges in Spectrum Sensing ............................................................................... 13

2.2. Related Works .................................................................................................................... 14

2.2.1. Spectrum Occupancy Prediction Using ANN ........................................................... 14

2.2.2. Spectrum Occupancy Prediction Using RNN ........................................................... 16

2.2.3. Spectrum Occupancy Prediction Using GRU ........................................................... 17

2.2.4. Spectrum Occupancy Prediction Using LSTM ........................................................ 18

2.2.5. Spectrum Occupancy Prediction Using HNN ........................................................... 21

2.2.6. Summary of Related Works ...................................................................................... 22

Chapter Three ............................................................................................................................... 23

3. Research Methodology ...................................................................................................... 23

3.1. Overview ............................................................................................................................ 23

3.2. Data Collection ................................................................................................................... 23

3.2.1. Description of Data Features ..................................................................................... 23

3.3. Data Pre-processing ...................................................................................................... 24

3.3.1. Feature Selection ....................................................................................................... 24

3.3.2. Data Cleaning ............................................................................................................ 25

3.3.3. Feature Engineering................................................................................................... 25

3.4. Setup the Data .................................................................................................................... 25

3.5. Design and Implementation Tools ..................................................................................... 26

3.5.1. Hardware Tools ......................................................................................................... 26

3.5.2. Software Tools........................................................................................................... 26

vi | P a g e
 3.6. Model Performance Evaluation Metrics ........................................................................... 27

3.6.1. Mean Squared Error .................................................................................................. 27

3.6.2. Root Mean Square Error ............................................................................................ 28

3.6.3. Mean Absolute Error ................................................................................................. 28

3.6.4. Mean Absolute Percentage Error............................................................................... 28

3.7. Confusion Matrix ............................................................................................................... 29

Chapter Four ................................................................................................................................. 31

4. Proposed Approach ............................................................................................................ 31

4.1. Overview ............................................................................................................................ 31

4.2. System Model..................................................................................................................... 32

4.2.1. Time-Series Model Analysis ..................................................................................... 33

4.3. Proposed Deep LSTM Architecture ................................................................................... 33

4.3.1. Proposed Deep Learning LSTM Basic Architecture ................................................. 35

4.3.2. Proposed Deep LSTM Architecture Components ..................................................... 36

Chapter Five.................................................................................................................................. 39

5. Implementation Details, Result and Discussion ................................................................ 39

5.1. Introduction ........................................................................................................................ 39

5.2. Data Set Description........................................................................................................... 39

5.2.1. Feature Description and Selection ............................................................................. 40

5.2.2. Data Analysis and Feature Engineering .................................................................... 41

5.2.3. Preparing The Dataset ............................................................................................... 45

5.2.4. Creating the Model .................................................................................................... 46

5.3. Implementation Details ...................................................................................................... 47

5.3.1. Working Environment and Tools .............................................................................. 47

5.3.2. Dataset ....................................................................................................................... 47

vii | P a g e
 5.3.3. Experiment Setup with .................................................................................................... 47

5.4. Results of the Study............................................................................................................ 49

5.4.1. Spectrum Occupancy Prediction Using the LSTM Model ........................................ 49

5.4.2. Spectrum Occupancy Prediction Using the Bi-LSTM Model................................... 53

5.4.3. Spectrum Occupancy Prediction Using the ConvLSTM Model ............................... 56

5.5. Discussion .......................................................................................................................... 59

5.5.1. Quantifying Improvements ........................................................................................ 59

5.5.2. Comparison of Results .............................................................................................. 61

Chapter Six ................................................................................................................................... 63

6. Conclusion and Future Work ............................................................................................. 63

6.1. Conclusion.......................................................................................................................... 63

6.2. Future Work ....................................................................................................................... 64

Reference ...................................................................................................................................... 65

Appendixes ................................................................................................................................... 70

Appendix A: Sample Code for Building, Creating, and Compiling the LSTM Model ............ 70

Appendix B: Sample Code for Building, Creating, and Compiling the Bi-LSTM Model ....... 73

Appendix C: Sample Code for Building, Creating, and Compiling the ConvLSTM Model .... 77

viii | P a g e
 List of Figures

Figure 2.1 Contextualizing the characterization of spectral decision stage in CRN ...................................................... 9

Figure 4.1 The Proposed Deep Learning Approach .................................................................................................... 31

Figure 4.2 Spectrum channels occupancy state modeling ........................................................................................... 32

Figure 4.3 Basic Architecture of the LSTM Model ..................................................................................................... 34

Figure 4.4 Learning and predicting long-term spectrum occupancy in LSTM ............................................................ 36

Figure 4.5 LSTM architecture with input Xt in time slot t ........................................................................................... 38

Figure 5.1 GSM900 MHz uplink band average spectrum utilization in four regional cities ....................................... 41

Figure 5.2 Addis Abeba City average spectrum utilization in four spectrum uplink bands ....................................... 42

Figure 5.3 GSM900 spectrum utilization percentage in five different days ................................................................ 43

Figure 5.4 GSM900 spectrum utilization percentage in four different seasons of a day ............................................. 44

Figure 5.5 Spectrogram of the 902.5-915 MHz for the CFD based spectrum sensing ................................................ 45

Figure 5.6 The dataset used for the spectrum occupancy state prediction model ........................................................ 46

Figure 5.7 Learning and predicting the long-term spectrum occupancy in the proposed LSTM network with
sequence-to-sequence architecture ..................................................................................................................... 46

Figure 5.8 Training and validation, loss and accuracy of the LSTM model short-term prediction ............................. 49

Figure 5.9 Training and validation, loss and accuracy of the LSTM model for 3-hours ahead prediction .................. 50

Figure 5.10 Training and validation, loss and accuracy of the LSTM model for 5-hours ahead prediction ................ 51

Figure 5.11 Spectrum measurement data for the duration of 5 hours .......................................................................... 52

Figure 5.12 Training and validation, loss and accuracy of the Bi-LSTM model short-term prediction ...................... 53

Figure 5.13 Training and validation, loss and accuracy of the Bi-LSTM model for 3-hours ahead prediction ...... 54

ix | P a g e
Figure 5.14 Training and validation, loss and accuracy of the Bi-LSTM model for 5-hours ahead prediction ........... 55

Figure 5.15 Training and validation, loss and accuracy of the ConvLSTM model for short-term prediction ............. 56

Figure 5.16 Training and validation, loss and accuracy of the ConvLSTM model 3-hours ahead prediction ............. 57

Figure 5.17 Training and validation, loss and accuracy of the ConvLSTM model for 5-hours ahead prediction ....... 58

Figure 5.18 Comparison of models in Accuracy, Precision, and F1-Score ................................................................. 62

Figure 5.19 Comparison of models in MSE, RMSE, MAE, and MAPE ..................................................................... 62

x|Page
 List of Tables

Table 1.1 RF Spectrum bands and their applications.....................................................................................................1

Table 2.1 Summary of Related Works......................................................................................................................... 22

Table 3.1 Description of spectrum measurement data attributes ................................................................................. 24

Table 3.2 Hardware tools used for implementation of the study ................................................................................. 26

Table 5.1 A sample data obtained from the spectrum measurement campaign ........................................................... 39

Table 5.2 The final features selected used to model the spectrum occupancy state prediction ................................... 40

Table 5.3 The hardware and software specifications used to conduct the experiments ............................................... 47

Table 5.4 Performance results of the LSTM model for short-term prediction ............................................................ 49

Table 5.5 Performance results of the LSTM model for 3-hours ahead prediction....................................................... 50

Table 5.6 Performance results of the LSTM model for 5-hours ahead prediction....................................................... 51

Table 5.7 Performance results of the Bi-LSTM model for short-term prediction ....................................................... 53

Table 5.8 Performance results of the Bi-LSTM model for 3-hours ahead prediction.................................................. 54

Table 5.9 Performance results of the Bi-LSTM model for 5-hours ahead prediction.................................................. 55

Table 5.10 Performance results of the ConvLSTM model for short-term prediction .................................................. 56

Table 5.11 Performance results of the ConvLSTM model for 3-hours ahead prediction ............................................ 57

Table 5.12 Performance results of the ConvLSTM model for 5-hours ahead prediction ............................................ 58

Table 5.13 Quantified improvements achieved in the spectrum occupancy state prediction ...................................... 61

xi | P a g e
 List of Abbreviations
ANN Artificial Neural Network
BP Backpropagation
Bi-LSTM Bidirectional Long Short-Term Memory
CDMA Code Division Multiple Access
CFD Cyclostationary Feature Detection
Conv-LSTM Convolutional Long Short-Term Memory
COR Channel Occupancy Rate
CR Cognitive Radio
CRN Cognitive Radio Network
CRIoT Cognitive Radio Internet of Things
CRU Cognitive Radio User
CSS Cooperative Spectrum Sensing
DC Duty Cycle
DL Deep Learning
DNN Deep Neural Network
ECA Ethiopian Communication Authority
ED Energy Detection
FCC Federal Communication Commission
FSA Fixed Spectrum Allocation
GA Genetic Algorithm
GPU General-Purpose Graphical Processing Units
GRU Gated Recurrent Unit
GSM Global System for Mobile Communication
HMM Hidden Markov Model
HNN Hybrid Neural Network
IMT International Mobile Technology
IOT Internet of Things
ISM Industrial Scientific and Medical
KM Killo Meter

xii | P a g e
LTE Long Term Evolution
LMR Land Mobile Radio
LSTM Long Short-Term Memory
MAC Media Access Control
MAE Mean Absolute Error
MAPE Mean Absolute Percentage Error
ML Machine Learning
MLP Multilayer Perceptron
MIMO Multiple Input Multiple Output
MAP Mean Square Error
PSD Power Spectral Density
PU Primary User
RBF Radial Basis Function
RF Radio Frequency
RMSE Root Mean Square Error
RNN Recurrent Neural Network
QoS Quality of Service
SDR Software Defined Radio
SMS Short Message Service
SNR Signal-to-Noise Ratio
SS Spectrum Sensors
SU Spectrum Utilization
SUs Secondary Users
TCI Telecommunication Intelligence
TV Television
UMTS Universal Mobile Telecommunications Service
USA United States of America
WLAN Wireless Local Area Network

xiii | P a g e
 List of Acronyms
2G Second Generation
3G Third Generation
4G Fourth Generation
5G Fifth Generation
ADAM Adaptive Moment Estimation
dB Decibel
ReLU Rectified Linear Unit
mm Millimeter
Wi-Fi Wireless Fidelity

xiv | P a g e
 Abstract

The fixed spectrum allocation (FSA) policy causes a waste of valuable and limited natural
resources because a significant portion of the spectrum allocated to users is unused. With the
exponential growth of wireless devices and the continuous development of new technologies
demanding more bandwidth, there is a significant spectrum shortage under current policies.
Dynamic spectrum access (DSA) implemented in a cognitive radio network (CRN) is an emerging
solution to meet the growing demand for spectrum that promises to improve spectrum utilization
that enables secondary users (SUs) to utilize unused spectrum allocated to primary users (PUs).
CRNs have capabilities for empowerment to spectrum sensing, decision-making, sharing, and
mobility. Spectrum sharing gets spectrum usage patterns from spectrum occupancy prediction to
determine the channel states as “idle” or “busy”. This study has addressed all the limitations of
the previous studies by implementing a comprehensive approach that encompasses reliable
spectrum sensing, potential candidate spectrum band identification, long-term adaptive prediction
modeling, and quantification of improvements achieved in the prediction model. The Long-Short
Term Memory (LSTM) Deep Learning (DL) model was proposed as a solution for this study to
address the challenge of capturing temporal dynamics in sequential inputs. The LSTM model
leverages a gating mechanism to regulate information flow within the network, allowing it to learn
and model long-term temporal dependencies effectively. The dataset used for this study was
obtained from a real-world spectrum measurement by employing the Cyclostationary Feature
Detection (CFD) approaches in the GSM900 mobile network uplink band, spanning a frequency
range of 902.5 to 915 MHz over five consecutive days. The dataset comprises a total of 225,000
data points. The five-day spectrum measurement data analysis yields an average spectrum
utilization of 20.47%. The proposed model has predicted the spectrum occupancy state for 5 hours
ahead in the future with an accuracy of 99.45% improved the spectrum utilization from 20.47% to
98.28% and reduced the sensing energy to 29.39% compared to real-time sensing.

Keywords: Spectrum, Spectrum Occupancy, Dynamic Spectrum Access, Deep Learning, and
Cognitive Radio

xv | P a g e
 Chapter One

1. Introduction

1.1. Background of the Study

The Radio Frequency (RF) spectrum stands as a limited and valuable natural resource used for
various wireless communication systems, encompassing voice radio, digital terrestrial television
(DTT), mobile telephony, and mobile broadband (MBB), all of which use the spectrum to facilitate
the transmission and reception of data[1]. The RF spectrum spans a wide range of electromagnetic
waves that demonstrate a direct relationship with wavelength, each spectrum band has distinct
advantages and disadvantages. Lower frequencies, for instance, can propagate over longer
distances and exhibit superior penetration through building walls. This characteristic makes them
well-suited for applications such as broadcasting in expansive geographic areas. On the other hand,
higher frequencies offer advantages in microelectronic devices like cell phones. Their shorter
wavelengths enable the use of proportionally smaller antennas, allowing these devices to transmit
larger volumes of data. Table 1.1 provides an overview of the diverse bands within the RF
spectrum and their corresponding applications [2] .

Table 1.1 RF Spectrum bands and their applications

Band Frequency Wavelength Propagation Application

VLF-Very Low Frequency 3-30 kHz 100-10 km Ground Long-range radio navigation

LF-Low Frequency 30-300 kHz 10-1 km Ground Radio beacons and navigational
locators
MLF-Medium Frequency 300-3000 kHz 1000-100 m Sky AM radio

HF-High Frequency 3-30 MHz 100-10 m Sky Citizens band (CB), ship/aircraft

VHF-Very High Frequency 30-300 MHz 10-1 m Sky and VHF TV, FM radio
line-of-sight
UHF-Ultra High Frequency 300-3000 100-10 cm Sky and UHF TV, cellular phones, paging,
MHz line-of-sight satellite
SHF-Super High Frequency 3-30 GHz 10-1 cm Line-of-sight Satellite
EHF-Extremely High Frequency 30-300 GHz 10-1 cm Line-of-sight Radar, satellite

1|Page
As wireless technologies continue to advance, the effective management and allocation of the RF
spectrum remain essential for sustaining the growth and reliability of these systems. The spectrum
used for wireless communication systems is a limited resource that cannot be simultaneously
utilized for different services. This is because the simultaneous use of the same spectrum channel
by different services can lead to interference. Therefore, careful spectrum channel management is
crucial to prevent interference. In recent years, the demand for spectrum has increased dramatically
due to the exponential growth of wireless devices and the continuous development of new
bandwidth-hungry technologies. This has resulted in a spectrum scarcity, further increasing its
commercial value. Spectrum stakeholders must, therefore, develop adaptable strategies to manage
and use the spectrum efficiently, meeting both current and future spectrum standards [1][2][3].
The current fixed spectrum allocation (FSA) policy is not addressing the growing demand for
spectrum due to its rigid command-and-control approach, which assigns channels to a single user
regardless of actual usage. This inefficient allocation can lead to the wastage of spectrum resources
and a decrease in the quality of service. However, studies have shown that significant portions of
the spectrum assigned to licensed users are unused, indicating the need for a more dynamic and
responsive spectrum allocation policy [4] [5] [6].

Different communication technologies have used various techniques to overcome spectrum

utilization challenges. Such as higher frequencies using millimeter waves(mmWave) and massive
multiple input multiple outputs (MIMO) in fifth-generation (5G), network densification, and
MIMO in fourth-generation (4G). However, all of these techniques have their drawbacks. For
example, mmWave and densification can only cover short ranges, and MIMO-based techniques
require more complex and power-hungry transmitters [7]. To meet the growing demand for the
spectrum and to change the conventional course of spectrum allocation, researchers are exploring
new solutions. Dynamic spectrum access (DSA) implemented in cognitive radio networks (CRNs)
has now emerged as a solution to improve spectrum utilization and reduce spectrum waste by
allowing secondary users (SUs) to share unused spectrum with primary users (PUs)[8] . Because
the CRNs are comprised of two types of users i.e., PUs and SUs, where PUs have a higher priority
than SUs in accessing the channels. The SUs logically divides the channels allocated to the PUs
into slots. Within each slot, the SUs has to sense the PU channel slot and accordingly access the
slot when idle. The idle slots are called spectrum holes or white spaces[4].

2|Page
Spectrum sharing requires knowledge of spectrum usage patterns, which can be obtained through
spectrum sensing. However, real-time spectrum sensing is considered unreliable due to its high
energy and time consumption. Spectrum occupancy state prediction, which infers the future states
of the spectrum channel, proactively forecasts these states and estimates the effective bandwidth
in the next slot. This allows SUs to adjust their data rates in advance for improved spectrum
sensing. Consequently, SUs can conserve energy and time by avoiding the busy portions of the
spectrum and focusing on idle portions during sensing [4][8]. This demonstrates that spectrum
occupancy state prediction is a key enabler for shared spectrum access in the DSA model, as it
allows SUs to identify and access idle spectrum channels before they become busy [8][9][10]. SUs
in CRNs search for idle spectrum channels to use temporarily. They are equipped with the
cognitive ability to effectively implement the CR system, which performs the following cycle of
functions: Sensing: to observe and sample spectral channels, Decision: to allocate suitable spectral
holes, Sharing: to contend access with other secondary users, and Mobility: to evacuate the spectral
hole when a PU is present[10].

1.2. Motivation of the Study

Spectrum is a limited resource essential for a wide range of private and national activities, faces
escalated demand driven by the rapid expansion of novel services and innovative technologies.
The escalating demand introduces challenges compelling stakeholders to proactively adjust the
formulation of new spectrum allocation and access policies. In response to these dynamic trends,
spectrum prediction has emerged as a hot research area, offering the potential to optimize spectrum
utilization. Predicting how the spectrum will be utilized in the future allows stakeholders to
anticipate and plan for evolving demands. In particular, SUs can access spectrum channels that are
not currently accessed by PUs. This proactive approach not only enhances spectrum utilization but
also mitigates interference between primary and secondary users. Therefore, spectrum prediction
can empower regulatory bodies, service providers, and technology developers to make informed
decisions regarding spectrum management. The rigidity of the command-and-control spectrum
allocation policy was considered the root cause of the problems of spectrum scarcity and
underutilization.

3|Page
Motivated by the imperative to overcome these challenges in managing spectrum resources, this
study focuses on the exploration of spectrum occupancy prediction. This study holds promise as it
offers a potential solution to the problems posed by traditional allocation policies. Specifically, it
empowers CRs to operate in underutilized spectrum bands, dynamically adapting to changing
conditions. This study has developed an analytical model to evaluate the spectrum utilization of
the candidate spectrum band and the spectrum utilization improvement obtained from spectrum
sharing. The significance of this study lies in mitigating the underutilization of spectrum that can
significantly degrade the quality of service in wireless communication. Policymakers stand to be
motivated by the prospect of spectrum sharing, recognizing it as a viable and innovative solution
for future CR deployments. Spectrum sharing promises to drive technological innovation and holds
the potential to maximize economic benefits and improve overall connectivity [5] [6].

1.3. Statement of the Problem

The statements of the problems in this study could be considered as four issues. The first is from
the spectrum sensing methods used to identify the state of the PU, whereas the second one is the
long-term spectrum occupancy state prediction i.e., the degree of predictability concerned with the
increased length of future predictions, and identifying the potential candidate spectrum band for
implementing a CR and quantifying the improvements achieved in the spectrum occupancy state
prediction model.

There is a parametric sensing approach that works based on some prior information about the states
of PU activity. In many real-world applications, the lack of prior information about the PU activity
has led to a preference for nonparametric sensing methods, with Energy Detection (ED) being a
common choice due to its low computational complexity and ease of implementation [11].
However, the wireless environment introduces challenges such as fading and hidden node
problems, resulting in an exponential decay of field strength over transmission in ED. This
drawback complicates the effective selection of threshold values, particularly when the signal-
noise ratio (SNR) is very low, making ED inefficient and susceptible to interference when used to
implement CRs [12][13][14].

4|Page
In contrast, the parametric sensing approach known as Cyclostationary Feature Detection (CFD),
offers advantages over non-parametric sensing approaches because it is not affected by noise. The
CFD sensing approach leverages the spectral correlation of cyclostationary signals, a property
absents in noise. This characteristic enables the CFD sensing approach to operate effectively in
regions of low SNR and provides better performance even when prior information about the PUs
signal is known. Cyclostationary feature detection exhibits superior performance and demonstrates
robustness to noise uncertainties in low SNR environments [14] [15].

Traditional statistical models and machine learning algorithms assume spectrum occupancy states
as stationary processes, implying that they remain constant over time and are suitable for short-
term predictions [6] [16]. Artificial Neural Networks (ANNs) are less effective for modeling
temporal data due to the absence of memory elements. Recurrent Neural Networks (RNNs) have
been employed for such tasks; however, they face challenges, such as the vanishing gradient
problem, hindering their ability to capture long-term dependencies [8] [11]. Long Short-Term
Memory (LSTM) neural networks have been introduced to address the vanishing problem. LSTMs
overcome the vanishing gradient problem by incorporating memory cells, allowing them to retain
information over extended periods. This feature is particularly advantageous for modeling
temporal data, such as spectrum occupancy [7] [8] [11] [16]. Long-term spectrum occupancy state
prediction plays a crucial role in anticipating the channel idle period duration, which reduces
channel switching costs and enhances channel selection in CRNs. Previous studies have
overlooked quantifying the improvements achieved in spectrum occupancy state prediction.
Accurately quantifying the improvements is essential for a comprehensive assessment of the
effectiveness of the models in spectrum occupancy state prediction.

1.4. Objective of the Study

1.4.1. General Objective

The general objective of this study is to design a model for spectrum occupancy state prediction
using deep learning algorithms for a cognitive radio network.

5|Page
1.4.2. Specific Objectives

To achieve the general objective, this study has carried out the following specific objectives:

 To select relevant attributes for building the spectrum occupancy state prediction model.

 To build the spectrum occupancy state model with the proposed deep learning algorithm.

 To evaluate the performance of the model.

 To quantify the improvements achieved in the spectrum occupancy state prediction model.

1.5. Scope and Limitations of the Study

1.5.1. Scope of the Study

The scope of this study is to predict the spectrum occupancy state of the GSM900 MHz mobile
network uplink band ranging from 902.5 to 915 MHz being used in Addis Abeba City.

1.5.2. Limitations of the Study

A CRN spectrum occupancy prediction involves the predictions of the occupancy state, radio
environment, and transmission rate. This study is limited to the spectrum occupancy state
prediction only.

1.6. Significance of the Study

This study will provide significant benefits for industry professionals and researchers, including:

• Improves national policymakers' and regulatory organs to perform spectrum management

techniques based on real spectrum utilization levels.
• Motivates policymakers to adjust existing policies to promote the deployment of new
services in underutilized bands.
• Informs legal and policy decisions by predicting future spectrum occupancy state, which
can lead to more efficient spectrum usage and promote spectrum sharing.

6|Page
1.7. Contribution of the Study

This study makes the following contributions to the field of cognitive radio:

1. Defining the PU channel characterization in a new mode: This study proposed a new
time-domain approach called CFD to characterize primary user states. This spectrum
sensing approach has a more accurate probability of channel detection than the ED, which
can reduce the interference between the PUs and SUs.
2. Long-term spectrum occupancy prediction: This study developed a long-term spectrum
occupancy prediction more than in the literature that can predict the spectrum occupancy state
of how long it will be busy and idle. This allows the SUs to improve spectrum access, reduce
channel-switching costs, and increase the CRN throughput.
3. Improvements achieved in the spectrum occupancy prediction: This study has
quantified the improvements achieved in the spectrum occupancy state prediction.

1.8. Organization of the Study

The study is organized as follows in seven chapters:

Introduction (Chapter One) describes the background and motivation of the study, statement of
the problem, objective of the study, objective of the study, scope, limitations, significance, and
contributions of the study.

Theoretical Background and Related Works (Chapter Two) gives an overview of what is a CR,
the main characteristics of a CR, spectrum sensing and access techniques, applications of a CR,
challenges in spectrum sensing, and a review of previous works of spectrum occupancy prediction
that have been done using DL models like ANN, RNN, GRU, LSTM, and HNN.

Research Methodology (Chapter Three) provides a brief background of methodology that can fit
the proposed system architecture's design, implementation, and evaluation and parallelly discusses
data selection, collection, and analysis in more detail the methods, procedures, and tools used to
analyze the study.

7|Page
Proposed Approach (Chapter Four) presents the proposed spectrum occupancy state prediction
and a brief description of the sequence-to-sequence LSTM deep neural network architecture used
to perform the spectrum occupancy state prediction.

Implementation Details (Chapter Five) discusses the details of the dataset used for this study,
implementation details like the setup of each experiment, the result obtained, and the evaluation
methods applied to measure the performance and the improvements achieved in the proposed
model.

Conclusion and Future Work (Chapter Six) concludes the study result and presents the future
work of the study.

8|Page
 Chapter Two

2. Theoretical Background and Related Works

2.1. Cognitive Radio

FSA policy leads to inefficient spectrum utilization as licensed users may not fully occupy all
allocated spectrum channels (or frequency), resulting in artificial scarcity. In contrast to the FSA
policy, where spectrum channels are statically assigned, CR has emerged as a solution to enable
opportunistic spectrum access for SUs in instances where spectrum channels are not used by PUs.
This approach enhances spectrum utilization, accommodating the growing demand for wireless
connectivity that enables more devices to be connected [12] [13]. The spectrum decision made in
CRNs can be characterized based on a comprehensive understanding of the spectrum occupancy
states as depicted in Figure 2.1[5].

PUs database
Spectrum characterization
(Modelling and Prediction)
PUs with assignment
of licensed spectrum
bands Historical data on
use of spectrum
bands

Request for
opportunistic Accessto the
spectrum

Dynamic spectrum management

Figure 2.1 Contextualizing the characterization of spectral decision stage in CRN

CRs have distinct characteristics that distinguish them from traditional radio systems and
Software-Defined Radios (SDRs). These distinctive characteristics are:

9|Page
 1. Cognitive Capability: CRs are equipped with the ability to sense the spectrum, allowing
them to gain insights into their environment. Through spectrum sensing, CR systems can
detect the occupancy and usage patterns of the spectrum bands. This cognitive capability
enables CRs to make informed decisions regarding spectrum utilization. For instance, they
can identify idle channels, and predict how long a channel is likely to remain idle. By
dynamically adapting to the spectral environment, they enhance spectrum efficiency and
optimize communication performance [17].
2. Reconfigurability: Unlike traditional radios that operate on fixed parameters, CR systems
can dynamically adjust their operating parameters. This adaptability includes the ability to
change frequency, modulation schemes, and transmission power levels without requiring
hardware modifications. This dynamic reconfigurability empowers CRs to respond to
changes in the radio frequency environment, mitigating interference issues and optimizing
communication quality [14] [15] [17].

CRNs are intelligent networks capable of autonomously learning and dynamically adapting to
optimize spectrum, this allows them to inherit the adaptability of the DL models that excel in
learning complex patterns from data to make informed decisions. The integration of the DL models
into CRNs with the capacity to analyze vast amounts of data and enhance the awareness of CRNs
about their operating environment holds significant potential in spectrum-sharing models. The DL
models can adapt and learn from the changing spectrum state conditions, allowing CRNs to
dynamically optimize their communication parameters, spectrum channels, and transmission
scenarios. Therefore, the synergy presents a relationship where DL's learning capabilities empower
CRNs with enhanced situational awareness for intelligent decision-making [7][18].

2.1.1. Spectrum Sensing Techniques

In CRNs, spectrum sensing techniques can be broadly categorized into two main types: non-
cooperative sensing and cooperative sensing [14] [15].

1. Non-cooperative Spectrum Sensing techniques

Non-cooperative spectrum sensing is a technique in which the CR independently decides the

presence or absence of the PU signal in any given specified spectrum band.
10 | P a g e
It does not require any coordination with other CRs. The non-cooperative sensing techniques are
further divided into ED, CFD, and matched filter detection (MFD) [15].

A. Energy Detection: is the simplest and most common sensing technique that works by
comparing the energy of the received signal to a predetermined threshold. If the energy is
above the threshold, then a PU is assumed to be present. ED is a low-complexity technique
that does not require any knowledge of the PU signal. However, it is not very reliable in
low SNR conditions, as it cannot distinguish between noise and the PU signal [13][14][15].
B. Cyclostationary Feature Detection: it works by exploiting the periodicity of the PU
signals that measure statistical parameters which are the time function. It is a more
sophisticated sensing technique that is more reliable in low SNR conditions because it is
robust to noise uncertainties. CFD is a more complex technique, due to the requirement of
prior knowledge about the PU signal [12] [13][14][15].
C. Matched Filter Detection: is the most complex sensing technique. It works by correlating
the received signal with a known template of the PU signal. MFD is the most reliable
sensing technique, but it requires knowledge of the PU signal for example, the type of
modulation used, the central frequency, and the bandwidth. Any bad information on the
signal will lead to the degradation of the detection performance [12] [13][14].

CFD is better than other techniques because it has a priori information about the signal and
performs well for a low value of SNR. However, ED performs lower than others because it does
not know a priori information about the signal, and the value of SNR significantly impacts the
performance [13].

2. Cooperative Spectrum Sensing

Transmission channels are inherently susceptible to challenges such as fading and hidden node
problems. Fading refers to the attenuation of signal strength as it traverses the transmission
channels, a phenomenon induced by environmental factors like multipath propagation. On the
other hand, the hidden node problem arises when an SU is unable to detect the presence of a PU
due to physical separation[12], [13] [13] [14]. To address these issues, Cooperative Spectrum
Sensing (CSS) techniques have been introduced. CSS involves multiple SUs collaboratively
sensing the spectrum to collectively determine the state of a PU.
11 | P a g e
This cooperative approach enhances the reliability of spectrum sensing by combining information
gathered from multiple SUs. CSS methods can be categorized into three main types: centralized,
distributed, and cluster-based [12] [14] [15].

A. Centralized CSS: In centralized CSS, a central node collects the sensing information from
all the SUs and decides the state of a PU. This is the simplest and most efficient way to
implement CSS, but it requires a central node that is always available [14][15].
B. Distributed CSS: In distributed CSS, the SUs make their own decisions about the state of
a PU, but they share their sensing information. This can improve the reliability of sensing,
but it requires more communication between the SUs [14][15].
C. Cluster-based CSS: In cluster-based CSS, the SUs is organized into clusters, and each
cluster has a cluster head. The cluster heads make decisions about the state of a PU based
on the sensing information from the SUs in their cluster. This can improve the reliability
of sensing and reduce the communication overhead, but it requires more coordination
between the SUs [15].

CSS is a technique for overcoming the challenges of spectrum sensing in wireless environments.
However, it still faces challenges, such as computational complexity and efficient information
sharing.

2.1.2. Spectrum Access Techniques

There are three main spectrum access techniques used in CRNs: underlay, overlay, and interweave.

1. Interweave: In interweave, the SUs opportunistically access the channel when it is not
being used by the PUs. This means that the SUs must sense the channel to determine
whether it is available. If the channel is available, the SUs can transmit their signals.
However, if the channel is being used by a PU, the SUs must stop transmitting and wait
until the PU is finished [19][20].

2. Underlay: In underlay, the SUs transmit their signals in the same channel as the PUs.
However, the SUs must ensure that their transmissions do not cause harmful interference.
This means that the SUs must transmit at a lower power level than the PUs [19].

12 | P a g e
 3. Overlay: In the overlay, the SUs transmit their signals in the same channel as the PUs, but
they also cooperate with the PUs. This means that the SUs may help relay the PUs' signals
or provide other types of assistance [19][20].

The choice of spectrum access technique depends on factors like the governing regulations of
spectrum use, the capabilities of the SUs, and the requirements of the PUs. The interweave
spectrum access technique is based on the prediction of the spectrum state by the PUs. The
underlay and overlay spectrum access techniques are more complex than the interweave spectrum
access technique. This is because they require the SUs to have more information about the PUs,
such as the PUs' transmission power and the channel state information [19][20].

2.1.3. Applications of Cognitive Radio

CR has the potential to revolutionize a wide range of industries, including Military, Healthcare,
Home appliances, Real-time surveillance, Vehicular networks, Addressing connectivity problems,
Content distribution networks, Smart cities, Campus-wide network coverage, Disaster relief,
Emergency networks, Weather forecasting and Traffic control [17].

2.1.4. Challenges in Spectrum Sensing

Spectrum sensing plays a pivotal role in CR systems that enable the identification of spectrum
channel states, which SUs can then opportunistically access. Nevertheless, spectrum sensing
encounters various challenges stemming from factors such as hardware constraints, hidden PUs
problems, spread spectrum PUs, security, and sensing duration in the radio frequency environment.

1. Hardware Constraints: CRs need to be able to sense multiple frequency bands to identify
unused spectrum. This requires high-performance hardware, which can be expensive[21].
2. Hidden PUs Problem: Factors like fading and shadowing cause these problems. This can be
addressed by using cooperative sensing techniques in which several cognitive radios cooperate
to enhance the overall sensing ability [14][15][21].
3. Spread Spectrum PUs: Spread spectrum PUs transmit their signals over a wide frequency
range. This makes it difficult for CRs to detect them, as they may not be able to receive enough
signal power to make a reliable detection. It can be detected with prior awareness of
synchronization pulses and hopping patterns [15][21].
13 | P a g e
4. Security Issues: A harmful user can change its air interface to mimic the PU. It gives some
wrong ideas about spectrum sensing. Public key encryption methods are used to avoid this
scenario [15].
5. Sensing Duration and Frequency: The sensing duration and frequency are the parameters
that determine how often and for how long a CR senses the spectrum. These parameters need
to be carefully chosen to ensure that the CR can reliably detect PUs without wasting too much
time or energy [15].

2.2. Related Works

Spectrum occupancy prediction is a critical problem in wireless communication technology. Many
studies have been conducted on spectrum occupancy prediction using different DL models. This
study reviews those studies that guided this study, focusing on the five selected models of ANN,
RNN, GRU, LSTM, and HNN.

2.2.1. Spectrum Occupancy Prediction Using ANN

B. G. Najashi et al. [22] proposed spectrum occupancy prediction using a neural network model
that aims to improve the reliability and accuracy of spectrum channel state prediction. The main
attraction of employing neural networks for spectrum channel state prediction in cognitive radios
is their ability to autonomously learn and adapt using historical data gathered by the cognitive
radio system, eliminating the need for complete system redesigns. First, the reliability of the model
was improved by using data obtained from cooperative spectrum occupancy measurements which
involved two spectrum analyzers that combine measurements from the two nodes instead of the
traditional method that involved a single device, because data captured in a spectrum occupancy
measurement done using a single device can be viewed as unreliable, especially in harsh channel
conditions. Cooperative spectrum sensing minimizes the effects of fast-fading and hidden terminal
problems, which can mislead SUs and cause interference with the PUs. The collaboration among
multiple spectrum sensors reduces the probability of false alarms and improves the overall
reliability of the spectrum sensing process. Secondly, the prediction accuracy of the model was
improved by employing a genetic algorithm for optimizing the weight selection of the model.

14 | P a g e
The data used for prediction was generated from five different services in a cooperative spectrum
occupancy measurement conducted in Abuja, Nigeria. The data obtained from the measurements
are combined using logical OR operators. The data used for prediction was generated from services
of 800 broadcasting, GSM900 uplink and downlink, and 3G 1865 uplink and downlink bands. The
process resulted in a total of 2,700 data sets. Each data set corresponds to a single slot having a
duration of 16 seconds, which is the time required for the spectrum analyzer to perform one
complete sweep. The accuracy of the model prediction was as follows: GSM 900 uplink: 96.77%,
GSM 900 downlink: 99.85%, 3G uplink: 98.25%, Broadcasting: 99.93%, and 3G downlink:
99.75%. The difference in prediction accuracy between the uplink and downlink bands arises
because base stations continuously transmit information to users in the downlink band, leading to
more stable and predictable patterns. In contrast, the uplink band is used more sparsely, exhibiting
less predictable patterns due to the randomness of occupancy.

D. D. Das et al. [23]proposed the design of a simple and faster spectrum prediction model known
as a Functional Link Artificial Neural Network (FLANN) to forecast the future spectrum usage
profile and to obtain utilization statistics for the industrial, scientific, and medical bands of 2.4–
2.5 GHz. The study conducted measurements indoors at the Swearingen Engineering Center,
University of South Carolina over five working days from Monday to Friday to present the
occupancy statistics and gather realistic data for assessing the performance of the proposed model.
Models such as MLP, RNN, and RBF are good for learning in CRNs, but they have high
computational cost and complexity due to the hidden layer. However, FLANN is a novel single-
layer model that eliminates the need for a hidden layer and is proposed to replace MLP due to its
faster convergence rate and lower complexity. It operates on the entire input pattern using a set of
linearly dependent functional expansions. Spectrum occupancy was estimated and predicted by
employing different ANN models including MLP, RNN, Chebyshev FLANN, and Trigonometric
FLANN. It is observed that the absence of a hidden layer in FLANN makes it more efficient than
other ANN models in predicting the occupancy in less computational time and with less
complexity, which achieved the best prediction accuracy of 99%. This suggests that FLANN is a
promising approach for an effective spectrum occupancy prediction to designing intelligent
learning systems for CR applications.

15 | P a g e
2.2.2. Spectrum Occupancy Prediction Using RNN

R. Fan et al. [24]proposed a multi-channel state prediction method based on the RNN model to
address the scarcity of spectrum resources and enhance spectrum utilization efficiency in CR. First,
they generated synthetic data to address the issue of how to model the primary user behavior in
CR considering each PU has M channels and B Hz bandwidth, which are logically mutually
independent ones.

Second, the historical information of the PU channel state behavior in CR by representing a binary
sequence has been used to perform spectrum occupancy state prediction using the RNN. The RNN
takes the multiple channel occupancy state information with a time step and uses the previous L
time slots to predict the channel state at the L+1 time slot. Then, the predicted channel state at the
L+1 time slot and the preceding historical information was used to predict the channel state at the
L+2 time slot, and so on until the required length of prediction was achieved. The proposed multi-
state predictor provides better performance when the time step increases. However, the
improvement does not continue to increase when the time step reaches a certain length. Because
the historical information strongly correlates with the channel state at the next moment for certain
time steps only. Therefore, when the prediction length increases, the prediction accuracy will
decrease due to information fading in learning the long-term dependence of RNN, leading to a
problem known as vanishing gradient. The vanishing gradient problem is encountered in training
RNNs when the gradient of the loss function for the parameters of the RNN tends to zero as the
time step increases. This makes it difficult for the RNN to learn long-term dependencies. To reveal
the proposed model advantage, it has been compared with other existing spectrum occupancy state
prediction models, numerical results illustrate this method has better performance than other
existing models and can provide sufficient information to the secondary users to utilize spectrum
holes.

Learning long-term dependence is the biggest challenge in RNN. Gated RNNs (GRUs) that have
memory blocks to control the flow of information within the network and were developed to
address this challenge, allow to selectively remember and forget information over long periods.

16 | P a g e
GRU has the same structure as the RNNs, but it has improved in its hidden layer to maintain long-
term memory achieved through memory blocks to regulate the flow of information into and out of
the hidden state. The GRU has a simple and faster structure consisting of only two gates, namely
the reset and update gates. The reset gate determines how much of the previous hidden state
information should be forgotten, while the update gate determines how much of the new input
information should be used to update the hidden state [25].

2.2.3. Spectrum Occupancy Prediction Using GRU

L. Xing et al. [26] proposed a GRU-based deep learning model for learning complex patterns in
spectrum usage and adopted a time series prediction for multiple channel states in CR networks.
This implementation of intelligent spectrum sensing and dynamic spectrum access enables the
prediction of spectrum usage, facilitating compliance with spectrum usage regulations and
improving overall spectrum utilization. They have investigated the use of channel state values from
historical time slots to predict the channel state values of future time slots. The study examined
the impact of input and output sequence lengths on prediction accuracy and discussed the
performance differences between multi-user joint prediction and single-user independent
prediction. After the channel state data is segmented into slices each slice represents channel states
over a specific time interval. Each channel state sequence is then input into the model to obtain the
corresponding output of the predicted future channel states. First, they evaluated the input
sequence length effect on the model's performance. More accurate predictions were achieved
within longer input sequence lengths containing more historical information. However, the
prediction accuracy improvement is limited to a certain length because the model has a finite length
of data dependence. Next, they compared the GRU-based model prediction accuracy with an MLP
model. The GRU-based model achieved better prediction accuracy than the MLP model, especially
when the input sequence length was too large. Second, they evaluated the output sequence length
effect on the model's performance. The longer the output sequence length, the more accurate the
prediction is for the immediate future. However, the prediction accuracy decreases for longer
output sequence lengths. Lastly, they evaluated the joint prediction of multiple channels by taking
four channels as one regarding them as independent and uncorrelated. The comparison with the
independent prediction shows that it has more accuracy than predicting each channel
independently because it can consider the dependencies between the channels.
17 | P a g e
2.2.4. Spectrum Occupancy Prediction Using LSTM

M. A. Aygul et al. [27] proposed an LSTM model to predict the next spectrum occupancy state
because identifying spectrum opportunities is crucial for efficient spectrum utilization in CR
systems. The aim is to improve spectrum occupancy prediction accuracy while reducing
computational complexity. Spectrum occupancy states correlated over time, and spectrum
prediction can be effectively treated as a time-series process.

This study demonstrates the advantage of leveraging both the correlation over frequency and the
correlation over time to enhance the accuracy of spectrum occupancy predictions. By exploiting
these correlations, this study investigates a more realistic scenario using a 2D LSTM model. This
model incorporates previous time and frequency spectrum measurements to predict future
spectrum occupancy states. The model was trained on data obtained from real-world spectrum
measurements collected over one hour in a city surrounded by commercial and residential areas.
The measurements were taken in the frequency range of 832-862 MHz, which is used by most
telecom operators in Turkey for the uplink band. The model predicts spectrum occupancy by using
known binary values to forecast the corresponding spectrum occupancy state at the next time
instant. Extensive experimental results demonstrated that the proposed approach outperformed
classical prediction methods in terms of accuracy, computational complexity and the efficiency of
time-based prediction using DL models. The proposed model was compared to classical prediction
models in terms of precision, recall, and F1-score. The model was also easy to train and converge
faster than the other models. Therefore, the proposed LSTM model is a promising approach to
improve the accuracy and computational complexity of the spectrum occupancy prediction.

A. Shenfield et al. [7] proposed an LSTM-based DL model to predict the availability of the
spectrum the next time to reuse the underutilized spectrum frequencies, which is a promising
approach to overcoming spectrum limitations. This allows the SUs to select more effectively the
channels with the highest probability of being free the next time. The proposed model was
evaluated using both synthetic and real data sets. The results show that the proposed model
outperforms the Markov chain-based model in predicting the availability of a channel for multiple
future steps. For example, the probability of success in selecting an idle channel reaches 90% for
two-step ahead predictions and 80% for five-step ahead predictions using a synthetic data set.

18 | P a g e
This indicates that the proposed model algorithm can also minimize the channel switching cost by
accurately selecting a free channel for more than one-time slot ahead. The proposed model, tested
on real datasets, achieves a selection accuracy of 80% for two-step-ahead predictions and 70% for
five-step-ahead predictions. This represents, on average, a 5% improvement in the probability of
success compared to other existing methods in the literature.

L.Yu et al. [28] built an LSTM model to predict future spectrum availability at arbitrary locations
by leveraging the intrinsic spectral-temporal correlations. This approach solves the problem of
spectrum availability in cognitive aerospace communications, which is more effective and efficient
than a model based on temporal correlations only. The model sets a relatively large number of
nodes to better characterize the dynamic states of spectrum occupancy and more effectively
capture the features of the input data. The model performance was evaluated using real-time
spectrum occupancy data gathered from the cities of New York and Vienna. The data covered the
frequency range of 3MHz to 5.4MHz, divided into 26 channels. The channel's availability was
predicted by setting the look-back window length, which is the number of past observations used
to make a prediction, from 15 to 90 time slots. The results showed that the model achieved an
accuracy of 94.06% when the look-back window length was 60. However, when the look-back
window length was set to 75, the model accuracy improved to 98.14%. This shows that the model
can learn long-term temporal correlations in the spectrum data, which helps to improve the
predictions' accuracy. The model was also compared to an ANN model. The results show that the
LSTM model outperforms the ANN model in terms of accuracy. This is because the LSTM model
can capture the long-term dependencies in the data.

L.Yu et al. [29] an LSTM-based model with multiple layers was constructed and trained through
supervised learning to predict frequency hopping sequences. This model is designed to resist
interference by accurately predicting the frequencies used by adversaries. By correctly predicting
these interference frequencies, the model enables effective avoidance, thereby enhancing the
robustness of communication systems against interference. The output of the LSTM network is
transformed into 0 or 1 which indicates the idle state or occupied compared to the threshold. The
simulation data used in this experiment consisted of frequency hopping patterns with a length of
160-time slots across ten channels. The model achieved a prediction accuracy of 90%.

19 | P a g e
They have compared the influence of the LSTM network depths and widths on the prediction
accuracy performance with a Back Propagation (BP) model. The prediction accuracy of the BP
model increases when the network depth increases, while the prediction accuracy of the LSTM
model increases when the network width increases. The experiment results show that the LSTM
model has better predictive performance than the BP model. The shallow and wide LSTM neural
network model captures the characteristics of the data set better than the deep and narrow LSTM
neural network model.

However, the prediction performance of the BP network did not change significantly under similar
situations. When there is something wrong with spectrum sensing, channel states with error and
correct channel states both can be used as learning labels. The simulation results indicate that
prediction accuracy for the latter one is about 4% better than the previous one.

B. S. Shawel et al. [30] proposed a Convolutional Long Short-Term Memory (ConvLSTM) model
to enhance long-term temporal prediction by learning the joint spatial-spectral-temporal
dependencies observed in spectrum usage. This approach seeks to alleviate the inefficient use of
the radio spectrum through DSA with CR enabling opportunistic users that share spectrums
without causing interference when these spectrum channels are not being used by their PUs. The
data used to train the ConvLSTM model was collected from a real-time environment measurement
in the UHF bands of 450-520 MHz to evaluate the prediction accuracy of the proposed network
for increasing future time steps and different spectrum channels. Five sparsely distributed sensors
were used to collect the data and the signal power level was sent to a central entity for decision-
making. The ConvLSTM model predicted the spectrum occupancy for the next 3 hours. The
prediction's accuracy was assessed utilizing the root mean square error (RMSE), yielding a value
of 5.012%. During prediction, an increase in future time steps leads to a higher accumulated error
from previous predictions, resulting in an increased RMSE. However, the RMSE values have
remained below 5.012, the benchmark set by the combined sample standard deviation. This
demonstrates that the predictions effectively capture spectral dependencies and long-term temporal
spectrum usage patterns, even though they may not be as accurate for short-term changes.

20 | P a g e
Spectrum sensing and prediction from a single SU perspective is unreliable under harsh
environmental conditions. This is because the SU does not have a clear view of the spectrum
channel state and is affected by problems such as hidden terminals, shadowing fading, and
multipath fading [8]. To address these challenges, cooperative spectrum sensing has been
proposed. In cooperative spectrum sensing, multiple SUs collaborates to sense the spectrum and
make a more reliable decision about the spectrum channel state. This approach can improve the
detection of the channels and reduce the probability of interference. However, cooperative
spectrum sensing faces challenges such as the need for more hardware-intensive devices for
sensing and the time required to combine the data collected from multiple devices [15].

2.2.5. Spectrum Occupancy Prediction Using HNN

S. S. Shirgan et al. [31] proposed a Hybrid Neural Network (HNN) model that combines Radial
Basis Function (RBF) and Multilayer Perceptron (MLP) techniques to enhance prediction
performance by proper learning for the deployment of CR, intending to achieve energy-efficient
and time-efficient spectrum analysis. The measurement results of spectrum occupancy utilization
analysis provide information only about the current occupancy status, not future usage. In the
context of spectrum sharing and mobility in CR, offering cognitive users a predefined channel list
in advance would be a more effective solution to mitigate interference with existing PUs.
Therefore, CR performance can be enhanced by adopting predictive methods to analyze spectrum
occupancy and determine the future occupancy status of PUs. First, the spectrum occupancy
analysis was performed using the statistical method of data collected from a seven-day
measurement campaign. The results show that the overall occupancy was 6.32% in the TV band,
45.24% in the GSM900 MHz, 36.91% in the GSM1800, and 13.09% in the UMTS2100. Second,
the HNN model predicts the spectrum occupancy at weekends and weekdays, the result shows the
HNN prediction outperformed the MLP and RBF model's prediction accuracy by combining the
strength of both MLP and RBF. Even though MLP is simple to understand and implement, it
requires more hidden layers for better approximation, this makes it difficult to train and slows
down the learning process. RBF has a single hidden layer with a Gaussian activation function,
which makes it faster to train than MLP. However, RBF is more complex to understand and
interpret.

21 | P a g e
 The model's performance was evaluated using the RMSE of the four bands' predictions: 1.5%,
4.5%, 0.5%, and 0.7%, respectively. They also tested the performance of the HNN model for seven
days. The mean expectation error was a minimum of 0.036. The model was found to perform
poorly on Sundays than on weekdays due to the unpredictable behaviour of the spectrum on
holidays. The performance of the model examined for popular bands at weekends and weekdays
was found to be accurate compared to MLP, and RBF. This proves that by combining MLP and
RBF, the HNN can achieve the best of both worlds. It is simple to understand and implement as
MLP, while also being as fast to train as RBF.

2.2.6. Summary of Related Works

Table 2.1 Summary of Related Works

Authors Methodology Objective Gaps

B. G. Najashi ANN To improve the reliability and accuracy of Short-term prediction.

et al. [22] spectrum occupancy prediction. Cooperative SS.

R. Fan RNN Synthetic data.

et al. [23] To predict the future channel state does not solve the vanishing
multiple steps ahead. gradient problem.
M. A. Aygul LSTM To improve the spectrum occupancy Short-term prediction.
et al. [27] prediction accuracy and computational
complexity.
B. S. Shawel ConvLSTM Long-term prediction. Computationally complex and
et al. [30] more hardware intensive.

S. S. Shirgan HNN To achieve better spectrum prediction Short-term prediction.

et al. [31] performance.

The table above summarizes some studies that predict spectrum occupancy. These studies have
achieved promising results, but they have limitations like performing short-term predictions that
do not consider the effect of channel state on future time slots, using unreliable spectrum sensing
methods, and not identifying potential candidate spectrum bands. This study addresses these
limitations by developing a reliable spectrum sensing method that can correctly detect the state of
the PU, identify the potential candidate spectrum bands based on a proper analysis of spectrum
utilization, and develop a long-term spectrum occupancy prediction model.

22 | P a g e
 Chapter Three

3. Research Methodology

3.1. Overview
This chapter elaborates on data selection, collection, pre-processing, details about data analysis,
and the tools used to conduct this study. First, the source of data and the collection methods applied
were explained. Next, it explains the data pre-processing activities like feature selection that have
been used to select essential features for building the model, cleaning the data, and transforming
the data into a format suitable for the model building where feature engineering and data analysis
were performed. Finally, the tools used to analyze the data and to conduct the study, and the
performance evaluation metrics used in this study have been presented.

3.2. Data Collection

The data used for this study was collected from a real-world spectrum measurement in Addis
Abeba, Ethiopia using the TCI spectrum monitoring system. The real-time measurement was
conducted in the GSM900 MHz mobile network uplink spectrum band ranging from 902.5 to 915
MHz for five consecutive days from January 28th to February 1st, 2021. The area of Bole was
selected for spectrum measurement due to its status as a commercial hub, which is expected to
have a higher spectrum demand than other areas [31]. The resolution bandwidth of each spectrum
channel was 100 kHz with 4 minutes resolution time. This resulted in a total of 450,000
measurement data points for five days. The GSM900 MHz uplink band presented to be a promising
potential candidate for the deployment of a CR due to its underutilization from the sparse use of
its users communicating on the network [22] [32] [31].

3.2.1. Description of Data Features

The data features contain spectrum measurements that include Channel, Frequency, Maximum
occupancy (%), Average occupancy (%), Maximum field strength, and Average field strength.
These attributes are time series data attributes that can be used for the implementation of the
spectrum occupancy state prediction.

23 | P a g e
The following table 3.1 is a description of all the spectrum measurement data features or attributes.

Table 3.1 Description of spectrum measurement data attributes

Attribute Type Description

Channel Numeric Channel number
Frequency Numeric Frequency of a channel
Max Occupancy Numeric Percentage of maximum occupied time
Average Occupancy Numeric Percentage of average occupied time
Max power gained Numeric Values of maximum power obtained
Average power gained Numeric Values of average power obtained

3.3. Data Pre-processing

Data pre-processing is all about cleaning the data and removing the noise, selecting the features to
reduce the data dimension, and transforming it into a format appropriate for the proposed model.
So, before it has been applied to the proposed model, the raw data has passed through several
transformation steps that provide a significant data quality to improve the model performance. In
this study, after the spectrum measurement data has been obtained from the spectrum monitoring
system, it has been arranged in a proper format. Data pre-processing was started with feature
selection parallelly used for data analysis, and after that data cleaning was performed. Finally, data
transformation has been applied.

3.3.1. Feature Selection

All the data features obtained from the spectrum measurement are not necessary to predict the
spectrum occupancy for the proposed model. The TCI spectrum monitoring system presents two
types of attributes, some have valuable information about spectrum occupancy while others don’t.
For this study, feature selection was performed by selecting only the features that have relevant
information for the spectrum occupancy state prediction model, and ignoring features that don’t
have valuable information [33]. The data collected from the TCI spectrum monitoring system have
six features which are Channels, Frequency, Maximum occupancy (%), Average occupancy (%),
Maximum field strength, and Average field strength.

24 | P a g e
For modeling spectrum occupancy prediction, three features were selected based on filter-based
feature selection [34]: Frequency, Average Occupancy (%), and Average Field Strength. These
features were chosen because they more effectively describe spectrum occupancy. In contrast, the
maximum values of occupancy and field strength represent only the peak values observed at a
single instance during the measurement period, which do not accurately reflect the overall state of
spectrum occupancy.

3.3.2. Data Cleaning

Cleaned data was used to improve the performance of the prediction model. Data quality problems
arise from mistaken data entry, missing values, redundant, and invalid data. Data cleaning was
performed by removing redundant data using appropriate techniques to remove inconsistencies
that helped to improve the performance of the proposed model.

3.3.3. Feature Engineering

Data found in the real world can be messy. Feature engineering transforms such data into a
common understandable format. It ensures that all numerical values are on the same scale to avoid
computational error, making the data easy to interpret and understand. It can also improve the
model performance by incorporating domain knowledge, such as classifying the raw numeric
values obtained from the spectrum measurement into two states: i.e., idle or busy [35]. In this
study, depending on the spectrum measurement numerical value, if the duty cycle is lower than
51% in the CFD-based SS, the channel can be considered “idle” because mostly the spectrum
channel was not in use.

3.4. Setup the Data

After all data pre-processing tasks have been completed, the collected dataset is divided into
training and validation sets with a ratio of 80%, and 20%, respectively. The data setup has 180,000
data points selected for training, and the other 45,000 data points were used for validating the
prediction algorithms. The validating points were used to validate the performance of the
prediction algorithm.

25 | P a g e
 3.5. Design and Implementation Tools

Many types of tools were used to conduct this study, starting from data collection and pre-
processing to the design and implementation of the long-term spectrum occupancy prediction. The
following design and implementation tools, which can be hardware or software, were used during
the study for the design, implementation, and reporting of the study document:

3.5.1. Hardware Tools

The following table shows the hardware tool used in this study along with its specific function
throughout the study.

Table 3.2 Hardware tools used for implementation of the study

S.no Device name Used in the Study

The TCI spectrum monitoring system was used to collect
1 TCI Server
the spectrum occupancy data.
2 Hard Disk Used as storage for large datasets.
3 GPU To increase the computation and to fasten the training.
4 RAM To accelerate the training process cooperatively with GPU.

3.5.2. Software Tools

Software tools and libraries were used in the study to write the code, debug the program, visualize
results, collect data, and report the study document.

Scorpio Client: is an application software that provides an interface to the measurement server
(Scorpio Server). The Scorpio client provides all functions necessary to arrange the measurement
parameters, to download the data from the server, and to select the appropriate data features from
the data.

Anaconda: is an application used to install Python programming language with its all modules. It
provides a navigator application to view different settings like Jupyter Notebook, spider, vs-code,
and modules installed in the environment.

26 | P a g e
Python: Programming language used to perform the implementation of the study.

Jupyter Notebook: An interactive web-based application that helps to configure, load Python
API, and to write Python code.

Keras: API with easy extensibility to work on CPU and GPU, that supports modularity, and works
with Python as a high-level framework.

TensorFlow: a machine learning library used to build a model and deploy it in client environments.
It supports ML, DL, and other types of flexible numerical computation.

Microsoft Office Packages: Tools used to write and organize the study like Word to write the
document, Excel to make the CSV dataset, Visio to draw diagrams and write equations, and
PowerPoint to make the presentation slides.

Mendeley Desktop: is a powerful reference manager tool that serves as an academic, social
network for referencing similar works.

Bib Guru: Excellent reference and citation generator tool that can quickly add sources paper and
make citations in IEEE and other styles.

3.6. Model Performance Evaluation Metrics

A model can be evaluated using different metrics to quantify its performance on a test set. These
metrics measure how the model fits the data, works on new data, and makes no discrimination
among the model results. Some metrics used for the prediction of time-series data include mean
squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), and mean
absolute percentage error (MAPE) [36] [37].

3.6.1. Mean Squared Error

It is the average of the square of the difference between the predicted values and actual values. It
has the same units as the actual and predicted values squared and is always positive.

𝑛
1
𝑀𝑆𝐸 = ∑( 𝑦′𝑡 _ 𝑦𝑡 )2 3.1
𝑛
𝑡=1

27 | P a g e
Where 𝑦′𝑡 is the predicted value, 𝑦𝑡 is the actual value, and n is the total number of values in the
test set. It is clear from the equation that MSE is more penalizing for larger errors or outliers.

3.6.2. Root Mean Square Error

It is the square root of the mean square error. It is also always positive and is in the range of the
data.

𝑛
1
𝑅𝑀𝑆𝐸 = √ ∑( 𝑦′𝑡 _ 𝑦𝑡 )2 3.2
𝑛
𝑡=1

Where y′t is the predicted value, yt is the actual value, and n is the total number of values in the
test set. It is in the power of unity and is more interpretable than MSE. RMSE is also more
penalizing for larger errors.

3.6.3. Mean Absolute Error

It is the average absolute difference between predicted and actual values. It has the same units as
the predicted and actual value and is always positive.

𝑛
1
𝑀𝐴𝐸 = ∑|𝑦′𝑡 _ 𝑦 𝑡 | 3.3
𝑛
𝑡=1

Where y′t is the predicted value, yt is the actual value, and n is the total number of values in the
test set.

3.6.4. Mean Absolute Percentage Error

It is the average absolute difference percentage between predicted values and actual values,
divided by the actual value.
𝑛
1 𝑦′𝑡 _ 𝑦 𝑡
𝑀𝐴𝑃𝐸 = ∑ | | ∗ 100% 3.4
𝑛 𝑦𝑡
𝑡=1

Where y′t is the predicted value, yt is the actual value, and n is the total number of values in the
test set.

28 | P a g e
3.7. Confusion Matrix

A confusion matrix is a table that shows how well a model is performing on a given dataset. It
compares the actual values of the data to the predicted values.

1. True positive (TP): This is the condition when both the actual and the predicted values
are true. This is known as the detection probability in a CRN.
2. True negatives (TN): This is the condition when both the actual and the predicted values
are false. This is known as the identification probability in a CRN.
3. False positives (FP): The condition when the actual value of the data is false whereas the
predicted value is true. This is known as the false alarm probability in a CRN.
4. False negatives (FN): The condition when the actual value of the data is true whereas the
predicted value is false. This is known as the miss-detection probability in a CRN.

The confusion matrix can be used to calculate several performance metrics, such as accuracy,
precision, specificity, recall, and F1 score. These metrics can be used to compare the performance
of different models and to identify areas where the model can be improved.

1. Accuracy: Accuracy is the percentage of predictions that are correct. It is calculated by

dividing the number of correct predictions by the total number of predictions.

𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 3.5
𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁

2. Precision: Precision is the percentage of predicted positives that are actually positive. It is
calculated by dividing the number of true positives by the number of true positives plus the
number of false positives.
𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃+𝐹𝑃 3.6

3. Recall: Recall is the true positive rate used to measure the fraction of positive values that
are correctly predicted.
𝑇𝑃
𝑅𝑒𝑐𝑎𝑙𝑙 = 𝑇𝑃+𝐹𝑁 3.7

29 | P a g e
4. Specificity: Specificity is the true negative rate used to measure the fraction of negative
values that are correctly predicted.

𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = 𝑇𝑁+𝐹𝑃 3.8

5. F1_score: The F1-score is a weighted average of precision and recall or the harmonic
mean between recall and precision, used to measure the model's overall accuracy.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙
F1_score = 2 ∗ 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑅𝑒𝑐𝑎𝑙𝑙 3.9

30 | P a g e
 Chapter Four

4. Proposed Approach

4.1. Overview
This chapter discusses the details of the proposed approach and the model architectures that
include the block diagram, flow charts, and the algorithms used to solve the study problem.
Moreover, a brief description of the sequence-to-sequence LSTM deep learning neural network
architecture used in this study to predict spectrum occupancy state from the frequency spectrum
and temporal perspectives used was explored.

Figure 4.1 The Proposed Deep Learning Approach

31 | P a g e
4.2. System Model

This study aims to predict the spectrum occupancy state used to model a CRN. The heterogeneous
spectrum occupancy state model was used to implement a CRN for spectrum sharing in a DSA
model [7]. The spectrum band has been divided into k contiguous frequencies or channels. The
channel states represent the spectrum channel state at time t and are denoted by a vector
matrix. Each element in the matrix represents the corresponding channel that is occupied, or idle
which is ready to be used by the opportunistic SU [28]. A heterogeneous spectrum occupancy state
model has multiple PUs and SUs, which are centrally controlled by a database that identifies the
spectrum occupancy states based on prediction[4]. This study has proposed a long-term spectrum
occupancy state prediction model in a stationary location by exploiting the spectral and temporal
correlation of the data. The occupancy of a channel is characterized by the presence of a PU signal,
while the presence of a spectrum hole characterizes the vacancy of a channel. These cases are
formally stated as hypotheses (H0) and (H1).

𝐻0 : 𝑦[𝑡] = 𝑤[𝑡] when there is no PU

𝐻1 : 𝑦[𝑡] = ℎ[𝑡]𝑥[𝑡] + 𝑤[𝑡] when PU’s the signal is present 4.1

where x[t] denotes the PU signal, w[t] is white noise and y[t] is the received signal at 𝑡𝑡ℎ time
instant. 𝐻0 , the null hypothesis indicates the noise samples while 𝐻1 , the alternate hypothesis
indicates the presence of PU signal along with noise 𝑡𝑡ℎ instant [6] [12] [17].

Figure 4.2 Spectrum channels occupancy state modeling

32 | P a g e
For the sequentially obtained time-series spectrum measurement data X1, X2, X3, X4,…, the long-
term spectrum channel state prediction using the deep learning model can be done within the
sequence-to-sequence neural network architecture in the LSTM deep learning model is defined as
(Xt-n,…., Xt-2, Xt-1, Xt) to (Xt+1, Xt+2, Xt+3,…., Xt+m) where n and m represent the historical observations
and the future instants in time, respectively [30].

4.2.1. Time-Series Model Analysis

Time series data is composed of sequence data points measured over regular intervals. Time series
models are designed to comprehend the patterns and trends inherent in the data that can feat the
natural temporal ordering. This shows that observations closer in time are more similar than further
apart observations because values in a time series data at a given time were derived from past
values. Time series analysis is a methodology used to study time series data that identifies
relationships and makes predictions of future values[38]. Time series data analysis begins by
identifying whether the data is stationary or non-stationary. A stationary time series has consistent
statistical properties, including mean, variance, and autocorrelation. This stability implies that the
underlying data-generating process is predictable and stable. On the contrary, non-stationary time
series data displays varying statistical properties over time, often influenced by trends, seasonality,
or other random fluctuations. Time series prediction is a technique that estimates future values
based on historical data. In spectrum occupancy state prediction, the objective is to forecast the
state of the spectrum occupancy state at the next time point [38][39].

Time series data should be selected carefully, considering the different variations that can occur at
different timescales. For example, a day can have four seasons (morning, afternoon, evening, and
night), while a week can have only two (weekday and weekend). The spectrum occupancy data
can vary significantly depending on the time of day, day of the week, and peak or trough times.
Therefore, representative data must be chosen for the specific period and timescale of interest [32].

4.3. Proposed Deep LSTM Architecture

This study proposed a long-term spectrum channel state prediction using a deep-learning LSTM
model. The method uses a sequence-to-sequence neural network architecture based on the LSTM
to predict the spectrum channel state based on preceding channel states.

33 | P a g e
The proposed model can accurately predict the spectrum channel state for the next time slot and
several time slots ahead. The proposed long-term spectrum channel state prediction model is then
implemented to facilitate a DSA. This model enables users to access spectrum channels that are
not used by PUs, enhancing overall spectrum utilization efficiency.

Figure 4.3 Basic Architecture of the LSTM Model

Conventional RNNs can model the temporal dynamic behavior of sequential inputs, but they have
difficulty learning long-range dependencies. This is because backpropagating errors through long
sequences can cause the vanishing gradient problem. LSTM networks solve this problem by using
a gating mechanism to control the information flow in the network. The gating mechanism consists
of three gates: the input gate, the forget gate, and the output gate. And the cell memory that is used
to store information over long periods [7] [8].

1) The input gate which controls how much new information is added to the LSTM's memory
state.

2) The forget gate which controls how much information is removed from the memory state.

34 | P a g e
3) The output gate which controls how much information from the memory state is output by the
LSTM cell.

The gates are mathematically formulated as:

𝑖𝑡 = σ(𝑊𝑖𝑥 𝑋𝑡 + 𝑊𝑖ℎ ℎ𝑡−1 + 𝑊𝑖𝑐 𝑐𝑡−1 + 𝑏𝑖 )

𝑓𝑡 = σ (𝑊𝑓𝑥 𝑋𝑡 + 𝑊𝑓ℎ ℎ𝑡−1 + 𝑊𝑓𝑐 𝑐𝑡−1 + 𝑏𝑓 )

𝑐𝑡 = 𝑓𝑡 ◦ 𝑐𝑡−1 + 𝑖𝑡 ◦ φ(𝑊𝑐𝑥 𝑋𝑡 + 𝑊𝑐ℎ ℎ𝑡−1 + 𝑏𝑐 )

𝑜𝑡 = σ(𝑊𝑜𝑥 𝑋𝑡 + 𝑊𝑜ℎ ℎ𝑡−1 + 𝑊𝑜𝑐 𝑐𝑡 + 𝑏0 )
ℎ𝑡 = 𝑜𝑡 ◦ φ (𝑐𝑡 ) 4.2

Where i, f, o, and c denote the input gate, forget gate, output gate, and cell state, respectively. All
these gates are on the same dimension as the hidden vector h is assumed to be of N ×1 dimension.
σ is a sigmoid function, and φ is a nonlinear function that maps the input to [-1, 1]. Wic, Wfc, and
Woc are the peephole connection matrices that connect cell state to their respective gates. Whereas,
Wix, Wfx, Wox, and Wcx are the weight matrices that connect the input vector 𝑋𝑡 and input gate,
forget gate, output gate, and cell state, respectively. Because the gates and the input vector Xt have
the dimensions of N ×1 and M × 1 respectively. The dimensions of matrices can have Wih, Wic,
Wfh, Wch, Woh, and Woc are all the same, which is N × N, and the dimensions of the matrices Wix,
Wfx, Wcx, and Wox are N × M.

4.3.1. Proposed Deep Learning LSTM Basic Architecture

In the model presented in [27], illustrated in Figure 4.3, an LSTM neural network architecture was
employed for long-term spectrum occupancy prediction. The model design focuses on a specific
set of frequencies, with known binary values provided for spectrum occupancy state prediction
within the model. The input and output data for the model are constructed using a sliding window
that traverses both the time and frequency axis’s. This process generates a 2D matrix from the
spectrum measurement data, where each element corresponds to a specific point in time and
frequency, along with its associated binary value. This matrix serves as the input dataset used to
train the model. During the validation stage, the proposed model demonstrates its operational
capability in real-time.

35 | P a g e
Spectrum measurements for time and frequency lags were utilized to predict the corresponding
spectrum occupancy state at the next instant. This prediction is achieved by inputting the binary
values, arranged as a grid, to the model, which then generates the binary values for the subsequent
instant [27].

Figure 4.4 Learning and predicting long-term spectrum occupancy in LSTM

In this study, the CRN model was developed to enhance spectrum resource utilization through
long-term spectrum occupancy prediction. The model is conceptually likened to a smart parking
management system that has two types of customers, illustrating its capability to optimize the use
of spectrum resources akin to the efficient utilization of parking spaces. Much like a smart parking
system identifies and directs vehicles to available parking spaces, the CRN can identify both
available and occupied spectrum channels. It then guides SUs to optimal spectrum channels,
providing information about their availability and duration. This approach aims to optimize
spectrum utilization in a manner analogous to how a smart parking management system optimizes
parking space usage. Therefore, the two technologies have similarities in their operation,
emphasizing resource optimization, dynamic adaptability, and the pursuit of key objectives.

4.3.2. Proposed Deep LSTM Architecture Components

The sequence-to-sequence LSTM model used for spectrum occupancy state prediction consists of
the following major components [28]. The main components of the LSTM units are:

36 | P a g e
1. The Main Framework

First the input goes through the LSTM layer to extract long-term dependencies from the input
sequence, and the output goes into the full connected LSTM dense layer, where a dropout process
is applied to avoid overfitting in training. Finally, the Sigmoid activation function is employed to
obtain the final prediction result. The input Xt in the time slot t is a matrix of dimension M × T,
which means that the system predicts the spectrum occupancy state in the forthcoming time slot
by exploiting the data in the most recent T time slots. In the input, each column represents the
spectrum occupancy state in a time slot.

2. LSTM Layer

The input sequence undergoes processing through an LSTM layer, a RNN tailored for capturing
and learning long-term dependencies within sequential data. This layer effectively analyses the
input sequence, enabling the extraction of long-term patterns. The final output of the LSTM layer,
representing the output of the last LSTM memory cell, is subsequently fed into a dense network
for further processing.

3. Dense Network with Dropout

The dense layer is used to reduce the dimensionality and to produce the prediction final output, in
the dense layer every node is connected to all nodes in the preceding layer. Specifically, it receives
the output from the LSTM layer and learns to map it to predict spectrum occupancy in the
subsequent time slot. To address overfitting, the model incorporates dropout, a regularization
technique that helps to prevent overfitting. Dropout randomly drops out some of the neurons in the
dense layer during training to prevent the model from overfitting and enhance the prediction
performance.

4. Activation Function and Loss Function

In the computation of the final prediction output, the activation function chosen is Sigmoid, applied
to the output of dense networks. Specifically, this activation function maps the output of the dense
networks into a vector of elements ranging between 0 and 1.

37 | P a g e
Each element in the vector represents the probability of a channel being occupied, and the sum of
all elements equals 1. To map the probabilities into binary series during training, an optimal
threshold is determined. This threshold minimizes the disparity between actual data and the
predicted values, which are either 0s or 1s. For learning and evaluation, a loss function is defined.
In this system, cross-entropy is used as the loss function. This choice is motivated by the
classification nature of the problem, where the goal is to categorize inputs into two labels. Cross-
entropy has been widely adopted and demonstrated to be effective in classification tasks, making
it a suitable choice for this work.

Figure 4.5 LSTM architecture with input Xt in time slot t

38 | P a g e
 Chapter Five

5. Implementation Details, Result and Discussion

5.1. Introduction

This chapter details the data analysis and feature description, the experimental setup and
configuration for each experiment, the evaluation methods applied to measure the model
performance, the results obtained, and discussions about the improvements achieved in long-term
spectrum occupancy prediction.

5.2. Data Set Description

The dataset used for this study was collected using the TCI spectrum monitoring system. The
dataset contains spectrum occupancy measurements from January 28, 2021, to February 1, 2021.
The dataset includes Channel, Frequency, Maximum occupancy (%), Average occupancy (%),
Maximum field strength, and Average field strength. These attributes are time series data attributes
that were used for the implementation of the spectrum occupancy state prediction.
Table 5.1 A sample data obtained from the spectrum measurement campaign

Channel Frequency Occupancy (%) Field Strength (dB)

(MHz)
Maximum Average Maximum Average

1 902.6 0 0 -102 -107

2 902.7 0 0 -113 -115
3 902.8 52 52 -82 -88
4 902.9 52 52 -82 -88
39 906.2 75 60 -91 -95
40 906.3 75 60 -91 -95
41 906.4 0 0 -111 -113
42 906.5 0 0 -111 -113
60 908.3 100 100 -72 -73
61 908.4 100 100 -72 -73
62 908.5 100 100 -88 -90
63 908.6 100 100 -88 -90

39 | P a g e
5.2.1. Feature Description and Selection

All the features stated in Table 5.1 were not used to model the spectrum occupancy because they
do not capture unique information and may negatively affect the generalization of the prediction
model. To address this, a feature reduction technique was employed to eliminate redundancies,
selecting only those features essential for describing the spectrum occupancy model. Feature
selection aims to reduce the number of input variables, preventing an undue increase in model
complexity and preserving its generalization capability. In this study, filter-based feature selection
was implemented for feature reduction. Filter methods rely on statistical measures, such as
information gain, to identify features that contribute the most information about the target variable.
Notably, filter methods exclusively consider the association between each feature and the class
label [34]. Following the feature reduction process, the selected spectrum occupancy state
prediction features comprised frequency, average occupancy, and average field strength. These
features were chosen based on their relevance and significance in the context of spectrum
occupancy state modelling. The refined set of features is detailed in Table 5.2, providing a clear
understanding of the variables considered in the predictive modelling of spectrum occupancy.

Table 5.2 The final features selected used to model the spectrum occupancy state prediction

Frequency Average Occupancy (%) Average Field Strength (dB)

902.6 0 -107
902.7 0 -115
902.8 52 -88
902.9 52 -88
906.2 60 -95
906.3 60 -95
906.4 0 -113
906.5 0 -113
908.3 100 -73
908.4 100 -73
908.5 100 -90
908.6 100 -90

These methods were applied to ascertain the average spectrum utilization within the designated
spectrum band.

40 | P a g e
The methods used to characterize the spectrum occupancy model are the average field strength (or
power spectral density) and average occupancy (or duty cycle). Average field strength indicates
the intensity of the signal, providing insight into the strength of signals present within the analysed
spectrum. On the other hand, average occupancy measures how often the signal is present or the
frequency with which signals are detected, reflecting the temporal presence of signals within the
specified spectrum band. In summary, the results presented in the table offer a comprehensive
understanding of the spectrum occupancy state through signal strength (or average field strength)
and temporal presence (or average occupancy) using the ED and CFD spectrums sensing methods
respectively.

5.2.2. Data Analysis and Feature Engineering

A spectrum measurement campaign conducted in the Ethio-telecom GSM900 MHz 2G mobile

network uplink band at four different regional cities in October 2021 under the project “Digital
Transformation for Ethiopia” was used as supplementary data in this study shows the
underutilization of the spectrum has an average utilization of 21.45% in Adama, 16.21% in Bahir
Dar, 18.87% in Hawassa, and 33.52% in Jijiga.

100
90
80
70
Utilization(%)

60
50
40 33.52
30 21.45 18.87
16.21
20
10
0
Adama Bahir Dar Hawassa Jijiga

City

Figure 5.1 GSM900 MHz uplink band average spectrum utilization in four regional cities

The same spectrum measurement campaign was also conducted in Addis Abeba within four
different mobile network uplink bands owned by Ethio-telecom.

41 | P a g e
The results showed the following average spectrum utilization values in the four spectrum bands:
14.72% in GSM 900MHz, 31.67% in UMTS 900MHz, 25.32% in LTE 1800MHz, and 6.75% in
LTE 2600MHz, which was also underutilized.

100
90
80
Utilization (%)

70
60
50
40 31.67
30 25.32
20 14.72
10 6.75

0
GSM900 UMTS900 LTE1800 LTE-A

Spectrum Bands

Figure 5.2 Addis Abeba City average spectrum utilization in four spectrum uplink bands

The result of the spectrum utilization analysis in the cities mentioned above indicates that the
spectrum resource is underutilized. Even though all of the spectrum bands are now underutilized,
subscribers would migrate from 2G to 3G, 4G, and 5G to get the latest technology features and
services through the load-handling capacity of these new-generation mobile networks. The 900
MHz GSM 2G is the first mobile network in Ethiopia that offers only voice and short message
services (SMS) [40]. Therefore, the 900 MHz 2G mobile network uplink band will become more
underutilized because of the subscribers' migration, making it a potential candidate for
implementing a CR.

The CFD-based spectrum sensing method was employed in this study to characterize primary user
(PU) channels and model spectrum occupancy prediction due to its superior performance in
challenging SNR conditions. Over five days, spectrum measurements using CFD-based spectrum
sensing achieved an average utilization rate of 20.47%.

42 | P a g e
The CFD-based spectrum sensing method was used in this study for defining PU channel
characterization and modeling spectrum occupancy prediction due to its enhanced performance in
challenging SNR conditions. In a five-day spectrum measurement campaign conducted using the
CFD spectrum sensing method, the average spectrum utilization for the GSM 900 MHz uplink
band was 20.47%.

The spectrum utilization analysis conducted in five days revealed distinct values for weekdays
(Thursday, Friday, and Monday) and weekends (Saturday and Sunday). Specifically, the average
spectrum utilization on weekdays was 19.13%, 19.03%, and 21.14%, respectively. In contrast, the
average spectrum utilization on weekends exhibits a variation, with values of 17.3% on Saturday
and a higher utilization rate of 25.76% on Sunday. This observed pattern suggests that spectrum
usage tends to be lower on Saturdays and higher on Sundays. This difference can be attributed as
the fact that Sunday is a holiday in Ethiopia.

100
90
80
70
Utilization(%)

60
50
40
30 25.76
19.13 19.03 21.14
17.3
20
10
0
Thursday Friday Saturday Sunday Monday

Day

Figure 5.3 GSM900 spectrum utilization percentage in five different days

An in-depth analysis was performed on spectrum utilization by dividing the day into four equal
six-hour periods to get detailed insights into the variation of spectrum usage at different times.

43 | P a g e
The result indicates a utilization rate of 8.27% during the night (00:00:00 - 06:00:00), 23.69% in
the morning (06:00:00 - 12:00:00), 28.67% during the day (12:00:00 - 18:00:00), and 21.06% in
the evening (18:00:00 - 00:00:00). The utilization has a visible difference over time, with the
lowest utilization at night and the highest during the day.

100
90
80
70
Utilization (%)

60
50
40
30 28.67
23.69
21.06
20
8.27
10
0
Night Morning Day Evening

Seasons of a day

Figure 5.4 GSM900 spectrum utilization percentage in four different seasons of a day

The spectrum utilization can be further effectively represented through a spectrogram, as depicted
in Figure 5.7. This visual depiction shows the frequency spectrum usage pattern over time. A
spectrogram illustrates the relationship between duty cycle and frequency, with line brightness
indicating the duty cycle for each corresponding frequency in the frame [41]. The presented figures
in the spectrogram cover a comprehensive dataset collected over five days, considering the
frequency range of 902.5-915 MHz. The x-axis denotes time, while the y-axis represents frequency
in the spectrum occupancy plots. The depicted information reveals the spectrum utilization over
the five days, indicating that the entire spectrum band experiences underutilization.

44 | P a g e
 Figure 5.5 Spectrogram of the 902.5-915 MHz for the CFD based spectrum sensing

5.2.3. Preparing The Dataset

The original dataset had too many unnecessary features for the spectrum occupancy state
prediction. Since then, irrelevant features that make the model complex and not necessary for the
spectrum occupancy state prediction have been removed. The proposed model now utilizes a
dataset generated from the CFD-based spectrum sensing, chosen due to its reliability-enhanced
performance in challenging SNR environment conditions.

45 | P a g e
 …

Figure 5.6 The dataset used for the spectrum occupancy state prediction model

5.2.4. Creating the Model

The Encoder-Decoder LSTM, a specialized type of RNN model, is designed to address sequence-
to-sequence problems in spectrum occupancy state prediction. This model handles multiple time
steps as both input and output data, making it suitable for many-to-many sequence prediction
problems [41]. Consequently, the Encoder-Decoder LSTM model has proven to be highly effective
for sequence-to-sequence prediction tasks, enabling multiple steps-ahead forecasting in spectrum
occupancy state prediction.

Figure 5.7 Learning and predicting the long-term spectrum occupancy in the proposed LSTM
network with sequence-to-sequence architecture

46 | P a g e
5.3. Implementation Details

5.3.1. Working Environment and Tools

The following tools were used for design and implementation in this study.

Table 5.3 The hardware and software specifications used to conduct the experiments

Specification of Machine Used for Experiments

Manufacturer HP
System Model HP ZBook Fury 17 G7 Mobile Workstation
Processor Intel(R) Core (TM) i7-10750H CPU @ 2.60GHz,
GPU NVIDIA Quadro T1000 with Max-Q Design
SSD Configured SSD is 1TB
Memory 16GB DDR4
Operating System Microsoft Windows 11 Pro (64- bit)
Software MATLAB R2021a and Python 3.10.9

5.3.2. Dataset

This study utilizes the data acquired through the CFD-based spectrum sensing method to define
the characteristics of PU channels and develop a predictive model for the spectrum occupancy
state. The dataset encompasses 225,000 data points, representing half of the five-day measurement
data. Subsequently, this dataset is divided into training and validation sets, maintaining an 80% to
20% ratio. Specifically, 180,000 data points are earmarked for training purposes, while the
remaining 45,000 are allocated for validating the predictive model. The selection of this split is
driven by the available data for this study, aiming to provide a substantial number of training
samples and validation data to ensure the effective generalization and robust evaluation of the
trained model.

5.3.3. Experiment Setup with

The spectrum occupancy state prediction model was implemented using Python programming
language and the Keras library, specifically using a DL algorithm. Training and validation
experiments were conducted to ensure optimal model performance. These experiments involved a
comprehensive exploration of hyperparameter values, specifically focusing on the learning rate,
the number of hidden units, and the dropout rate.
47 | P a g e
The model designed for learning the long-term time-series data dependencies in spectrum
occupancy state prediction was structured as a sequential model comprising seven layers. The
model architecture comprises three LSTM layers with 128 units, followed by two dropout layers
with a dropout rate of 0.1. Additionally, two dense layers with 128 and 64 units, were incorporated,
along with a final output layer. The model was configured with an activation function of rectified
linear unit (ReLU) for all hidden layers and sigmoid for the output layer, adaptive moment
estimation (ADAM) as an optimizer, binary cross-entropy as a loss function, 0.001 learning rate,
128 batch size, and 400 Epochs. The model configuration was systematically evaluated based on
key metrics like loss function and accuracy. Various combinations of hyperparameters were
vigorously tested to minimize the loss percentage while enhancing the accuracy to identify the
most effective model. Hyperparameter values were determined depending on the results of the
iterative experiments. Therefore, the selection of the hyperparameters was guided by the dual
goals of minimizing the loss function percentage and improving the accuracy simultaneously. The
final spectrum occupancy state prediction model was chosen based on its consistent demonstration
of the lowest loss and highest accuracy across the iterative experimentation phases. During the
training phase, a dynamic scheduler plays a pivotal role in enhancing the model's performance.
The scheduler continuously monitors the model's losses and dynamically adjusts hyperparameters
to optimize prediction accuracy. The adaptive nature of the scheduler ensures that the learning rate
is tuned automatically, reflecting a proactive approach to improving the model's accuracy.

Three distinct deep learning models, namely LSTM, Bi-LSTM, and ConvLSTM, were employed
to compare the performance of spectrum occupancy state prediction, and three sets of experiments
were conducted. The first set of experiments focused on short-term prediction, where the objective
was to predict spectrum occupancy state one step ahead. In the second set of experiments, the focus
shifted to long-term prediction. The models predicted the spectrum occupancy state for three and
five hours ahead. This required the systems to forecast the spectrum occupancy patterns over
extended time horizons, providing insights into the system's ability to capture and generalize
temporal dependencies over a more extended period. These experiments were aimed to evaluate
and compare the effectiveness of the models in different prediction scenarios.

48 | P a g e
 5.4. Results of the Study

The models were evaluated using the performance evaluation metrics mentioned in 3.6 and 3.7.
The results obtained in each model have been presented in the below sections from 5.4.1 to 5.4.3.

5.4.1. Spectrum Occupancy Prediction Using the LSTM Model

Figure 5.8 Training and validation, loss and accuracy of the LSTM model short-term prediction

Table 5.4 Performance results of the LSTM model for short-term prediction

Metric Value
Accuracy 0.99446
Precision 0.97619
Recall 1
F1-Score 0.987952
Specificity 0.992832
MSE 0.00554017
RMSE 0.0744323
MAE 0.00554017
MAPE 0.554017

49 | P a g e
 Figure 5.9 Training and validation, loss and accuracy of the LSTM model for 3-hours ahead
prediction

Table 5.5 Performance results of the LSTM model for 3-hours ahead prediction

Metric Value
Accuracy 0.99446
Precision 0.97619
Recall 1
F1-Score 0.987952
Specificity 0.992832
MSE 0.00554017
RMSE 0.0744323
MAE 0.00554017
MAPE 0.554017

50 | P a g e
 Figure 5.10 Training and validation, loss and accuracy of the LSTM model for 5-hours ahead
prediction

Table 5.6 Performance results of the LSTM model for 5-hours ahead prediction

Metric Value
Accuracy 0.99446
Precision 0.97619
Recall 1
F1-Score 0.987952
Specificity 0.992832
MSE 0.00554017
RMSE 0.0744323
MAE 0.00554017
MAPE 0.554017

The performance results presented in Tables 5.4 to 5.6 demonstrate that the spectrum occupancy
state predictions obtained from the LSTM model consistently exhibit equal results across all
evaluated metrics.

51 | P a g e
 Figure 5.11 Spectrum occupancy state prediction in LSTM for the 904.1MHZ channel

Figure 5.11 compares the spectrum occupancy state actual data, training, and testing prediction
performed for the 904.1MHZ channel. The LSTM network is trained to adapt new spectrum
occupancy states, as shown in Figure 5.11, using the five-day spectrum measurement data for a
one-channel 904.1MHz. In the graph, the brown dotted line representing the training data and the
green dotted line representing the testing data closely match the actual observations depicted by
the blue solid line. The test performance indicates an accuracy of 96%.

52 | P a g e
5.4.2. Spectrum Occupancy Prediction Using the Bi-LSTM Model

Figure 5.12 Training and validation, loss and accuracy of the Bi-LSTM model short-term
prediction

Table 5.7 Performance results of the Bi-LSTM model for short-term prediction

Metric Value
Accuracy 0.99446
Precision 0.97619
Recall 1
F1-Score 0.987952
Specificity 0.992832
MSE 0.00554017
RMSE 0.0744323
MAE 0.00554017
MAPE 0.554017

53 | P a g e
Figure 5.13 Training and validation, loss and accuracy of the Bi-LSTM model for 3-hours ahead
prediction

Table 5.8 Performance results of the Bi-LSTM model for 3-hours ahead prediction

Metric Value
Accuracy 0.99446
Precision 0.97619
Recall 1
F1-Score 0.987952
Specificity 0.992832
MSE 0.00554017
RMSE 0.0744323
MAE 0.00554017
MAPE 0.554017

54 | P a g e
Figure 5.14 Training and validation, loss and accuracy of the Bi-LSTM model for 5-hours ahead
prediction

Table 5.9 Performance results of the Bi-LSTM model for 5-hours ahead prediction

Metric Value
Accuracy 0.99446
Precision 0.97619
Recall 1
F1-Score 0.987952
Specificity 0.992832
MSE 0.00554017
RMSE 0.0744323
MAE 0.00554017
MAPE 0.554017

The performance results presented in Tables 5.7 to 5.9 demonstrate that the spectrum occupancy
state predictions obtained from the Bi-LSTM model consistently exhibit equal results across all
evaluated metrics.

55 | P a g e
5.4.3. Spectrum Occupancy Prediction Using the ConvLSTM Model

Figure 5.15 Training and validation, loss and accuracy of the ConvLSTM model for short-term
prediction

Table 5.10 Performance results of the ConvLSTM model for short-term prediction

Metric Value
Accuracy 0.99723
Precision 0.987952
Recall 1
F1-Score 0.993939
Specificity 0.996416
MSE 0.00277008
RMSE 0.0526316
MAE 0.00277008
MAPE 0.277008

56 | P a g e
Figure 5.16 Training and validation, loss and accuracy of the ConvLSTM model 3-hours ahead
prediction

Table 5.11 Performance results of the ConvLSTM model for 3-hours ahead prediction

Metric Value
Accuracy 0.99723
Precision 0.987952
Recall 1
F1-Score 0.993939
Specificity 0.996416
MSE 0.00277008
RMSE 0.0526316
MAE 0.00277008
MAPE 0.277008

57 | P a g e
 Figure 5.17 Training and validation, loss and accuracy of the ConvLSTM model for 5-hours
ahead prediction

Table 5.12 Performance results of the ConvLSTM model for 5-hours ahead prediction

Metric Value
Accuracy 0.99723
Precision 0.987952
Recall 1
F1-Score 0.993939
Specificity 0.996416
MSE 0.00277008
RMSE 0.0526316
MAE 0.00277008
MAPE 0.277008

The performance results presented in Tables 5.10 to 5.12 demonstrate that the spectrum occupancy
state predictions obtained from the ConvLSTM model consistently exhibit equal results across all
evaluated metrics.
58 | P a g e
 5.5. Discussion
5.5.1. Quantifying Improvements

The long-term spectrum occupancy state prediction model used to implement a CRN in this study
has improved spectrum utilization and reduced sensing energy.

1. Improvement in Spectrum Utilization

The long-term spectrum occupancy prediction model improves spectrum utilization by allowing
SUs to select and use idle PU channels. This allows the SUs to choose the appropriate channels
efficiently to reduce channel-switching costs and increase network throughput. In a CRN the PUs
have two states, but the SUs can sense only one channel at a time. The CRN has two types of SUs
the CRsense and the CRpredict. The CRsense randomly selects a channel at every slot and senses
the state of the channel, while the CRpredict senses only channels with an idle prediction state.
According to [4], spectrum utilization (SU) in the CRN can be defined as the ratio of the number
of idle slots correctly identified and utilized by the SUs to the total number of idle slots available
in the CRN.

Number of idle slots sensed

𝑆𝑈 = Total number of idle slots in the band 5.1

The improvement in spectrum utilization due to spectrum prediction can be expressed as

𝑆𝑈𝑠𝑒𝑛𝑠𝑒 − 𝑆𝑈𝑝𝑟𝑒𝑑𝑖𝑐𝑡
𝑆𝑈𝑖𝑚𝑝(%) = 5.2
𝑆𝑈𝑠𝑒𝑛𝑠𝑒

Where 𝑆𝑈𝑠𝑒𝑛𝑠𝑒 and 𝑆𝑈𝑝𝑟𝑒𝑑𝑖𝑐𝑡 represent the spectrum utilization for the CRsense and CRpredict,
respectively. Substituting (5.1) in (5. 2), 𝑆𝑈𝑖𝑚𝑝(%) can be given by

𝐼𝑠𝑒𝑛𝑠𝑒 − 𝐼𝑝𝑟𝑒𝑑𝑖𝑐𝑡
𝑆𝑈𝑖𝑚𝑝(%) = 5.3
𝐼𝑠𝑒𝑛𝑠𝑒

Where 𝐼𝑠𝑒𝑛𝑠𝑒 and 𝐼𝑝𝑟𝑒𝑑𝑖𝑐𝑡 represent the number of idle channels sensed by the CRsense and the
number of idle channels predicted by the CRpredict respectively.

59 | P a g e
This analysis can be translated into a machine learning model and becomes equal with the
specificity that measures the true negative rate, which is the fraction of negative values that were
correctly predicted and can be calculated as expressed in equation 5.4.

𝑇𝑁
𝑆𝑈𝑖𝑚𝑝(%) = 𝑆𝑝𝑒𝑐𝑖𝑓𝑐𝑖𝑡𝑦 = 𝑇𝑁+𝐹𝑃 5.4

The 𝑆𝑈𝑖𝑚𝑝(%) improves the network throughput which shows the data rate achieved in the network
due to the availability of more channels and can be calculated as expressed in equation 5.5.

Throughput = 𝑆𝑈𝑖𝑚𝑝(%) * the number of channels in the spectrum band 5.5

2. Reduction in Sensing Energy

The prediction model reduces the sensing energy required by the SUs, because the SUs senses
only the idle channels. The CRsense senses all the channels whereas the CRpredict only senses the
channel that is predicted idle. In other words, when the channel state is predicted to be busy, the
sensing operation is not performed to reduce energy. Assuming that one unit of sensing energy is
required to sense one slot [4], then the total sensing energy required for a CRsense in a finite
duration of time can be calculated as expressed in equation 5.6.

SEsense=(Toatal number of slots in the duration)∗(unit sensing enegry) 5.6

while the total sensing energy required by the CRpredict can be given by

SEpredict = (SEsense − (Bpreidct )) ∗ (Unit sensing energy) 5.7

Where 𝐵𝑝𝑟𝑒𝑖𝑑𝑐𝑡 is the total number of busy slots predicted by the CRpredict.

Therefore, using (5.6) and (5.7), the percentage reduction in the sensing energy can be given by

𝑆𝐸𝑠𝑒𝑛𝑠𝑒 −𝑆𝐸𝑝𝑟𝑒𝑑𝑖𝑐𝑡 𝐵𝑝𝑟𝑒𝑑𝑖𝑐𝑡

𝑆𝐸𝑟𝑒𝑑(%) = = 𝑇𝑜𝑡𝑎𝑙 𝑛𝑜.𝑜𝑓 5.8
𝑆𝐸𝑠𝑒𝑛𝑠𝑒 𝑖𝑑𝑙𝑒 𝑠𝑙𝑜𝑡𝑠

This can be translated to a machine learning model that measures a value by dividing the true
positive value by the true negative plus the false positive values, even it doesn’t have an equivalent
machine learning metric it can be calculated and expressed as shown in equation 5.9.
60 | P a g e
 𝑇𝑃
𝑆𝐸𝑟𝑒𝑑(%) = 𝑇𝑁+𝐹𝑃 5.9

The improvements achieved in spectrum utilization, the network throughput obtained, and the
reduction of sensing energy on three models for all predictions results calculated based on
equations 5.4, 5.5, and 5.9 are presented in the above Table 5.13.

Table 5.13 Quantified improvements achieved in the spectrum occupancy state prediction

Model Length of Prediction 𝑺𝑼𝒊𝒎𝒑(%) Throughput 𝑺𝑬𝒓𝒆𝒅(%)

LSTM Short-Term 99.28 124.1 29.39
Long-Term (3 hrs.) 99.28 124.1*45 29.39
Long-Term (5 hrs.) 99.28 124.1*75 29.39
Bi- LSTM Short-Term 99.28 124.1 29.39
Long-Term (3 hrs.) 99.28 124.1*45 29.39
Long-Term (5 hrs.) 99.28 124.1*75 29.39
Conv-LSTM Short-Term 99.64 124.55 29.39
Long-Term (3 hrs.) 99.64 124.55*45 29.39
Long-Term (5 hrs.) 99.64 124.55*75 29.39

5.5.2. Comparison of Results

Across all three models, each term of prediction has equal performance results. However, the short-
term prediction achieved its targeted accuracy earlier than the long-term 3-hour and 5-hour
predictions. The 3-hour long-term prediction achieved its targeted accuracy earlier than the 5-hour
long-term prediction. The Bi-LSTM model differs from the others by achieving its targeted
accuracy earlier across all prediction scenarios. This may be attributed to its ability to process data
in both forward and backward directions, potentially improving its ability to learn temporal
relationships. Consequently, short-term predictions tend to achieve better accuracy earlier than
long-term predictions. This tendency may arise from factors such as predicting the immediate
future requires fewer data points and simpler relationships, and the available data for training the
model can be enough for short-term prediction than for long-term prediction.

In the comparative analysis performed on the spectrum occupancy prediction conducted on LSTM,
Bi-LSTM, and ConvLSTM models, performance evaluation metrics outlined in sections 3.6 and
3.7 were employed.

61 | P a g e
The comparative analysis is visually represented in Figures 6.20 to 6.21. Notably, both LSTM and
Bi-LSTM exhibited identical results achieving 99.45% accuracy, 97.62% precision, 98.8% F1-
Score, MSE 0.554017%, RMSE 7.44323%, MAE 0.554017%, and MAPE of 55.4017%. However,
the ConvLSTM model outperformed that achieved 99.72% accuracy, 98.8% precision, and 99.39
F1-Score, MSE 0.277008%, RMSE 5.26316%, MAE 0.277008%, and a MAPE of 27.7008%.

100
99
98
97
96
95
Short- Long-Term Long-Term Short- Long-Term Long-Term Short- Long-Term Long-Term
Term (3hrs) (5hrs) Term (3hrs) (5hrs) Term (3hrs) (5hrs)
LSTM Bi-LSTM ConvLSTM
Accuracy 99.45 99.45 99.45 99.45 99.45 99.45 99.72 99.72 99.72
Precision 97.62 97.62 97.62 97.62 97.62 97.62 98.8 98.8 98.8
F1-Score 98.8 98.8 98.8 98.8 98.8 98.8 99.39 99.39 99.39

Accuracy Precision F1-Score

Figure 5.18 Comparison of models in Accuracy, Precision, and F1-Score

Figure 5.19 Comparison of models in MSE, RMSE, MAE, and MAPE

62 | P a g e
 Chapter Six

6. Conclusion and Future Work

6.1. Conclusion
The primary challenge in spectrum utilization arises from the rigidity of the FSA policy, which
makes the underutilization of the available scarce spectrum resource. This study addresses the
challenges posed by the rigidity of FSA policies and paves the way for more effective and efficient
spectrum utilization that optimizes scarce spectrum resources. This study has improved spectrum
utilization by predicting spectrum occupancy state using a long-term adaptive LSTM model and a
historical dataset obtained through a reliable spectrum sensing method. Going beyond
conventional prediction approaches, this study identified potential candidate spectrum bands
through a comprehensive spectrum utilization and techno-economic analysis. The temporal data is
input into the LSTM model through an input gate, enabling the network to capture and learn
patterns over extended periods. The model undergoes extensive training steps comprising 400
epochs, each having 128 batches. The data passes over multiple layers of the LSTM model, each
designed to extract and process relevant features for optimal prediction outcomes.

The comparative analysis performed on LSTM, Bi-LSTM, and ConvLSTM models reveals the
proposed model has remarkable performances for spectrum occupancy state prediction. Although
the proposed LSTM and the Bi-LSTM model exhibit strong performance, the ConvLSTM model
stands out with a superior accuracy of 99.72%, surpassing the 99.45% accuracy achieved by the
LSTM and Bi-LSTM models. The proposed model demonstrates notable accuracy evidenced by
the low error values of 0.554017% MSE, 7.44323% RMSE, 0.554017% MAE, and 55.4017%
MAPE. Employing five days' worth of historical spectrum measurement data, the model
successfully predicted the spectrum occupancy state for the subsequent five hours with an accuracy
of 99.45%. The low error values, indicating close alignment between the model's predictions and
actual spectrum occupancy state, underscore the model's accuracy in predicting future spectrum
occupancy states. Therefore, the proposed LSTM model and its counterparts are considered robust
models for spectrum occupancy state prediction.

63 | P a g e
This study presented significant progress in enhancing the reliability of spectrum sensing and the
accuracy of spectrum occupancy state prediction. Employing DL models for spectrum occupancy
state prediction for the CRN has improved spectrum utilization boosts from a mere 20.47% to an
impressive 98.28% and reduced the sensing energy by 29.39% compared to real-time sensing. This
significant enhancement underscores the inefficiency of the traditional FSA policy that
inadvertently leads to now artificial spectrum scarcity. Integrating DL models with reliable
spectrum sensing methods and identifying potential candidate spectrum bands presents a
promising avenue for augmenting the efficiency of deploying CRs. This synergy addresses the
limitations of conventional approaches and unlocks new possibilities for optimizing spectrum
utilization and overcoming challenges associated with spectrum scarcity. The study's findings
suggest that embracing advanced technologies such as DL in the context of CR can pave the way
for more adaptive and spectrum resource-efficient wireless communication technologies.

6.2. Future Work

Spectrum occupancy state prediction for deploying a CR becomes a foundation for application
development like smart city. The applications of a smart city can be traffic management, efficient
parking solutions, public safety, waste management, remote monitoring, environmental well-
being, and improvements in public transport systems. Future studies can focus on enhancing the
predictability of further occupancy lengths up to days for integrating CR with the Internet of
Things, which creates a synergistic system known as the CR Internet of Things (CRIoT). This
integrated IoT and CR amplifies smart cities’ capability, providing a comprehensive and
interconnected infrastructure for effective and efficient urban management.

Introducing a new technology, such as a smart city, into a specific area necessitates a
comprehensive techno-economic analysis to assess the feasibility based on technical, economic,
environmental, social, and legal criteria. The area of Bole, where spectrum measurement data
collection has been conducted, emerges as a feasible location for implementing a smart city project.
This feasibility is attributed to the infrastructural development in the Bole area, which enhances
the practicality of conducting a techno-economic analysis compared to other parts of the city.

64 | P a g e
 Reference
[1] J. Kalliovaara, “Field Measurements in Determining Incumbent Spectrum Utilization and
Protection Criteria in Wireless Co-existence Studies,” Turku University of Applied
Sciences, 2017. [Online]. Available: DOI:10.13140/RG.2.2.33351.80805
[2] F. Weidling, D. Datla, V. Petty, P. Krishnan, and G. J. Minden, “A Framework for R.F.
Spectrum Measurements and Analysis,” in First IEEE International Symposium on New
Frontiers in Dynamic Spectrum Access Networks, 2005. DySPAN 2005., 2005. [Online].
Available: DOI: 10.1109/DYSPAN.2005.1542672
[3] T.Jeacock, “A Standard Approach for Assessing the Spectrum Management Needs of
Developing Countries,” Report on Telecommunication Development Sectror: ITU, 2016.
[Online]. Available: www.itu.int
[4] V. K. Tumuluru, P. Wang, and D. Niyato, “A Neural Network Based Spectrum Prediction
Scheme for Cognitive Radio,” in IEEE International Conference on Communications,
2010. doi: 10.1109/ICC.2010.5502348.
[5] L.J.H.Viveros, D.A.L.Sarmiento, and N.E.V.Parra, “Modeling and Prediction Primary
Nodes in Wireless Networks of Cognitive Radio Using Recurrent Neural Networks,”
Contemporary Engineering Sciences, 2018, doi: 10.12988/ces.2018.84164.
[6] N.G.Bara’u, W.Feng, and M. Almustapha, “Spectrum Hole Prediction Based on Historical
Data: A Neural Network Approach,” Computer Science, Engineering, 2014, [Online].
Available: Corpus ID: 6991758
[7] A. Shenfield, Z. Khan, and H. Ahmadi, “Deep Learning Meets Cognitive Radio: Predicting
Future Steps,” in 2020 IEEE 91st Vehicular Technology Conference (VTC2020-Spring),
2020. [Online]. Available: http://shura.shu.ac.uk/25962/
[8] B. S. Shawel, D. H.Woledegebre, and S.Pollin, “Deep-learning based Cooperative
Spectrum Prediction for Cognitive Networks,” in 2018 International Conference on
Information and Communication Technology Convergence (ICTC), 2018.
[9] A. Saad, B. Staehle, and R. Knorr, “Spectrum Prediction using Hidden Markov Models for
Industrial Cognitive Radio,” in 2016 IEEE 12th International Conference on Wireless and
Mobile Computing, Networking and Communications (WiMob), 2016. [Online]. Available:
DOI: 10.1109/WiMOB.2016.7763231

65 | P a g e
[10] T. Wysocki and B. J. Wysocki, “Spectrum Occupancy Prediction Using a Hidden Markov
Model,” in 2016 IEEE 12th International Conference on Wireless and Mobile Computing,
Networking and Communications (WiMob), 2016. Accessed: Mar. 12, 2024. [Online].
Available: DOI: 10.1109/ICSPCS.2015.7391772
[11] N.Balwani, D.K.Patel, B.Soni, and M.L.Benítez, “Long Short-Term Memory based
Spectrum Sensing Scheme for Cognitive Radio,” in 2019 IEEE 30th Annual International
Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 2019.
[Online]. Available: DOI: 10.1109/PIMRC.2019.8904422
[12] M. Sardana and A. Vohra, “Analysis of Different Spectrum Sensing Techniques,” in 2017
International Conference on Computer, Communications and Electronics (Comptelix),
2017. [Online]. Available: DOI: 10.1109/COMPTELIX.2017.8004006
[13] P. S. Yawada and A. J. Wei, “Comparative Study of Spectrum Sensing Techniques Base on
Techniques Non-cooperative in Cognitive Radio Networks,” in 2016 5th International
Conference on Computer Science and Network Technology (ICCSNT), 2016. [Online].
Available: DOI: 10.1109/ICCSNT.2016.8070212
[14] M. Z. Alom, T. K. Godder, and M. N. Morshed, “A Survey of Spectrum Sensing Techniques
in Cognitive Radio Network,” in 2015 International Conference on Advances in Electrical
Engineering (ICAEE), 2015. [Online]. Available: DOI: 10.1109/ICAEE.2015.7506821
[15] V. Amrutha and K. V. Karthikeyan, “Spectrum Sensing Methodologies in Cognitive Radio
Networks: A Survey,” in 2017 International Conference on Innovations in Electrical,
Electronics, Instrumentation and Media Technology (ICEEIMT), 2017. [Online]. Available:
DOI: 10.1109/ICIEEIMT.2017.8116855
[16] M. A. Aygül, M. Nazzal, M. İ. Sağlam, D. B. da Costa, H. F. Ateş, and H. Arslan, “Efficient
spectrum Occupancy Prediction Exploiting Multidimensional Correlations through
Composite 2D-LSTM Models,” Sensors (Switzerland), 2020, doi: 10.3390/s21010135.
[17] N. Muchandi and R. Khanai, “Cognitive radio spectrum sensing: A survey,” in 2016
International Conference on Electrical, Electronics, and Optimization Techniques
(ICEEOT), 2016. [Online]. Available: DOI: 10.1109/ICEEOT.2016.7755301
[18] B. Soni, D. K. Patel, and M. L.Benitez, “Long Short-Term Memory based Spectrum Sensing
Scheme for Cognitive Radio using Primary Activity Statistics,” IEEE Access , 2020. doi:
10.1109/ACCESS.2017.
66 | P a g e
[19] A. Sharmila and P. Dananjayan, “Spectrum Sharing Techniques in Cognitive Radio
Networks-A Survey,” in 2019 IEEE International Conference on System, Computation,
Automation and Networking (ICSCAN), 2019. [Online]. Available: DOI:
10.1109/ICSCAN.2019.8878714
[20] G.Ding et al., “Spectrum Inference in Cognitive Radio Networks: Algorithms and
Applications,” IEEE Communications Surveys & Tutorials, 2017. [Online]. Available:
DOI: 10.1109/COMST.2017.2751058
[21] J. N. Javed, M. Khalil, and A. Shabbir, “A Survey on Cognitive Radio Spectrum Sensing:
Classifications and Performance Comparison,” in 2019 International Conference on
Innovative Computing (ICIC), 2019. [Online]. Available: DOI:
10.1109/ICIC48496.2019.8966677
[22] B. G. Najashi and F. Wenjiang, “Cooperative Spectrum Occupancy based Spectrum
Prediction Modeling,” Journal of Computational Information Systems, 2014, doi:
10.12733/jcis10167.
[23] D. Das, D. W. Matolak, and S. Das, “Spectrum Occupancy Prediction based on Functional
Link Artificial Neural Network (FLANN) in ISM Band,” Neural Comput Appl, 2018, doi:
10.1007/s00521-016-2653-5.
[24] R. Fan, H. Guo, L. Di, and X. Ling, “Spectrum Occupancy State Predictor based on
Recurrent Neural Network,” in Journal of Physics: Conference Series, Institute of Physics
Publishing, 2019. doi: 10.1088/1742-6596/1345/4/042020.
[25] W. Zhang, G. Huang, G. Wang, and Y. Wang, “Prediction High Frequency Parameters
based on Neural Network,” in IOP Conference Series: Materials Science and Engineering,
Institute of Physics Publishing, Nov. 2019. doi: 10.1088/1757-899X/631/5/052035.
[26] L. Xing, M. Li, Y. Wan, and Q. Wan, “Spectrum Prediction in Cognitive Radio Based on
Sequence- to- Sequence Neural Network,” in International Conference on Advanced
Hybrid Information Processing, Springer, 2019, pp. 343–354. doi: 10.1007/978-3-030-
36405-2_34.
[27] M.A.Aygül et al., “Spectrum Occupancy Prediction Exploiting Time and Frequency
Correlations Through 2D-LSTM,” in 2020 IEEE 91st Vehicular Technology Conference
(VTC2020-Spring), 2020. [Online]. Available: DOI: 10.1109/VTC2020-
Spring48590.2020.9129001
67 | P a g e
[28] L. Yu, Q. Wang, Y. Guo, and P. Li, “Spectrum Availability Prediction in Cognitive
Aerospace Communications: A Deep Learning Perspective,” in 2017 Cognitive
Communications for Aerospace Applications Workshop (CCAA), IEEE, 2017. [Online].
Available: DOI: 10.1109/CCAAW.2017.8001877
[29] L. Yu, J. Chen, and G. Ding, “Spectrum Prediction via Long Short- Term Memory,” in 2017
3rd IEEE International Conference on Computer and Communications (ICCC), 2017.
[Online]. Available: DOI: 10.1109/CompComm.2017.8322623
[30] B. S. Shawel, D. H. Woldegebreal, and S. Pollin, “Convolutional LSTM-based Long-Term
Spectrum Prediction for Dynamic Spectrum Access,” in 2019 27th European Signal
Processing Conference (EUSIPCO), 2019. [Online]. Available: DOI:
10.23919/EUSIPCO.2019.8902956
[31] S. S. Shirgan and U. L. Bombale, “Hybrid Neural Network Based Wideband Spectrum
Behavior Sensing Predictor for Cognitive Radio Application,” Sens Imaging, Dec. 2020,
doi: 10.1007/s11220-020-00293-4.
[32] L. Yin, S. Yin, W. Hong, and S. Li, “Spectrum Behavior Learning in Cognitive Radio based
on Artificial Neural Network,” in 2011 - MILCOM 2011 Military Communications
Conference, 2011. [Online]. Available: DOI: 10.1109/MILCOM.2011.6127671
[33] N.K. Chauhan and K.Singh, “A Review on Conventional Machine Learning vs Deep
Learning,” in 2018 International Conference on Computing, Power and Communication
Technologies (GUCON), 2018. [Online]. Available: DOI: 10.1109/GUCON.2018.8675097
[34] J. Cai, J. Luo, S. Wang, and S. Yang, “Feature Selection in Machine Learning: A New
Perspective,” Neurocomputing, 2018, doi: 10.1016/j.neucom.2017.11.077.
[35] C. Janiesch, P. Zschech, and K.Heinrich, “Machine Learning and Deep Learning,”
Electronic Markets, 2021, doi: 10.1007/s12525-021-00475-2/Published.
[36] D. Zhou, X. Zuo, and Z. Zhao, “Constructing a Large-Scale Urban Land Subsidence
Prediction Method Based on Neural Network Algorithm from the Perspective of Multiple
Factors,” Remote Sens (Basel), Apr. 2022, doi: 10.3390/rs14081803.
[37] R. Chandra, S. Goyal, and R. Gupta, “Evaluation of Deep Learning Models for Multi-Step
Ahead Time Series Prediction,” IEEE Access, 2021, doi: 10.1109/ACCESS.2021.3085085.
[38] Z. Wang and S. Salous, “Spectrum Occupancy Statistics and Time Series Models for
Cognitive Radio,” J Signal Process Syst, Feb. 2011, doi: 10.1007/s11265-009-0352-5.
68 | P a g e
[39] A. Agarwal, A. S. Sengar, and R. Gangopadhyay, “Spectrum Occupancy Prediction for
Realistic Traffic Scenarios: Time Series versus Learning-Based Models,” Journal of
Communications and Information Networks, vol. 3, no. 2, Jun. 2018, doi: 10.1007/s41650-
018-0013-6.
[40] B. B. H. D. Negash, “Techno-economic Analysis of LTE Deployment Scenarios for
Emerging City: A Case of Adama, Ethiopia Addis Ababa, Ethiopia,” in International
Conference on Information and Communication Technology for Development for Africa,
2018. [Online]. Available: DOI:10.1007/978-3-030-26630-1_18
[41] A. Ghosh, S. Kasera, and J. V.Merwe, “Spectrum Usage Analysis And Prediction using
Long Short-Term Memory Networks,” in Proceedings of the 24th International Conference
on Distributed Computing and Networking, 2023. [Online]. Available:
https://doi.org/10.1145/3571306.3571412

69 | P a g e
 Appendixes

Appendix A: Sample Code for Building, Creating, and Compiling the

LSTM Model
import numpy as np
import pandas as pd
import tensorflow as tf
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, mean_squared_error,
mean_absolute_error, confusion_matrix
import matplotlib.pyplot as plt
from tensorflow.keras.layers import LSTM, Dropout, Dense
from tabulate import tabulate
from tensorflow.keras.optimizers import Adam
colnames = ['Timestamp', '902.6', '902.7', '902.8', '902.9', '903.0', ...] # Include all column names
data = pd.read_csv('addisu_main.csv', names=colnames)
# Data Preprocessing
X = data.iloc[:, 1:-1].values # Exclude the 'Timestamp' column and target column
y = data.iloc[:, -1].values # Assuming the last column is the target
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
X = scaler.fit_transform(X)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
model = tf.keras.Sequential([
LSTM(128, input_shape=(X_train.shape[1], 1), activation='tanh', return_sequences=True),
Dropout(0.2),
LSTM(128, activation='tanh', return_sequences=True),
Dropout(0.2),
LSTM(128, activation='tanh'), # Additional LSTM layer
Dense(128, activation='relu'), # Additional dense layer
Dropout(0.2),
Dense(64, activation='relu'),
Dropout(0.2),
Dense(1, activation='sigmoid')
])
optimizer = Adam(learning_rate=0.001)
70 | P a g e
model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])
X_train = X_train.reshape(X_train.shape[0], X_train.shape[1], 1)
X_test = X_test.reshape(X_test.shape[0], X_test.shape[1], 1)
def lr_scheduler(epoch, lr):
if epoch % 10 == 0 and epoch > 0:
lr *= 0.85 # Reduce the learning rate by 15% every 10 epochs
return lr
lr_scheduler_callback = tf.keras.callbacks.LearningRateScheduler(lr_scheduler) # Use tf.keras.callbacks.
# Train the model with learning rate scheduler
history = model.fit(
X_train, y_train,
epochs=400,
batch_size=128,
validation_split=0.2,
verbose=2,
callbacks=[lr_scheduler_callback]
)
y_pred = model.predict(X_test)
y_pred = (y_pred > 0.5).astype(int)
accuracy = accuracy_score(y_test, y_pred)
precision = precision_score(y_test, y_pred)
recall = recall_score(y_test, y_pred)
f1 = f1_score(y_test, y_pred) # Add F1 Score calculation
mse = mean_squared_error(y_test, y_pred) # Calculate Mean Squared Error
mae = mean_absolute_error(y_test, y_pred) # Calculate Mean Absolute Error
conf_matrix = confusion_matrix(y_test, y_pred)
specificity = conf_matrix[0, 0] / (conf_matrix[0, 0] + conf_matrix[0, 1])
TN = conf_matrix[0, 0]
FP = conf_matrix[0, 1]
FN = conf_matrix[1, 0]
TP = conf_matrix[1, 1]
FPR = FP / (FP + TN)
FNR = FN / (FN + TP)
FRN = TP / (TN+ FP)
metrics_data = [
["Accuracy", accuracy],
["Precision", precision],

71 | P a g e
 ["Recall", recall],
["F1 Score", f1],
["Specificity", specificity],
["False Positive Rate (FPR)", FPR],
["False Negative Rate (FNR)", FNR],
["False Rate Negative (FRN)", FRN],
["Mean Squared Error (MSE)", mse],
["Root Mean Squared Error (RMSE)", np.sqrt(mse)],
["Mean Absolute Error (MAE)", mae],
["Mean Absolute Percentage Error (MAPE)", mae * 100] # MAPE is usually expressed as a percentage
]
table = tabulate(metrics_data, headers=["Metric", "Value"], tablefmt="grid")
print(table)
plt.figure(figsize=(12, 6))
plt.subplot(1, 2, 1)
plt.plot(history.history['loss'], label='Training Loss')
plt.plot(history.history['val_loss'], label='Validation Loss')
plt.legend()
plt.title('Loss Over Epochs')
plt.xlabel('Epochs')
plt.ylabel('Loss')
plt.subplot(1, 2, 2)
plt.plot(history.history['accuracy'], label='Training Accuracy')
plt.plot(history.history['val_accuracy'], label='Validation Accuracy')
plt.legend()
plt.title('Accuracy Over Epochs')
plt.xlabel('Epochs')
plt.ylabel('Accuracy')
plt.savefig('training_validation_accuracy.png')
plt.show()

72 | P a g e
 Appendix B: Sample Code for Building, Creating, and Compiling the Bi-
LSTM Model

import numpy as np

import pandas as pd

import tensorflow as tf

from sklearn.model_selection import train_test_split

from sklearn.metrics import accuracy_score, precision_score, recall_score, f1_score, mean_squared_error,

mean_absolute_error, confusion_matrix

import matplotlib.pyplot as plt

from tensorflow.keras.layers import Bidirectional, LSTM, Dropout, Dense

from tabulate import tabulate

from tensorflow.keras.optimizers import Adam

from sklearn.preprocessing import StandardScaler

colnames = ['Timestamp', '902.6', '902.7', '902.8', '902.9', '903.0', ...] # Include all column names

data = pd.read_csv('addisu_main.csv', names=colnames)

# Data Preprocessing

X = data.iloc[:, 1:-1].values # Exclude the 'Timestamp' column and target column

y = data.iloc[:, -1].values # Assuming the last column is the target

scaler = StandardScaler()

X = scaler.fit_transform(X)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

model = tf.keras.Sequential([

73 | P a g e
Bidirectional(LSTM(128, activation='tanh', return_sequences=True), input_shape=(X_train.shape[1], 1)),

Dropout(0.2),

Bidirectional(LSTM(128, activation='tanh', return_sequences=True)),

Dropout(0.2),

Bidirectional(LSTM(128, activation='tanh')),

Dense(128, activation='relu'),

Dropout(0.2),

Dense(64, activation='relu'),

Dropout(0.2),

Dense(1, activation='sigmoid')

])

optimizer = Adam(learning_rate=0.001)

model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])

X_train = X_train.reshape(X_train.shape[0], X_train.shape[1], 1)

X_test = X_test.reshape(X_test.shape[0], X_test.shape[1], 1)

def lr_scheduler(epoch, lr):

if epoch % 10 == 0 and epoch > 0:

lr *= 0.85 # Reduce the learning rate by 15% every 10 epochs

return lr

lr_scheduler_callback = tf.keras.callbacks.LearningRateScheduler(lr_scheduler) # Use tf.keras.callbacks.

history = model.fit(

X_train, y_train,

74 | P a g e
 epochs=400,

batch_size=128,

validation_split=0.2,

verbose=2,

callbacks=[lr_scheduler_callback]

y_pred = model.predict(X_test)

y_pred = (y_pred > 0.5).astype(int)

accuracy = accuracy_score(y_test, y_pred)

precision = precision_score(y_test, y_pred)

recall = recall_score(y_test, y_pred)

f1 = f1_score(y_test, y_pred) # Add F1 Score calculation

mse = mean_squared_error(y_test, y_pred) # Calculate Mean Squared Error

mae = mean_absolute_error(y_test, y_pred) # Calculate Mean Absolute Error

conf_matrix = confusion_matrix(y_test, y_pred)

specificity = conf_matrix[0, 0] / (conf_matrix[0, 0] + conf_matrix[0, 1])

TN = conf_matrix[0, 0]

FP = conf_matrix[0, 1]

FN = conf_matrix[1, 0]

TP = conf_matrix[1, 1]

FPR = FP / (FP + TN)

FNR = FN / (FN + TP)

75 | P a g e
FRN = TP / (TN+ FP)

metrics_data = [

["Accuracy", accuracy],

["Precision", precision],

["Recall", recall],

["F1 Score", f1],

["Specificity", specificity],

["False Positive Rate (FPR)", FPR],

["False Negative Rate (FNR)", FNR],

["False Rate Negative (FRN)", FRN],

["Mean Squared Error (MSE)", mse],

["Root Mean Squared Error (RMSE)", np.sqrt(mse)],

["Mean Absolute Error (MAE)", mae],

["Mean Absolute Percentage Error (MAPE)", mae * 100

table = tabulate(metrics_data, headers=["Metric", "Value"], tablefmt="grid")

print(table)

76 | P a g e
 Appendix C: Sample Code for Building, Creating, and Compiling the
ConvLSTM Model

from tabulate import tabulate

import matplotlib.pyplot as plt

colnames = ['Timestamp', '902.6', '902.7', '902.8', '902.9', '903.0', ...] # Include all column names

data = pd.read_csv('addisu_main.csv', names=colnames)

X = data.iloc[:, 1:-1].values # Exclude the 'Timestamp' column and target column

y = data.iloc[:, -1].values # Assuming the last column is the target

scaler = StandardScaler()

X = scaler.fit_transform(X)

X = X.reshape(X.shape[0], X.shape[1], 1)

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

model = tf.keras.Sequential([

Conv1D(filters=64, kernel_size=3, activation='relu', input_shape=(X_train.shape[1], 1)),

LSTM(128, activation='tanh', return_sequences=True),

Dropout(0.1),

LSTM(128, activation='tanh', return_sequences=True),

Dropout(0.1),

LSTM(128, activation='tanh'),

Dense(128, activation='relu'),

Dropout(0.1),

Dense(64, activation='relu'),

77 | P a g e
 Dropout(0.1),

Dense(1) # Assuming a single-output regression task

])

optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)

model.compile(loss='binary_crossentropy', optimizer=optimizer , metrics=['accuracy'])

# Implement a learning rate scheduler

def lr_scheduler(epoch, lr):

if epoch % 10 == 0 and epoch > 0:

lr *= 0.85 # Reduce the learning rate by 15% every 10 epochs

return lr

lr_scheduler_callback = tf.keras.callbacks.LearningRateScheduler(lr_scheduler)

history = model.fit(

X_train, y_train,

epochs=400,

batch_size=128,

validation_split=0.2,

verbose=2,

callbacks=[lr_scheduler_callback]

model.save('occupancy_prediction_model.h5')

y_pred = model.predict(X_test)

rmse = np.sqrt(mean_squared_error(y_test, y_pred))

78 | P a g e
print(f"Root Mean Squared Error (RMSE) on Test Set: {rmse}")

# Predict occupancy on the test set

y_pred = model.predict(X_test)

y_pred = (y_pred > 0.5).astype(int)

accuracy = accuracy_score(y_test, y_pred)

precision = precision_score(y_test, y_pred)

recall = recall_score(y_test, y_pred)

f1 = f1_score(y_test, y_pred) # Add F1 Score calculation

mse = mean_squared_error(y_test, y_pred) # Calculate Mean Squared Error

mae = mean_absolute_error(y_test, y_pred) # Calculate Mean Absolute Error

conf_matrix = confusion_matrix(y_test, y_pred)

specificity = conf_matrix[0, 0] / (conf_matrix[0, 0] + conf_matrix[0, 1])

TN = conf_matrix[0, 0]

FP = conf_matrix[0, 1]

FN = conf_matrix[1, 0]

TP = conf_matrix[1, 1]

FPR = FP / (FP + TN)

FNR = FN / (FN + TP)

FRN = TP / (TN+ FP)

metrics_data = [

["Accuracy", accuracy],

["Precision", precision],

79 | P a g e
 ["Recall", recall],

["F1 Score", f1],

["Specificity", specificity],

["False Positive Rate (FPR)", FPR],

["False Negative Rate (FNR)", FNR],

["False Rate Negative (FRN)", FRN],

["Mean Squared Error (MSE)", mse],

["Root Mean Squared Error (RMSE)", np.sqrt(mse)],

["Mean Absolute Error (MAE)", mae],

["Mean Absolute Percentage Error (MAPE)", mae * 100] # MAPE is usually expressed as a percentage

table = tabulate(metrics_data, headers=["Metric", "Value"], tablefmt="grid")

80 | P a g e