ARID ZONE JOURNAL OF ENGINEERING, TECHNOLOGY & 

ENVIRONMENT 
AZOJETE December 2022. Vol. 18(4):575-586 

Published by the Faculty of Engineering, University of Maiduguri, Maiduguri, Nigeria.   

Print ISSN: 1596-2644, Electronic ISSN: 2545-5818 
www.azojete.com.ng  

 

Corresponding author’s e-mail address: edirin.omoze@uniben.edu       575 

ORIGINAL RESEARCH ARTICLE 
 

ARTIFICIAL NEURAL NETWORK (ANN) PATHLOSS PREDICTION MODEL 

FOR LTE NETWORK FOR MICROCELLS IN AN URBAN ENVIRONMENT 

 

E. L. Omoze* and J. O. Emagbetere 
Department of Electrical and Electronic Engineering, University of Benin 

 *Corresponding author’s email address: edirin.omoze@uniben.edu 

1.0 Introduction 

The demand trend for mobile communication services has shown that as population 

increases coupled with technological advancement, the demand for mobile communication 

services increases (Bi et al., 2001).  The mobile communication technology has advanced 

from first generation (1G) to (4G) in Nigeria. The brand name for the fourth generation of 

mobile wireless cellular communication standard is long term evolution (LTE). LTE, a 

successor to both 2G and 3G developed by the 3rd generation partnership project (3GPP) 

committee to improve Universal mobile telecommunications services (UMTS) 

communication standard is a highly flexible radio interface. The demand for 

telecommunication services necessitates the need for proper planning of cellular networks. 

The performance of cellular networks depends on the propagation model deployed during 

planning stage. Due to the complex and diverse nature of the radio propagation 

environment, the power of a propagating signal fluctuates with respect to time and space. 

Thus, a model that will work best is that which is developed using propagation data gotten 

from that environment. The use of empirical and deterministic models for deployment of 

mobile networks presents a compromise between accuracy and simplicity (Ojo et al., 2021). 

Example of empirical models in literature are the Okumura-Hata model, Walfisch Ikegami 

model, Stanford University interim model, Egli model and COST 231 Hata model (Zhang et 

al., 2019 and Jo et al., 2020). 

 

ARTICLE 

INFORMATION 
 

 

ABSTRACT 
 

One of the challenges faced by telecommunication service providers is 

capacity. The capacity of a cell depends on the model used in planning 

the cell during deployment stage. This paper presents Artificial Neural 

Network (ANN) Pathloss Prediction model for Long Term Evolution 

(LTE) network for microcells in Benin city, Edo state, Nigeria. Received 

signal strength data collected with respect to distance, mobile station 

antenna height and base station antenna height was used to create a data 

base with pathloss estimated from the measured data. A feature variable 

consisting of  80% of measurement data collected was used to train a 4-

32-1 multilevel perceptron artificial neural network with pathloss as the 

label using python language on Scientific python development 

environment interface (SPYDER).  The result of the 20% test data shows 

that artificial neural network is suitable for predicting pathloss in Benin 

City. This is as a result of root mean square error (RMSE) values and 

mean absolute error values of 0.59𝑑𝐵 and 0.42𝑑𝐵 obtained respectively 

which are less than the 6𝑑𝐵 error value allowed for network planning. It 

is recommended that the result of the study be adopted in designing or 

expanding LTE network capacity in the areas studied in this work. 
 
© 2022 Faculty of Engineering, University of Maiduguri, Nigeria. All rights reserved. 

 

Submitted 9 September, 2021 

Revised 28 October, 2021 

Accepted 5 November, 2021 

  
 

 

Keywords: 

Pathloss 

Artificial neural network 

(ANN) 

Received signal strength 

Root mean square error 

(RMSE) 

Long term evolution 

(LTE) 

http://www.azojete.com.ng/
file:///C:/Users/i3/Downloads/edirin.omoze@uniben.edu
mailto:edirin.omoze@uniben.edu


 

Arid Zone Journal of Engineering, Technology and Environment, December, 2022; Vol. 18(4):575-586. ISSN 1596-2644;  

e-ISSN 2545-5818; www.azojete.com.ng 

 

Corresponding author’s e-mail address: gedirin.omoze@uniben.edu     576 

 

Although the empirical models are simple to use because of the few parameters needed to 

predict the path loss in a given environment, they provide unsatisfactory results when 

applied to a more general radio propagation environment since they are mostly site specific 

(Zhang et al., 2019 and Jo et al., 2020). Deterministic models are more accurate but quite 

complex to develop and lack computational accuracy with an example being the ray tracing 

method. In order to develop models that can predict pathloss with high accuracy and 

computationally efficient, researchers in recent times have looked into machine learning 

based path loss prediction models. 

A lot of researchers have contributed knowledge in the area of developing pathloss models 

for cellular network planning. While some have develop models for different mobile 

network technologies, others have simulated cellular networks at different frequencies. 

Shoewu et al. (2019) presented the result on a study conducted in wetland and dry land 

areas in Lagos state, Nigeria using received signal strength data collected with TEMS 

investigation software installed on mobile phone via drive test for 4G cellular networks. 

Their result show that existing empirical model such as ECC-33 model, COST 231 Hata 

model, Free space model, Okumura model and Davidson-Hata model are not suitable for 

pathloss prediction in Lagos. Sulyman et al. (2014), Sree Vardhan et al. (2016), Ubaidillah et 

al. (2020) and Atanasoy et al. (2017) also carried out propagation studies on different 

locations. Sharma et al. (2018) presented a report on optimization of propagation pathloss 

model in 4G wireless communication system was presented. The authors in this study 

carried out data collection at 1800MHz in India and compared their result with predictions 

made by COST-231 Hata, Egli, Walfich-bertoni, ECC-33 and SUI models with pathloss 

values obtained from the measured field strength data. The authors reported that COST-

231 Hata, is best model based on the closest agreement to the measured path loss 

exponent. Thus, the COST 231 Hata model was optimized. Among the various studies 

reviewed, it is worthy of note that propagation models are environment specific and existing 

empirical models are mostly not suitable for use in planning cellular networks in 

environment where the terrain characteristics is not similar with that in which the model is 

built for.  

In wireless communication, pathloss prediction is treated as a supervised learning task with 

the received signal strength, distance between the mobile station and the base station, 

mobile station antenna height and base station antenna height as the feature variables and 

the path loss values as the label. Artificial Neural Network (ANN) is a one of the artificial 

intelligence algorithms that can be used in developing pathloss models based on data 

collected in a typical radio propagation environment and trained with it (Eichie et al., 2017). 

The flexibility of ANN to learn from a given data set without being explicitly programmed 

can be seen from researches in recent times. 

An artificial neural network (ANN) is a learning algorithm whose design is inspired by the 

behavior of neurons in the human brain. An ANN is based on a collection of connected 

units or nodes called artificial neurons. A neuron is the fundamental processing element of a 

neural network. Each neuron transmits its output via a connection of weights like the 

synapses in a biological neural network to another neuron in the next layer. Weights are 

assigned to a neuron based on its relative importance against other inputs. Shown in Figure 

1 is the structure of an artificial neuron.  

file:///C:/user/Downloads/azojete143/www.azojete.com.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng


 

Omoze and Emagbetere: Artificial Neural Network (ANN) Pathloss Prediction Model for LTE Network for Microcells in 

an Urban Environment. AZOJETE, 18(4):575-586. ISSN 1596-2644; e-ISSN 2545-5818, www.azojete.com.ng 

Corresponding author’s e-mail address: edirin.omoze@uniben.edu     577 

 

A neural network consists of several layers. The outer most layers that receive the input to 

the network are called the input layers. The next layer is called the hidden layer. This layer 

does further processing of the signal and passes it to the output layer. Shown in Figure 1 is 

the structure of an artificial neuron with weights and bias. 

 

Figure 1:  Analysis of a single neuron with weights and bias 

A neuron (𝑎 ) is presented in Figure 1 with a vectorial input = [𝑥1, 𝑥2, … , 𝑥𝑛] . These inputs 

are multiplied with weights 𝑊 = [𝑤𝑗1, 𝑤𝑗2, … , 𝑤𝑗𝑛]  alongside the summation of a bias term 

to produce a weighted sum as in equation (1). 

𝑂𝑗 = ∑ 𝑥𝑖𝑤𝑗𝑖 + 𝑏𝑛
𝑖,𝑗=1                                             (1)                                                                     

The weighted sum in equation (1) is fed into a non-linear mapping function 𝑓 called the 

activation function which produces a scalar output 𝑎1,𝑛 as in equation (2) [11]. 

𝑎1,𝑛 = 𝑓(∑ (𝑥𝑖𝑤𝑗𝑖)
𝑛
𝑖,𝑗=1 + 𝑏)                                    (2) 

The ANN learning is obtained by updating the weights along the MLP artificial neural 

network in consecutive iterations of feed forward and back-propagation procedures with 

the matrix of neurons in the hidden layer as in equation (3).  

𝑧𝑛
(𝑙) = 𝑓(𝑎𝑛

(𝑙)) = [𝑓(𝑎1,𝑛
(𝑙)), 𝑓(𝑎2,𝑛

(𝑙)), 𝑓(𝑎3,𝑛
(𝑙)), … , 𝑓(𝑎𝑘,𝑛

(𝑙))]
𝑇
            (3) 

where 𝑎𝑛
(𝑙) = {

𝑤(𝑙,𝑛). 𝑋𝑛     𝑓𝑜𝑟 𝑙 = 1

𝑤(𝑙,𝑛). 𝑧𝑛
(𝐿−1)   𝑓𝑜𝑟 2…𝐿 − 1

 

For m features, the weight matrix for layer 𝑙 = 1 and for layers 𝑙 = 2,3, … , 𝐿 − 1 associated 

with 𝑋𝑛=[𝑥1,𝑛, 𝑥2,𝑛 , … , 𝑥𝐷,𝑛] are given by 

𝑤(1,𝑛) =

[
 
 
 
 𝑤1,1

(1,𝑛)
𝑤1,2

(1,𝑛)

𝑤2,1
(1,𝑛)

𝑤2,2
(1,𝑛)

⋯
𝑤1,𝑚

(1,𝑛)

𝑤2,𝑚
(1,𝑛)

⋮ ⋱ ⋮

𝑤𝑘,1
(1,𝑛)

𝑤𝑘,2
(1,𝑛) ⋯ 𝑤𝑘,𝑚

(1,𝑛)
]
 
 
 
 

 and 

http://www.azojete.com.ng/
file:///C:/Users/HP/Documents/Engr%20Oyeniyi/azojete/AZOJETE%20ARCHIVE/UPLOAD/VOL%2018%20NO%204/edirin.omoze@uniben.edu


 

Arid Zone Journal of Engineering, Technology and Environment, December, 2022; Vol. 18(4):575-586. ISSN 1596-2644;  

e-ISSN 2545-5818; www.azojete.com.ng 

 

Corresponding author’s e-mail address: gedirin.omoze@uniben.edu     578 

 

𝑤(𝑙.𝑛) =

[
 
 
 
 𝑤1,1

(𝑙,𝑛)
𝑤1,2

(𝑙,𝑛)

𝑤2,1
(𝑙,𝑛)

𝑤2,2
(𝑙,𝑛)

⋯
𝑤1,𝑚

(𝑙,𝑛)

𝑤2,𝑚
(𝑙,𝑛)

⋮ ⋱ ⋮

𝑤𝑘,1
(𝑙,𝑛)

𝑤𝑘,2
(𝑙,𝑛) ⋯ 𝑤𝑘,𝑚

(𝑙,𝑛)
]
 
 
 
 

 

Using equation 3, the feed forward computation of Figure 1 is performed as given in 

equation 4 as; 

𝑧𝑛
(𝐿−1) = 𝑓 (𝑤(𝐿−1,𝑛). 𝑓 (𝑤𝐿−2,𝑛 . 𝑓 (… 𝑤2,𝑛 . 𝑓(𝑤(1,𝑛). 𝑋𝑛))))               (4) 

The prediction value 𝑦𝑝 from the final layer of the feedforward procedure is a linear output 

of  𝑧𝑛
(𝐿−1) and 𝑤(𝐿.𝑛) as in equation (5). 

𝑦𝑝
(𝐿) = 𝑤(𝐿.𝑛). 𝑧𝑛

(𝐿−1)                                         (5) 

Where: 

 𝑤(𝐿.𝑛) = [𝑤1,1
(𝐿), 𝑤1,2

(𝐿), 𝑤1,3
(𝐿), … , 𝑤1,𝑘

(𝐿)] 

f: is the activation function from one neuron to another 

𝑋𝑛: represents the nth column of the 𝑚 × 𝑛 elmatrix in the input layer 

𝑎𝑛
(𝑙): represents the output of neuron 𝑎 in layer 𝑙 

𝑦𝑝: is the output of the neural network 

2. Materials and method 

This section presents the description of the study area as well as materials and method 

adopted in this work which includes the data measurement setup, data acquisition process, 

data pre-processing and model development. 

2.1 Description of study area 

Edo state is one of the 36 states in Nigeria. It lies on latitude 6°30'N and 6º00'E. The state is 

in the south geopolitical zone of Nigeria, bounded by Kogi state to the North East, Anambra 

State to the East, Delta State to the South and Ondo State to the west with a total land area 

of 17.802km2 as shown in Figure 2. Edo State has 18 local government areas with Benin city 

as the State capital and the largest urban Centre. NIFOR is one of rural communities in Edo 

state. For the purpose of this study, Benin City will be referred to as location 1 and NIFOR 

location 2. 

file:///C:/user/Downloads/azojete143/www.azojete.com.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng


 

Omoze and Emagbetere: Artificial Neural Network (ANN) Pathloss Prediction Model for LTE Network for Microcells in 

an Urban Environment. AZOJETE, 18(4):575-586. ISSN 1596-2644; e-ISSN 2545-5818, www.azojete.com.ng 

Corresponding author’s e-mail address: edirin.omoze@uniben.edu     579 

 

  
Figure 2: Map of Nigeria showing location of Edo state (https://www.guardian.ng/news 

accessed 10th July, 2020) 

 

2.2 Measurement Set-Up and Data Collection 

In order to develop suitable model for LTE cellular network situated in Benin city, Benin city 

was divided into ten sub locations which reflect the major areas of the city. This is because 

of the large number of base stations in Benin City. An experimental set up was designed and 

implemented to measure and record received signal strength data from one of the mobile 

network providers. The base station transmits signals using sectorial antennas spaced at 

120° apart with antenna heights ranging from 32 to 38 meters. The measurement set-up 

consists of base stations, an acer smart phone, a Global Positioning System and a personal 

computer as shown in Figure 3. 

 

Figure 3: Received Signal Strength Measurement Set up 

http://www.azojete.com.ng/
file:///C:/Users/HP/Documents/Engr%20Oyeniyi/azojete/AZOJETE%20ARCHIVE/UPLOAD/VOL%2018%20NO%204/edirin.omoze@uniben.edu
https://www.guardian.ng/news


 

Arid Zone Journal of Engineering, Technology and Environment, December, 2022; Vol. 18(4):575-586. ISSN 1596-2644;  

e-ISSN 2545-5818; www.azojete.com.ng 

 

Corresponding author’s e-mail address: gedirin.omoze@uniben.edu     580 

 

2.2.1 Measurement Campaign 

In order to study the network from subscribers’ point of view, a network cell info lite 

software installed on an acer phone was used to collect RSS data at intervals of 50m away 

from the base station during a download from the network. The measured data were 

transferred to the laptop through a Universal Serial Bus (USB) interface. In addition, a 

handheld GPS (GARMIN GPS 76CS) receiver was used to obtain the spatial coordinates of 

the measuring points in degrees. RSS collected was used to calculate the pathloss of the 

transmitted signals. (Omoze, 2021). 

2.3 Data Pre-processing 

Data pre-processing is an important step in preparing data set for ANN algorithm to 

enhance performance of the learning algorithm. The step taken to pre-process the field data 

collected in this work is feature normalization. 

2.3.1 Feature Normalization 

Normalization is a data pre-processing step in which the data collected are rescaled to lie 

between 0 and 1. This is necessary to avoid weight appropriation bias. In this work, the 

normalization technique used is the standard scaler technique presented in equation (6). 

�̅�𝑖 =
𝑥𝑖−�̅�

𝜎
                                                          (6) 

where �̅�𝑖 is the standardized of 𝑥𝑖 , �̅� is the expected value of the feature given denoted as 

E(𝑥); 

𝐸(𝑥) = �̅� =
1

𝑛
∑ 𝑥𝑖

𝑛
𝑖=1                                         (7) 

and 𝜎 is the standard deviation of a feature given as ; 

𝜎 = √
1

𝑛
∑ (𝑥𝑖 − �̅�)2𝑛

𝑖=1                                          (8) 

where the subscript 𝑖 denotes the sample number and n the number of samples. 

Feature scaling is done on the features of the dataset using the StandardScaler module 

imported from sklearn library using the fit transform command on the feature.  

2.4 Model Development 

In order to develop model with high accuracy and generalization ability, the RSS data 

collected with respect to distance (𝑑), base station antenna height ℎ𝑡𝑥 and mobile station 

antenna height ℎ𝑟𝑥 was  divided into 80% training data set and 20% test data set using the 

stratified shuffle split technique in the Scientific Python Development Environment software 

(SPYDER). Figure 4 is the flow chart of model development adopted in this study.  

file:///C:/user/Downloads/azojete143/www.azojete.com.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng


 

Omoze and Emagbetere: Artificial Neural Network (ANN) Pathloss Prediction Model for LTE Network for Microcells in 

an Urban Environment. AZOJETE, 18(4):575-586. ISSN 1596-2644; e-ISSN 2545-5818, www.azojete.com.ng 

Corresponding author’s e-mail address: edirin.omoze@uniben.edu     581 

 

 
Figure 4:  Flow chart of model development adopted in this study 

 

2.4.1 Artificial Neural Network (ANN) Training 

In this work, a 4-32-1 multi-level perceptron neural network was used to train the training 

data set. The input to the network is a feature matrix of distance (𝑑)𝑚, received signal 

strength values (𝑅𝑆𝑆)(𝑑𝐵𝑚) , base station antenna height (ℎ𝑡𝑥)𝑚  and mobile station 

antenna height (ℎ𝑟𝑥)𝑚 while the output is the pathloss values (𝑃𝐿)𝑑𝐵. Where pathloss (𝑃𝐿) 
is given in equation (9). 

𝑃𝐿(𝑑𝐵)
= 𝑃𝑡𝑥(𝑑𝐵𝑚) − 𝑃𝑟(𝑑𝐵𝑚)                                         (9) 

where 𝑃𝑡𝑥  is the transmitted power by the base station antenna and 𝑃𝑟  is the measured 

power (RSS). Figure 5 shows the artificial neural network architecture used in this work.  

http://www.azojete.com.ng/
file:///C:/Users/HP/Documents/Engr%20Oyeniyi/azojete/AZOJETE%20ARCHIVE/UPLOAD/VOL%2018%20NO%204/edirin.omoze@uniben.edu


 

Arid Zone Journal of Engineering, Technology and Environment, December, 2022; Vol. 18(4):575-586. ISSN 1596-2644;  

e-ISSN 2545-5818; www.azojete.com.ng 

 

Corresponding author’s e-mail address: gedirin.omoze@uniben.edu     582 

 

 
Figure 5: Block diagram of a MLP-artificial neural network with 4 features 

 

The hidden layer activation function is the rectified linear unit (relu) as presented in 

equation (10) while the output layer activation function is the linear activation function as 

given in equation (11). 

𝑓(𝑥) = {
𝑥             𝑥 > 0
0             𝑥 ≤ 0

                                       (10) 

𝑓(𝑥) = 𝑎𝑥                                                        (11) 

where 𝑎 = 1 

2.4.2 Cross validation 

Cross Validation (CV) is a process that provides continuous monitoring of the generalization 

performance of the model over the learning process. CV requires splitting the data into 2–3 

non-overlapping subsets: the training set which is used for parameter learning, the validation 

set on which the prediction error is used to approximate the generalization error, and the 

test set on which the prediction error is used to approximate the unbiased generalization 

error after the learning is completed. In this work, Early Stopping technique was adopted. 

To utilise early stopping, 10% of the training set was used as a validation set with a patience 

of 10 and the validation mean square error was monitored during the training process. 

2.5 Model Validation 

The validation of the developed model is a vital aspect of the modeling process; essential to 

its evaluation are the values of the goodness of fit indices obtained from the analysis of the 

given model. In this section, the artificial neural network model is validated to ascertain its 

suitability for the data collected in the environment investigated.  The Root mean square of 

errors (RMSE) and mean absolute error (MAE) are computed using equation (12) and 

equation (13) and are used to evaluate the goodness of fit of the developed model. 

file:///C:/user/Downloads/azojete143/www.azojete.com.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng


 

Omoze and Emagbetere: Artificial Neural Network (ANN) Pathloss Prediction Model for LTE Network for Microcells in 

an Urban Environment. AZOJETE, 18(4):575-586. ISSN 1596-2644; e-ISSN 2545-5818, www.azojete.com.ng 

Corresponding author’s e-mail address: edirin.omoze@uniben.edu     583 

 

RMSE=√
1

𝑛
∑ (𝑦𝑖 − 𝑦𝑝(𝑖))

2𝑛
𝑖=1                                         (12) 

𝑀𝐴𝐸(𝑦𝑖, 𝑦𝑝) =
1

𝑛
∑ |𝑦𝑖 − 𝑦𝑝(𝑖)|

𝑛
𝑖=1                             (13) 

Where 𝑦𝑖  is the pathloss estimated from the measured received signal strength data and 

𝑦𝑝(𝑖) is the pathloss predicted using the developed artificial neural network model. 

By computing equations (12) and (13), RMSE and MAE error values of 0.59𝑑𝐵 and 0.42𝑑𝐵 
respectively. 

3. Results and Discussion 

This section presents the results obtained from training received signal strength data 

collected in real time from LTE (4G) mobile networks in Benin city. Because of the problem 

of overfitting, the data collected for the 10 zones were merged and used to train the 

artificial neural network. Table 1 shows descriptive statistics of the data collected in this 

study.  

Table 1: Table of Descriptive Statistics of Data Collected for LTE (4G) Cellular Network 

Index 𝑑 𝑅𝑆𝑆 𝐻𝑟𝑥 𝐻𝑡𝑥 𝑃𝐿 
count 294 294 294 294 294 

mean 309.18 -77.33 1.5 35.22 107.83 

std 178.69 7.16 0 2.86 7.16 

min 50 -96 1.5 30 92.17 

25% 150 -82.86 1.5 34 102.74 

50% 300 -78.19 1.5 36 108.19 

75% 450 -72.74 1.5 36 112.86 

max 800 -62.17 1.5 40 126 

It presents received signal strength data collected from 294 sample points for a maximum 

cell radius of 800m with average RSS value of -77.33dBm. The average base station antenna 

height at this location for the 700MHz network was 35.22m with 126dBm pathloss value 

obtained at the cell edge. It is clear that the 700MHz cells in location 2 are micro cells.  

From literature, pathloss prediction for mobile communication systems is often treated as a 

supervised learning problem, thus the data collected for this investigation was divided into 

feature DataFrame and target DataFrame using DataFrame indexing techniques in the 

scientific python development environment. 

In order to develop models with high generalization ability on unseen samples, the 80% 

training data was trained using a 4-32-1 multi-layer perceptron feedforward neural network 

built using the keras functional API on tensorflow backend and learning was done by back 

propagating the errors. The input to the neural network is the normalized values of the 

feature DataFrame with the distance, received signal strength, mobile station antenna height 

and base station antenna height as feature variables. In the neural network architecture 

perspective, three key factors are considered, the type of activation function, the number of 

hidden layers and the number of hidden nodes on each layers. 

http://www.azojete.com.ng/
file:///C:/Users/HP/Documents/Engr%20Oyeniyi/azojete/AZOJETE%20ARCHIVE/UPLOAD/VOL%2018%20NO%204/edirin.omoze@uniben.edu


 

Arid Zone Journal of Engineering, Technology and Environment, December, 2022; Vol. 18(4):575-586. ISSN 1596-2644;  

e-ISSN 2545-5818; www.azojete.com.ng 

 

Corresponding author’s e-mail address: gedirin.omoze@uniben.edu     584 

 

In order to avoid overfitting of the training dataset, dropout regularization technique was 

adopted in the hidden layer with a dropout probability of 0.3. Since the aim of training is to 

predict real value number as the pathloss, the output layer is made up of a single neuron 

with the linear activation function. 

The artificial neural network model was compiled using the Adams optimizer with the mean 

square error as the loss function. Training was done using a batch size of 32 for dataset for 

100 epochs. 

Figure 6 shows a comparative analysis of prediction made using artificial neural network 

(ANN) and pathloss values obtained from measured data with respect to distance while 

Figure 7 shows training history of artificial neural network for 100 epochs. 

 
Figure 6: Plot of Measured Pathloss and Pathloss predicted using Artificial Neural Network 

 
Figure 7: Performance curve of trained neural network history for pathloss prediction for 

700MHz network at location 2. 

 

 

file:///C:/user/Downloads/azojete143/www.azojete.com.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng


 

Omoze and Emagbetere: Artificial Neural Network (ANN) Pathloss Prediction Model for LTE Network for Microcells in 

an Urban Environment. AZOJETE, 18(4):575-586. ISSN 1596-2644; e-ISSN 2545-5818, www.azojete.com.ng 

Corresponding author’s e-mail address: edirin.omoze@uniben.edu     585 

 

3.1. Discussion 

The artificial neural network model was compiled using the Adams optimizer with the mean 

square error as the loss function. Training was done using a batch size of 32 for dataset for 

100 epochs. Figure 6 is a plot of pathloss measured with respect to distance. This plot 

shows that pathloss increases with respect to distance. It also shows that artificial neural 

network was able to learn and predict the behaviour of the network.  

Figure 7 is a plot of Performance curve of trained neural network history for pathloss 

prediction for 700MHz network. Here, the network was trained for 100 epochs. At 0 

epoch, the network had not learnt the variation in the measured data thus the high loss, 

mean absolute error and mean square error values. As the number of epoch increases, the 

network learns more about the data set and this can be seen in the lower high loss, mean 

absolute error and mean square error values obtained at 70 epoch. The plot also shows that 

at 70 epoch, the network has fully learnt the variation in the data set presented. 

4. Conclusion 

This study presents ANN pathloss model development for LTE network in an urban 

environment. Using received signal strength data measured with respect to base station 

antenna height, mobile station antenna height and distance from the base station, pathloss 

was estimated. A feature matrix consisting of 4× 240 elements was used to train a 4-32-1 
multilevel perceptron artificial neural network with pathloss values as the label. The hidden 

layer activation function is the rectified linear unit activation function while the activation 

function for the output layer is the linear activation function. By performing cross validation 

with a validation set of 10%, ANN training was achieve at the 74th epoch with a mean 
square error value of 0.0058dBm and a validation mean square error value of 0.00086dBm. 

The result obtained from this study shows that artificial neural network with   RMSE and 

MAE values of 0.59𝑑𝐵  and 0.42 𝑑𝐵  respectively were obtained when compared with 
pathloss estimated from the measured data. This shows that the developed artificial neural 

network algorithm will provide good of LTE networks if deployed during design stage in the 

Benin City.  

References 

Atanasov, Ph. and Kiss'ovski, Zh. 2017. Optimization of path loss models based on signal 

level measurements in 4G LTE network in Sofia. Bulgarian Journal of Physics, 44(02): 145-

154. 

Bi, Q., Zysman, GL. and Menkes, H. 2001. Wireless mobile communications at the start of 

the 21st century. IEEE communications magazine, 39(1): 110-116. Doi:10.1109/35.8943 

Eichie, JO., Oyedum, OD., Ajewole, MO. And Aibinu, AM. 2017. Comparative analysis of 

basic models and artificial neural network based model for path loss prediction. Progress in 

Electromagnetic Research, 61: 133-146. 

Jo, H., Park, C., Lee, E., Choi, HK. and Park, J. 2020. Path loss prediction based on machine 

learning techniques: principal component analysis, artificial neural network, and gaussian 

process. Sensors, 20: 1927., Doi:10.3390/s20071927. Available online 

www.mdpi.com/journal/sensors. 

http://www.azojete.com.ng/
file:///C:/Users/HP/Documents/Engr%20Oyeniyi/azojete/AZOJETE%20ARCHIVE/UPLOAD/VOL%2018%20NO%204/edirin.omoze@uniben.edu


 

Arid Zone Journal of Engineering, Technology and Environment, December, 2022; Vol. 18(4):575-586. ISSN 1596-2644;  

e-ISSN 2545-5818; www.azojete.com.ng 

 

Corresponding author’s e-mail address: gedirin.omoze@uniben.edu     586 

 

Ojo, S., Imoize, A. and Alienyi, D. 2021. Radial basis function neural network path loss 

prediction model for LTE networks in multi-transmitter signal propagation environments. 

International Journal of Communication Systems, 34 (03): e4680, doi:10.1002/dac.4680. 

Omoze, EL. 2021. Radio signal characterization of cellular networks and machine learning 

based path loss model development. PhD thesis, Department of Electrical and Electronics 

Engineering, Faculty of Engineering, University of Benin. 

Sharma, PK., Sharma, D. and Sai, TV. 2018. Optimization of propagation path loss model in 

4G wireless communication systems. Proceedings of the Second International Conference 

on Inventive Systems and Control (ICISC 2018), 19-20 January, Coimbatore, India: 1245-

1248, doi:10.1109/icisc.2018.8399004 

Shoewu, O., Akinyemi, LA. and Oborkhale, L. 2019. Modelling path loss in mobile 

communication 4G network system for dryland and wetland Terrains. Proceedings of 

Southern Africa Telecommunication Networks and Applications Conference (SATNAC), 1-
4 September, Fairmont Resort, Ballito, KwaZulu-Natal, South Africa: 44-49. 

Sree Vardhan, C., Venkat Ratnam, D., Bhargyasree, N. and Hema Dattu, A. 2014. Analysis of 

path loss models of 4G femtocells. Proceedings of the Eleventh International Conference on 

Wireless and Optical Communications Networks (WOCN), 11-14 September, Vijawada, Andhra 

Pradesh, India: 1-6, doi: 10.1109/WOCN.2014.6923071 

Sulyman, AI., Nassar, AT., Samimi, MK., MacCartney, GR., Rappaport, TS. And Alsanie, A. 

2014. Radio propagation path loss models for 5G cellular networks in the 28 GHz and 38 

GHz millimeter-wave bands.  IEEE communications magazine, 52(09): 78-86. 

 

Ubaidillah, A., Ulum, M. and Pramono, S. 2020.  Planning of path loss in large scale signal 

propagation model at 5G network by using linear regression methods. Journal of Physics: 

Conference Series, 1569(3):  032080.  

Zhang, Y., Wen, J., Yang, G., He Z. and Wang, J.  2019. Path loss prediction based on 

machine learning: principle, method, and data expansion.  Applied Sciences, 09: 1908, 

Doi:10.3390/app9091908. Available online @www.mdpi.com/journal/appl 

file:///C:/user/Downloads/azojete143/www.azojete.com.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng
file:///C:/Users/HP/AppData/Roaming/Microsoft/Word/giagbara@unimaid.edu.ng