2022 International Conference on Broadband Communications for Next Generation Networks and Multimedia Applications (CoBCom) | 978-1-6654-8598-2/22/$31.

00 ©2022 IEEE | DOI: 10.1109/COBCOM55489.2022.9880654

Downlink Throughput Prediction in LTE Cellular

Networks Using Time Series Forecasting
Ali Mostafa1, ∗ , Mustafa A. Elattar1, † , Tawfik Ismail1,2, ‡ ID ,
1
Communications & Information Technology (MCIT), Nile University, Giza, Egypt
2
Department of EAL, National Institute of Laser Enhanced Sciences, Cairo University, Giza, Egypt
Email addresses: ∗ [email protected], † [email protected], ‡ [email protected]

Abstract—Long-Term Evolution (LTE) cellular networks have team relied on the traditional method using a conventional
transformed the mobile business, as users increasingly require linear regression [4]. In order to forecast cell downlink traffic
various network services such as video streaming, online gaming, and predict UE DL throughput for the upcoming year, it is
and video conferencing. A network planning approach is required
for network services to meet user expectations and meet their not sufficient to depend only on downlink traffic itself. As
needs. The User DownLink (UE DL) throughput is considered the previously stated, this will not continue to be successful.
most effective Key Performance Indicator (KPI) for measuring the The main contribution of this paper is the use of deep
user experience. As a result, the forecast of UE DL throughput learning and machine learning to combine multiple KPIs in
is essential in network dimensioning for the network planning UE DL throughput prediction by combining regression neural
team throughout the network design stage. The proposed system
employs several KPIs to predict UE DL throughput by combining network [5] and seasonal ARIMA [6] on a daily live network
machine learning and deep learning framework for a time series cells KPIs to produce a more accurate UE DL throughput pre-
forecasting rather than the traditional statistical technique based diction to improve accuracy. The paper covers all kinds, from
on downlink traffic only. The proposed scheme identifies the most selecting the most important KPIs affecting UE DL throughput
significant KPIs that affect UE DL throughput and provides to optimizing the regression neural network. This helps the
accurate results based on prediction.
engineers for better network planning and better selection for
Index Terms—Mobile Networks, Capacity Planning, Regression spectrum expansion solutions based on the predicted upcoming
Neural Network, seasonal Auto-Regressive Integrated Moving Av- year UE DL throughput verses the minimum accepted UE DL
erage (ARIMA), Throughput Forecast, Machine Learning (ML).
throughput defined by the management team.
The rest of the paper is organised as follows. Section II illus-
trates the system model architecture considering the machine
I. INTRODUCTION learning and Neural Network . Section III shows the simulation
results. Finally,Section IV presents conclusions and discusses
Increasing demand for data services forces users to rely on
future work.
cellular LTE networks for a range of services such as video
streaming and online gaming [1], which become essential to II. SYSTEM MODEL
people. Currently, operators consider the spectrum to be one of
the most precious resources available, and they must maximize The model is divided into two main parts that integrate
the utilization of the spectrum with providing the best possible to implement the overall system. As shown in Fig. 1, the
user experience. This has resulted in new features and strategies regression neural network uses non-historical data of Cells
for spectrum sharing between different technologies, And the KPIs, and the ARIMA model, which uses historical data for
development of new technologies. cells for which we want to predict their UE DL throughput.
Mobile operators must have great confidence in their selec-
tion of LTE cells that require spectrum expansion to reduce A. Regression Neural Network
network bottlenecks. In order to avoid investing in new hard- Regression neural network is used to predict optimal features
ware that is not completely exploited, as in the case of adding a while attempting to keep the Root Mean Square Error (RMSE)
new sector or expanding the number of Multiple Input Multiple as low as possible. This system is divided into three major
output (MIMO) antennas, or using spectrum sharing solutions sections: data collection, feature selection, and training and
which may affect the user KPIs with the improper planning. validation.
As a result, mobile operators are considering the UE DL • Data collection: The data-set is consists of 157K LTE cell
throughput as one of the most essential LTE key performance KPIs records for 15K live network cells in Egypt, the cells
indicators for network planning [2]. are Frequency Division Duplex (FDD) operating on LTE
The importance of using the UE DL throughput for network Band 3. The records were collected on 20 days between
dimensioning is crucial. As cellular networks display downlink August and September 2021 over different regions from
throughput fluctuations more significant than the other wired Etisalat Misr operator. Each record includes A features
transmission technologies due to rapid changes in radio con- for the cell that may affect the UE DL throughput from
ditions and user mobility [3] for network operators, it is very available spectrum represented by available number of
important to reflect the user experience accurately. The majority Physical Resource Block (PRB), Channel Quality Indica-
of the operator’s throughput predictions are based strictly on tor (CQI), MIMO rank 2%, downlink modulation scheme
traffic forecasting and do not consider the unique characteristics to E-UTRAN Radio Access Bearer (ERAB drop rate) and
of each cell in the network. In this particular case, the capacity Uplink Block Error Rate (UL BLER).

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on April 06,2024 at 02:38:02 UTC from IEEE Xplore. Restrictions apply.
Fig. 1. Overall System Diagram

• Feature selection: The model employs two stages of se- the MIMO RANK 2 PER, the percentage of modulation
lection: a filter method (Pearson Correlation) for remov- scheme 64QAM QAM64 PER, the available number of PRB,
ing features that have a low correlation with UE DL the DownLink Traffic in gigabytes DL TRAFFIC GB.
throughput, and an embedded method (Random Forest)
for obtaining each feature importance score and removing TABLE I
features with low importance, in order to obtain a final set F EATURE CORRELATION WITH UE DL T HROUGHPUT
of C features where C < A. KPI Abs correlation
• Training and validation: With the help of A features, we DL PRB UTI 0.70
may train and optimize the regression Neural Network to DL IBLER 0.69
CQI AVERAGE 0.62
obtain the lowest RMSE from the training and validation CCE USAGE PER 0.59
datasets, and we can use backward elimination for features MIMO RANK 2 PER 0.58
to obtain the final D features where D < C.

B. Seasonal ARIMA Time-Series Prediction TABLE II

A seasonal ARIMA is utilized to provide predicted values F EATURES PAIRS C ORRELATION
for P features, which are used as input for a regression neural Pair# KPI1 KPI2 Correlation
network to predict the future values of UE DL throughput [7]. 1 UL TRAFFIC GB DL TRAFFIC GB 0.93
Seasonal ARIMA input is comprised of daily historical data per 2 DL RETRANS DL IBLER 0.89
3 CCE USAGE PER DL PRB UTI 0.84
cell collected since the year 2019 till the end of 2021. Each 4 UL TRAFFIC GB DL PRB UTI 0.80
cell trend is collected and sampled on a weekly basis then
investigated to obtain the best seasonal ARIMA model param-
eters for both the non-seasonal model portion (p, d, q) and the
seasonal model part (P, D, Q)m where p is the non-seasonal B. Regression Neural Network
auto-regression order, d is the non-seasonal differencing,q is A 3−layers regression neural network model with Adam
the non-seasonal moving average order, P is the seasonal auto- optimizer was developed with input dimension C. For data
regression order, D is the seasonal differencing, and Q is the validation, neural network optimization and features reduction
seasonal moving average order. are improved in order to obtain the optimal neural network
parameters while maintaining an acceptable RMSE. Features
III. S IMULATION R ESULTS AND D ISCUSSIONS with a C = 7 were reduced to D where D = 3 features.
A. Feature selection These features are used as the regression neural network
After applying the feature selection technique to a dataset of inputs, the 3 features are the Downlink Traffic in Gigabytes
A features and N = 32 features, the dataset was reduced to DL TRAFFIC GB, the Downlink Physical Resource Block
B where B = 12 features after removing low correlated ones Utilization DL PRB UTI and the available of Downlink Re-
with UE DL throughput, and with one feature of the highly source Block number PRB, after the learning process for the
correlated feature pairs being removed to reduce the system neural network using these 3 features as input B and the UE
complexity. Table I provides a sample of feature correlations DL throughput as output Y the Mean Square Error (MSE) for
with UE DL throughput. In Table II, we can see a sample of test and validation reaches as low as 4.8Mbps which equal to
feature pairs correlation, and after reducing features with low 2.18Mbps RMSE as shown in Fig. 2.
importance, resulting in C where C = 7 features.
The remaining Y features are the DownLink Physical Re- C. Features Prediction using Seasonal ARIMA
source Block Utilization DL PRP UTI, the DownLink Block The seasonal ARIMA model was used to predict 2 out of 3
Error Rate (DL BLER), the Channel Quality Indicator (CQI), inputs for the regression neural network for each cell, with the

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on April 06,2024 at 02:38:02 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. DL TRAFFIC GB Forecast

Fig. 2. Neural network RMSE vs no. of epochs

third feature being the available number of PRB, which is a

configured parameter per cell, As a result, seasonal ARIMA is
applied to predict the DL PRB UTI and DL TRAFFIC GB
Following that, we calculate the estimated DL PRB UTI as
follows
DL PRB UTI = 100×(Used DL PRBs /Available DL PRBs)
(1)
Seasonal ARIMA input was developed using daily historical
data for each feature gathered since the beginning of 2019.
The data were sampled on a weekly basis for a total of 52
values annually. Auto-Correlation Factor (ACF) and Partial
Auto-Correlation Factor (PACF) were determined for each trend
to identify the presence of Auto Regressive (AR) and Moving
Average (MA) components in the residuals. It is also necessary
to obtain the seasonal component, after which iterations on
the seasonal and non-seasonal parameters are performed. Then Fig. 4. UE DL Throughput Predicted Forecast for sample cells with RMSE
using these parameters to predict feature values till the end of (a) 1.3Mbps (b) 1.7Mbps (c) 1.8Mbps (d) 2.7Mbps (e) 2.9Mbps (f) 3.6Mbps
2023. As example in Fig. 3 for DL TRAFFIC GB prediction
for a sample cell with (p, d, q) & (P, D, Q)m parameters
equal to (1, 1, 1) & (1, 1, 1)52 with RMSE of 4.6GBs for the considered major KPIs for representing the user experiences.
predicted period from May to December 2021. Integrated machine learning and deep learning techniques are
used to achieve more accurate results and improve the perfor-
D. Model Prediction Results mance of currently used statistical methodologies. Following
For each cell, the seasonal ARIMA predicted the features the three-stage feature selection procedure, three network KPIs
DL TRAFFIC GB, no. of PRBs, and DL PRB UTI. According (availability of PRBs, downlink PRB utilization, and downlink
to the proposed system architecture, these features were used traffic) are selected as the most effective KPIs of the UE DL
as inputs for the trained NN to predict the UE DL throughput throughput for the network planning team to use as input for
for the targeted cell. The predication results were evaluated the prediction. The Mean Absolute Error (MAE), MSE, and
by using RMSE between the actual and predicted UE DL RMSE metrics are performed to evaluate the two algorithms
throughput for the period from May 2021 to January 2022. studied in this research: seasonal ARIMA KPIs predictions and
Average RMSE for predicted UE DL throughput for the tested Regression neural network. The proposed system has recorded
samples is 2.3Mbps, which represent almost 10% of the average around 10% RMSE. These results will significantly improve the
actual UE DL throughput of the tested samples which is accuracy of the whole LTE cellular network planning process.
22.7Mbps. Six sample cells with actual and predicted UE DL On the subject of future work, it is possible to improve the study
throughput indicated by the RMSE is shown in Fig. 4. With by exploring various machine learning algorithms to forecast
RMSE varying between 1.3Mbps to 3.6Mbps the UE DL throughput, which significantly impacts network
planning. In addition, new network KPIs, such as downlink
IV. C ONCLUSION AND F UTURE W ORK BLER and CQI, can be introduced into these studies to show
The proposed framework in this paper could be used to the effectiveness of these KPIs on the user experience, which
predict UE DL throughput in LTE cellular networks, which are will be beneficial to the network planning team.

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on April 06,2024 at 02:38:02 UTC from IEEE Xplore. Restrictions apply.
 ACKNOWLEDGEMENT
This work would not have been possible without the support
of Etisalat Misr Company for providing me the academic time
and all the support to pursue my career goals. I am especially
indebted to Tarek AbuBakr, Head of Customer Experience
Department for his continued support and guidance.

R EFERENCES
[1] H. Elsherbiny, H. M. Abbas, H. Abou-zeid, H. S. Hassanein, and
A. Noureldin, “4g lte network throughput modelling and prediction,” in
GLOBECOM 2020 - 2020 IEEE Global Communications Conference,
2020, pp. 1–6.
[2] D. Lee and J. Lee, “Machine learning and deep learning for throughput
prediction,” in 2021 Twelfth International Conference on Ubiquitous and
Future Networks (ICUFN), 2021, pp. 452–454.
[3] Y. Liu and J. Y. B. Lee, “An empirical study of throughput prediction in
mobile data networks,” in 2015 IEEE Global Communications Conference
(GLOBECOM), 2015, pp. 1–6.
[4] S. Cheusheva, “How to forecast in excel: lin-
ear and non-linear forecasting methods,” Ablebits. [On-
line]. Available: https://www.ablebits.com/office-addins-blog/2019/03/20/
forecast-excel-linear-exponential-smoothing-forecasting-models/
[5] F. Y. Wu and K. K. Yen, “Applications of neural network in regression
analysis,” Computers & Industrial Engineering, vol. 23, no. 1, pp. 93–95,
1992.
[6] B. Zhou, D. He, and Z. Sun, “Traffic modeling and prediction using
arima/garch model,” in Modeling and Simulation Tools for Emerging
Telecommunication Networks, 2006, pp. 101–121.
[7] e. a. Shu, Yantai, “Wireless traffic modeling and prediction using seasonal
arima models,” IEICE transactions on communicationsg, vol. 88, no. 10,
pp. 3992–3999, 2005.

Authorized licensed use limited to: Institut Teknologi Bandung. Downloaded on April 06,2024 at 02:38:02 UTC from IEEE Xplore. Restrictions apply.