Analysis of the Utilization of Mobile Network Base Stations Using Traffic Load Predictions

Abstract

Every day, mobile network traffic is increasing in an unprecedented manner all over the world, resulting in growing demand from the network operators to deploy more base stations in order to be able to serve more devices while maintaining a satisfactory level of service quality. Base stations are considered the leading energy consumer in a network infrastructure; consequently, increasing the number of base stations will increase power consumption. By finding a method to predict the traffic load on base stations, network optimization techniques can be applied to put inactive base stations into sleep mode thereby decreasing energy consumption. This research explores methods capable of predicting traffic load on base stations. The most common time series forecasting techniques are examined in this research on a public dataset that provides records of traffic loads of several base stations over the span of one week. Because of the limited number of records that exist in the dataset for each base station, and to avoid a common problem often raised in training forecasting algorithms, different base stations are grouped together while building the prediction model. Due to the different behavior of the base stations, forecasting the traffic load of multiple base stations together becomes challenging. Our proposed solution involves clustering the base stations according to their behavior and forecasting the load on the base stations in each cluster individually. Different clustering algorithms along with several similarity metrics are compared to each other using a variety of evaluation methods. The base stations load prediction task is formulated as a time series forecasting problem. The most common statistical techniques, machine learning techniques, and deep learning techniques used for time series forecasting are applied on the clusters. Consequently, the clusters’ performances are evaluated and compared. Lastly, two popular online tools, created specifically for the time series forecasting task, are employed to the dataset. The results generated by these online tools are used in this research as a benchmark to compare to the performance of the proposed prediction model. Our findings demonstrate that Deep Recurrent Neural Networks such as the Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) networks perform better than other time series forecasting techniques for large data sets. Our findings offer clustering the time series data according to their behavior before using them for the forecasting algorithms as a solution for the problem of the dissimilar behavior of the time series when they are trained together.