Prediction of Load Demand on Base Stations for application of Energy Conservation Using Deliberated Informer

With the growing popularity and accessibility of mobile smart phones in every nook and corner of the world, cellular traffic demand has seen an exponential increase. Consequently, more number of Base Station (BS) are required to cater to this ever increasing demand. But, a close observation leads us to a very interesting fact that load demand on base stations is not constant throughout the day, i.e., there are time periods in the day where usually load demand is at its peak and as day comes to an end, it is seen load demand decreases. But the base stations are continuously in the ON mode which leads to massive energy wastage and telecommunication companies have to pay a heavy financial cost. This forms the basis for our motivation to reduce energy consumption or in other words ‘Energy Saving’ in Radio Access Networks(RAN). Recent studies have confirmed the possibility of RAN energy savings by the dynamic on/off of various BSs. In order to determine the correct time for applying the sleeping mechanisms, we have to predict the off-peak traffic hours. In this work, we propose a novel open-source dataset which contains the load demand versus time data for base stations for 6000 hours to facilitate training of various deep learning models. We propose an efficient transformer based model named Deliberated Informer and evaluate its performance on our dataset which gives state of the art performance in the task of Long sequence time-series forecasting (LSTF).