Neural Networks for Cellular Base Station Switching

The adoption of base station sleep modes is considered one of the most effective approaches for the reduction of the energy consumption of radio access networks. Sleep modes allow the switch-off of base stations in periods of low traffic, and their successive switch-on when traffic increases. The selection of the appropriate instants to switch base stations on and off requires an accurate prediction of the traffic loads in the near future. In this paper we explore the performance of machine learning techniques for traffic prediction and for the selection of the instants when to switch off and on base stations, considering a heterogeneous network portion comprising one macro cell and six small cells within the macro cell coverage. We experiment with two machine learning approaches. The first aims at short-term traffic load estimation, and from this derives the best combination of switching decisions. The second performs both traffic estimation and switching optimization at one time. For both approaches we develop artificial neural network implementations based on the Dense Neural Network and Recurrent Neural Network paradigms. Testing the two approaches on real traffic data, we observe very good performance in terms of both quality of service and energy saving.