International Journal of Computer Theory and Engineering

Abstract

The rapid deployment of 5G mobile communications has drastically increased the energy consumption of radio access networks, and Base Stations (BSs) account for 60%–80% of network energy consumption. This work presents a low-complexity predictive base station sleep mode framework based on Long Short-Term Memory (LSTM) traffic forecasting and a threshold-based control loop. It is based on proactive forecasting that is able to automatically transition BSs to sleep mode during low-demand state transitions in the predicted low-demand periods of the network, thereby minimizing frequent state-switching overhead and extending low-power intervals. Based on the real-world Telecom Italia Mobile Big Data Challenge data set (Milan and Trentino), our framework realized energy savings of up to 45.8% over the always-on baseline and 32.0% over the reactive baseline. Most importantly, our system does not breach Quality of Experience (QoE) service at the optimal 100 Mbps speed, and it requires much less computational effort than state-of-the-art reinforcement learning structures. A multi-objective sensitivity analysis presents the Pareto front to capture energy efficiency and service reliability. These results indicate that the lightweight predictive engineering frameworks can be a very practical and scalable way to manage the green 5G networks.