Explainable Decentralized Federated Learning for Energy-Efficient Base Station Sleep Control

Abstract

Given the capacity and performance boosts offered by the 5 G cellular networks, energy consumption at the base stations (BSs) has increased tremendously. This paper proposes a decentralized federated learning (DFL)-based intelligent BS switching, integrated with explainable artificial intelligence (XAI) methods, to mitigate the concerns for energy consumption in dense 5G networks. This entails collaboration among distributed but interconnected networks to learn the best policies for BS switching without any central controller, so that knowledge sharing can be ensured while privacy and communication efficiency are maintained. Very importantly, we further researched the XAI techniques to provide better transparency on the decisionmaking of the switching control agent and create some trust in the learned policies. Such explainability allows us to derive the most important factors affecting BS switching decisions and how these contribute in enabling energy savings while maintaining quality of service (QoS). Extensive simulations conducted to validate our proposed framework in presenting valuable XAI analysis have elaborately provided the basis for understanding the learned strategies and key factors driving energy-efficient BS management.