Joint optimization of multiple energy-saving strategies based on DRL in 5G railway networks

The energy consumption of 5G mobile networks has emerged as a critical challenge, largely driven by the application of multi-antenna systems and the dense deployment of base stations (BSs). This issue becomes particularly serious in railway environments, where traffic is highly sporadic and BSs experience long idle periods. Consequently, there is a compelling need to design energy-efficient schemes specifically suited for railway scenarios. In this paper, we propose a deep reinforcement learning (DRL)-based approach to achieve a reduction of energy consumption for private 5G networks in railway scenarios (5GR). In particular, BS sleeping, antenna selection, and symbol shutdown are jointly optimized while maintaining quality of service (QoS) constraints for the unique characteristics of railway traffic. Considering the varying energy-saving response cycles of different strategies, we develop a framework that operates across multiple time granularities. Moreover, a Conv-LSTMDDQN network (CLDN) is designed to capture the spatiotemporal correlations of train traffic to maximize long-term cumulative rewards, enabling the framework to make decisions with historical traffic information. Furthermore, a DRL approach based on an action space reduction network (ASRN) is proposed to iteratively solve the energy minimization problem at each time granularity. ASRN effectively narrows the action space, enabling more efficient exploration. Simulation results indicate that the proposed scheme markedly reduces power consumption in 5G-R systems even in the absence of instantaneous traffic information. The achieved energy savings are up to 17.66%, and the computation time is significantly reduced compared to traditional heuristic methods.