Decentralized Federated Deep Reinforcement Learning Framework for Energy-Efficient Base Station Switching Control

G cellular networks have achieved significant improvements in capacity and performance, while they have incurred a substantial increase in energy consumption at the base station (BS) level. To address this escalating energy consumption in dense 5G networks, this paper proposes a decentralized federated learning (DFL)-enhanced Deep Reinforcement Learning (DRL) framework for intelligent BS switching. Our approach leverages the distributed nature of 5G networks to collaboratively learn optimal BS on/off policies without centralized control. By combining DFL with DRL, we enable efficient knowledge sharing among BSs while maintaining privacy and reducing communication overhead. The cost function designed in this paper aims to balance energy savings as well as Quality of Service ($\mathbf{Q o S}$) requirements. Moreover, to enhance exploration and accelerate convergence, we incorporate an exploration network into the DRL agent and adopt a novel approach of model training. Performed simulations demonstrate the effectiveness of our proposed framework in achieving significant energy reduction of over 23% and total cost of 19%, while maintaining satisfactory QoS performance compared to the existing methods.