Advanced Cooling Optimization for 5G Base Station via a Three-Stage Hybrid Learning Approach

As data traffic and information services surge, 5G base stations (BSs) have become primary energy consumers in wireless networks, with cooling accounting for 40% of BS power use and driving high operational expenditure. This underscores the need for energy-efficient optimization of cooling control. However, existing model-based and AI-driven optimization methods face critical limitations in real-world deployments, including limited generalizability across heterogeneous BSs resulting from their high scenario dependence, the inherent tradeoff between control stability and dynamic adaptability under complex operating conditions, and high deployment costs due to computational demands. To address these challenges, this article develops a three-stage hybrid learning (TSHL) approach that integrates imitation learning, ensemble learning, and deep reinforcement learning into a three-stage offline-to-online architecture, enabling expert knowledge transfer from data-rich BSs to generate high-quality initial policies for data-scarce ones, while online learning ensures continuous adaptation to local dynamics. In addition, we propose a cost-efficient deployment mechanism that leverages only existing monitoring data without additional hardware costs, employs hybrid experience policy updates within a deep Dyna-Q-based architecture to enhance learning efficiency, and incorporates a safety-constrained exploration to enhance policy reliability. Extensive evaluations on both a simulation testbed and a real-world 5G BS demonstrate that TSHL achieves over 18.36% cooling energy savings and outperforms baseline methods in cold-start effectiveness, online adaptability, operational reliability, and overall cost-efficiency. These results highlight TSHL as a practical solution for sustainable 5G BS operations, especially for data-scarce BSs, such as retrofitted or newly established sites, offering a scalable pathway to network-wide energy savings.