Deep reinforcement learning for energy-efficient air-cooled servers: Algorithms and applications

Air cooling remains the predominant thermal management strategy for data centers, yet standard control methods often fail to balance energy efficiency with complex thermal coupling in high-density servers. This study proposes a safety-aware Twin Delayed Deep Deterministic Policy Gradient (TD3) framework, optimized via the Taguchi method, to regulate a multi-fan 1U server system. Unlike conventional black-box approaches, a dual-threshold reward mechanism is introduced to strictly enforce junction temperature limits (Max Allowable T_lim = 100 ℃ for Type A / 80 ℃ for Type B heaters) while optimizing power consumption below a critical threshold (T_c = 80 ℃/60 ℃). The agent achieved convergence within 1,500 training iterations. Comparative analysis against a baseline fixed 40% duty cycle strategy demonstrates that the proposed method effectively eliminates hot spots and significantly reduces energy consumption. Specifically, energy savings of 18.7%, 34.2%, and 55.4% were achieved for three typical heat source layouts in Cases 1, 2, and 4, respectively. Furthermore, under a dense, non-uniform layout in Case 4, retraining the algorithm yielded a maximum energy saving of 78.2%. These results confirm that the TD3-based approach outperforms fixed-frequency strategies by implicitly learning non-linear flow-thermal correlations, offering a robust solution for next-generation green data centers.