Energy saving of fans in air-cooled server via deep reinforcement learning algorithm

The present paper aims at using an artificial intelligence algorithm to minimize the fan power consumption in air-cooled servers. The proposed algorithm can handle the complex thermal environments within the servers to tailor the influences and interactions amid numerous heat sources, airflow, bypass phenomenon, fan operation, and the transient operations. Modified correlations are first proposed to effectively predict the thermal-hydraulic performance of heat sinks and the corresponding predictive ability against Nusselt number and pressure drop is within 5.0% and 10%, respectively. Without the algorithm control, the maximum deviation between the prediction and the experimental data is within 2.0 °C. By introducing the deep reinforcement learning (DRL) algorithm subject to the interactions of complex thermal environments, the fan power consumption can be saved by 55.7%, 40.3% and 26.3%, respectively, in comparison with the strategy with 100% fan duty. Yet the DRL agent still offers 16.7% energy saving when compared to a fixed 40% fan duty.