High-speed Search for Optimal Operation Parameters of Air Conditioning System in Data Center by Using Regression Prediction and Deep Reinforcement Learning on CFD Simulation

Demand for data centers has been increasing year by year, and so has the urgency to reduce their power consumption. In particular, air conditioning power consumption is still in strong demand for reduction. For this purpose, setting the optimum operation parameters of air conditioning is crucial for creating desired environmental state of data center. In this paper, we propose a procedure to search operation parameters such as operating conditions of air conditioning and transmittance of meshes installed in the aisles in a data center. Specifically, the proposed procedure consists of regression prediction by deep neural network (DNN) and search by means of deep reinforcement learning (DRL). Regression prediction by DNN using Computational Fluid Dynamics (CFD), based on the airflow distribution results of low mesh number, determines that of high mesh number in about 20 seconds. As a result, we established a search procedure for optimal operation parameters that realize the desired environmental state in less than 4 hours for a data center that accommodates around 500 servers. The demand of the server is also taken into account when dynamically controlling the operation parameters. These results indicate that the proposed procedure exhibits promising potential for a practical tool not only for designing a data center but also for efficiently managing environmental state control in case of module expansion and contraction, as well as real-time operation.