A Deep Reinforcement Learning Based Cooling Control Approach for Data Center

As a type of high energy consuming industry, the dramatic evolution of data centers requires escalating improvements in energy efficiency. With cooling systems accounting for a significant portion of data center energy consumption, it is imperative to implement effective control strategies to reduce energy consumption while maintaining temperature constraints for safe operation of IT equipmet. In this study, a simulation environment for a computer room is established by EnergyPlus. Based on this simulation environment, the Double Deep Q-Network (DDQN) algorithm is employed to control the cooling systems in data center. Incorporating temperature constraints of IT equipment and considering the impact of time-of-use electricity pricing, the objective is to minimize operating costs of the cooling system. The numerical tests demonstrate that, compared to the Rule-Based Control (RBC) method and the standard Deep Q-Network (DQN) algorithm, the DDQN algorithm achieves a more balanced control strategy. This strategy effectively reduces operating costs of the cooling system while ensuring compliance with IT operating temperature constraints, thereby showcasing superior optimization performance overall.