Adaptive Cooling System Control in Data Center with Reinforcement Learning

Data center cooling system is consuming large amounts of power, which requires effective control to reduce operational costs and deliver optimal server performance. The high power consumption occurs because traditional cooling methods struggle to adapt dynamically to workloads, causing wasteful power consumption. Therefore, this study aimed to explore the use of machine learning methods to improve energy efficiency for data center cooling system. For the experiment, an RL (Reinforcement Learning) model was designed to adjust cooling parameters with dynamic environmental changes. The method focused on optimizing energy efficiency while maintaining stable temperature and humidity control. By applying RL-based control method to PAC system, this study contributed original results that validated the effectiveness of RL-simulated data center environments. Specifically, the stages included developing system model, creating simulations using the PAC control system, and training an RL model with environmental conditions. Data were collected from simulations and analyzed to test the model performance, and the outcomes were presented using a real-time monitoring interface with Flask. The results showed that the RL model achieved an average reward of 4.76 (between -5 and 5), a convergence rate 13.2, a sampling efficiency 10.15, and a stability score 2.6. The model effectively reduced temperature and increased humidity during stressed data center operations. When compared with a fixed cooling system, RL showed superior adaptability to workload variations and reduced unnecessary energy consumption. However, scalability to real data center remained an issue, which required more than simulation validation. In conclusion, the RL-based method optimized efficiency of cooling system, showing the potential to improve energy savings and operational resilience in data center environments.