This study analyzes the causes of storage-tank pressure fluctuations in HVAC system based on real plant operational data, and proposes a reinforcement-learning-based parallel compressor control strategy to mitigate such fluctuations. For policy learni...
This study analyzes the causes of storage-tank pressure fluctuations in HVAC system based on real plant operational data, and proposes a reinforcement-learning-based parallel compressor control strategy to mitigate such fluctuations. For policy learning, we adopt a model-free reinforcement learning approach with TD3 that can directly optimize control policies from data. Training was first conducted using the May operational dataset, which represents high-disturbance conditions, so that the agent could learn a stabilization policy under severe pressure fluctuations. The learning environment was then switched to the February dataset to verify adaptability under distribution shifts caused by seasonal and operating-mode changes. Generalization performance was evaluated on the April dataset, which was not used for training. As a result, on the test period the tank-pressure error was reduced to an MAE of 3.44 kPa and the peak-to-peak variation within 100 s windows was reduced to 12.5 kPa, demonstrating that the RL-based policy can contribute to tank-pressure stabilization even in environments with frequent operator interventions and operating-mode changes.