Multi-agent reinforcement learning for chiller system prediction and energy-saving optimization in semiconductor manufacturing

Energy consumption in cooling systems is one of the major environmental burdens in semiconductor manufacturing. Energy-saving measures not only help reduce energy costs but also effectively decrease carbon emissions. These improvements enhance the operational efficiency of the entire supply chain and ultimately benefit downstream enterprises, thereby promoting the sustainable development of the semiconductor supply chain. This study aims to optimize the energy savings in chiller systems in the semiconductor manufacturing. We investigate the interactions between various devices and show how the chiller's operational status affects the temperature setpoint. This study proposes a meta-prediction model to simulate the dynamic behavior of the chiller system, and also employ multi-agent reinforcement learning to support the multi-setpoint control for energy optimization. An empirical study of a semiconductor manufacturer in Taiwan was conducted to validate the proposed model. The results indicate that our developed solution successfully reduced the kilowatts per refrigerated ton (KW/RT) by approximately 2.78% in a practical application.