Field evaluation of a soft actor–critic-based control strategy for a centralized geothermal heat pump system

Reinforcement learning (RL) has attracted increasing attention for building energy system control. However, its real-world application remains limited owing to sensor constraints, legacy hardware, and operational uncertainties. This study presents the development and field evaluation of a Soft Actor–Critic (SAC)-based optimal control strategy for a centralized ground-source heat pump cooling system in an existing commercial building. The proposed controller optimizes the number of operating heat pumps and chilled-water supply temperature by considering the outdoor conditions, ground-source temperature, buffer-tank temperature, and Time-of-Use electricity pricing without relying on real-time indoor temperature measurements or hardware modifications. A physics-informed simulation environment was constructed using simplified thermal models of the building, buffer tank, and geothermal loop. The SAC policy was trained using historical summer weather data representative of the study site. The trained controller was then deployed in a real building and evaluated through field A/B testing conducted over multiple months against a conventional rule-based control strategy. Fair comparison was ensured by normalizing weather-induced performance variations using the outdoor air enthalpy. Field results demonstrated that the SAC-based control reduced electricity consumption by 11.75% and electricity costs by 13.27% compared with the rule-based baseline, while maintaining comparable cooling energy delivery. The results confirm that reinforcement learning-based control can be effectively implemented in legacy ground-source heat pump systems and can deliver tangible energy and cost benefits under real operating conditions.