Offline reinforcement learning for optimal control of a greenhouse with ground source heat pump

An offline reinforcement-learning supervisory on/off controller for greenhouse heating was developed using TRNSYS–Python co-simulation of a ground-source heat pump with thermal energy storage and benchmarked against a fixed-threshold rule-based baseline. A conservative Q-learning policy was trained from an offline dataset generated under diverse behavior policies and evaluated at a 10 min control interval from December 1 to February 1. Under the primary Yesan weather scenario in South Korea with a 15 °C indoor setpoint, the learned policy maintained near-baseline temperature regulation while reducing heat-pump electricity use by 9–38% per week, averaging 17% over nine weeks, and improving the seasonal average COP from 3.09 to 3.23. Across regional tests, electricity consumption decreased by 19% in Cheorwon, 17% in Yesan, and 23% in Jeju, consistent with longer on/off intervals and storage-aware scheduling that suppresses unnecessary cycling. This study is limited by the supervisory binary action space and the winter-focused evaluation period. Future work will extend the framework to richer control actions and broader seasonal operation and will incorporate additional greenhouse constraints such as humidity and condensation risk.