A teacher-student reinforcement learning framework for temperature control in solar greenhouses

Solar greenhouses are a cornerstone of modern agriculture in northern China, which are passive structures relying solely on solar heating and ventilation systems. The highly fluctuating winter temperatures, however, lead to suboptimal control, which impairs crop growth and increases thermal energy loss. Effectively managing these issues often involves balancing multiple conflicting goals, such as optimizing crop growth while minimizing energy use. Therefore, an advanced control system that can efficiently optimize multiple objectives and enhance ventilation strategies under extreme climate conditions is essential. Here, we propose a control strategy that integrates model predictive control (MPC) with proximal policy optimization (PPO) within a teacher-student framework (MPC-PPO). In this framework, MPC generated high-quality control experience to guide the PPO agent's learning process. An adaptive dynamic weighting mechanism was employed to balance the influence of MPC experiences during PPO training. Evaluation experiments conducted in a simulated environment using data from solar greenhouses across three provinces in northern China (Beijing, Hebei, and Shandong) demonstrated that: (1) the MPC-PPO method achieved the highest temperature control performance with a 5.46-point improvement in temperature control compared to the PPO alone; (2) the MPC-PPO method achieved a ventilation control reward of 99.19±0.54, optimizing ventilation window operations with intelligent time-differentiated strategies that reduced energy loss during non-optimal hours; (3) feature analysis revealed that historical window opening, air temperature, and historical temperature were the most influential features for effective control; and (4) cross-regional tests confirmed that MPC-PPO consistently outperformed other methods across all tested regions.