Behavior cloning-guided multi-agent deep reinforcement learning for greenhouse-microgrid control

Greenhouse cultivation requires precise climate control and efficient energy management to maintain crop yields. However, this control task is challenging because climate variables are highly coupled and influenced by outdoor weather conditions. Although model predictive controllers have been applied, their performance relies strongly on system model accuracy. This study proposes a behavior cloning-guided multi-agent deep reinforcement learning (MADRL) framework for coordinated control of a smart greenhouse integrated with a renewable energy microgrid. The control problem is formulated as a multi-agent system in which each agent undertakes a specific function for climate control and energy storage. First, the agent policies are pretrained via behavior cloning of data generated by a nonlinear model predictive controller that ensures safe learning and accelerates convergence. Subsequently, a sequential multi-phase training strategy is employed to reduce non-stationarity during training. The control framework was validated using a physics-based simulator for tomato cultivation. The simulation results demonstrated improved tracking performance for lighting and relative humidity, while maintaining temperature and CO₂ regulation comparable to that of baseline controllers. In addition, a reduction of approximately 3–5 W m−2 in equivalent electrical power demand across different tomato growth stages was achieved. These findings demonstrate the potential of MADRL framework for coordinated climate control and energy management in greenhouse integrated microgrids under varying environmental conditions.