Multi-Agent Deep Reinforcement Learning Based Adaptive Control for Smart Greenhouse Integrated Microgrid

This paper proposes a multi-agent deep reinforcement learning (MADRL) framework for adaptive control of a smart greenhouse integrated into a renewable-based microgrid. The system jointly regulates temperature, humidity, CO2 concentration, lighting, water pumping, and battery storage to reduce the electric grid dependence. To enhance training efficiency, the actor networks of Twin Delayed Deep Deterministic Policy Gradient (TD3) agents are pretrained using imitation learning on datasets generated from nonlinear model predictive control (NMPC). Simulation results across multiple tomato growth stages show that MADRL achieves performance comparable to NMPC in microclimate and energy regulation while operating almost an order of magnitude faster in online computation. Moreover, under uncertain weather forecasts with noise and bias, MADRL outperforms NMPC in both tracking accuracy and energy efficiency, demonstrating strong robustness to forecast errors and adaptability to uncertainty. These findings highlight the potential of MADRL as a computationally efficient and resilient alternative for greenhouse-microgrid operation, paving the way for real-time applications and future extensions to profit-driven, year-round control.