Energy-efficient AI-based Control of Semi-closed Greenhouses Leveraging Robust Optimization in Deep Reinforcement Learning

As greenhouses are being widely adopted worldwide, it is important to improve the energy efficiency of the control systems while accurately regulating their indoor climate to realize sustainable agricultural practices for food production. In this work, we propose an artificial intelligence (AI)-based control framework that combines deep reinforcement learning techniques to generate insights into greenhouse operation combined with robust optimization to produce energy-efficient controls by hedging against associated uncertainties. The proposed control strategy is capable of learning from historical greenhouse climate trajectories while adapting to current climatic conditions and disturbances like time-varying crop growth and outdoor weather. We evaluate the performance of the proposed AI-based control strategy against state-of-the-art model-based and model-free approaches like certainty-equivalent model predictive control, robust model predictive control (RMPC), and deep deterministic policy gradient. Based on the computational results obtained for the tomato crop's greenhouse climate control case study, the proposed control technique demonstrates a significant reduction in energy consumption of 57% over traditional control techniques. The AI-based control framework also produces robust controls that are not overly conservative, with an improvement in deviation from setpoints of over 26.8% as compared to the baseline control approach RMPC.