Model Predictive Control and Machine Learning for Greenhouse Energy and Crop Production Optimization

This study proposes a data-driven robust model predictive control framework to maximize both crop production and energy usage efficiency in the greenhouse. To develop the state-space model (SSM), we start by studying the dynamics of the greenhouse and crop based on the building geometry, weather dynamics, and input control signals. Additionally, we incorporate crop growth as one of the states controlled by meteorologic and agricultural disturbances. Afterward, we use data-driven techniques and system identification methods to construct the SSM. We further construct uncertainty sets to represent the uncertainty space on the basis of historical weather forecast error data, which are captured by a group of disjunctive uncertainty sets using a k-means clustering algorithm. Machine learning approaches are used to construct each basic uncertainty set and reduce the conservatism of resulting robust control action under disturbances. A robust MPC framework is further developed and applied to a case study, leading to an increase of 3.61 % to 7.32 % in power usage efficiency and an increase of 0.76 % to 16.25 % in crop production.