Toward Data-Enabled Economic Predictive Control for Controlled Environment Agriculture

Abstract

Agricultural production is under increasing pressure from population growth and climate change, making efficient and high-yield farming essential. Greenhouse cultivation offers a controlled environment to meet these demands. This work investigates Data-Enabled Predictive Control (DeePC), a model-free method, for optimizing lettuce yield in a simulated greenhouse. We benchmark its performance against a nonlinear Model Predictive Control (NMPC) approach, which relies on a detailed physics-based model. Both controllers were tasked with maximizing final yield over a 40-day growth cycle under winter and summer weather conditions. In the winter scenario, which matched its training data, DeePC achieved 94% of NMPC's final biomass (125.81 g/m² vs. 133.30 g/m²) and a slightly lower economic return (2.164 vs. 2.516 Hfl/m²). In addition, DeePC was five times more computationally efficient. However, when the winter-trained DeePC was applied to the summer scenario, it frequently violated critical temperature constraints, despite achieving a high yield. These findings demonstrate that DeePC is a promising alternative for economic greenhouse control, but its reliability and safety depend heavily on the representativeness of its training data.