The urgent need for sustainable intensification of food production, driven by a growing global population and increasing climate variability, has positioned greenhouse cultivation and advanced climate control as critical areas of innovation in agriculture. Greenhouses enable precise environmental management, significantly boosting crop yields. However, realizing these benefits requires climate and crop models that accurately represent greenhouse dynamics while remaining interpretable and suitable for real-time control.
This thesis addresses greenhouse climate prediction by developing a hybrid modeling approach combining the strengths of physics-based system identification with modern machine learning techniques. In collaboration with Hoogendoorn Growth Management, a hybrid model was developed utilizing a physics-informed Sparse Identification of Nonlinear Dynamics (SINDy) method to capture primary mechanisms influencing temperature and humidity, augmented by an Long Short-Term Memory (LSTM) neural network to account for residual, unmodeled effects. A key contribution of this research is demonstrating the feasibility of transfer learning, successfully adapting a model trained initially on simulation data to real-world greenhouse scenarios, yielding reliable predictions for both air temperature and humidity under operational conditions.
Empirical results validate that transfer learning effectively bridges simulation and practi cal greenhouse environments, underscoring the practicality of data-driven climate models for industry applications. Integrating the model into an Model Predictive Control (MPC) framework illustrates its operational viability; the controller accurately tracks temperature setpoints but experiences challenges maintaining humidity within desired limits. These findings emphasize the necessity for accurate predictive models and underscore the importance of carefully formulating MPC strategies. Further research into modeling humidity dynamics, expanding the model’s state space, and refining targeted retraining methodologies is essential to ensure robust, year-round practical deployment.
Overall, this thesis advances greenhouse climate modeling and control by showcasing how hybrid modeling combined with transfer learning can effectively close the gap between simulation based development and operational implementation, thereby laying the groundwork for adaptive and efficient greenhouse management practices