Interpretable modelling of greenhouse environment through SINDy approach

Modelling of greenhouse environment is crucial for maintaining the optimal environment for crop growth. Digital twins are widely used in greenhouse modeling due to their real-time reflection of reality. However, most existing Digital twins for greenhouse are based on data-driven models. This method is highly dependent on the dataset and lacks interpretability and generalizability. In this study, a knowledge-infused data-driven model was built by using sparse identification of nonlinear dynamical systems (SINDy). We collected 1927 sets of environmental data from a solar greenhouse in the cold region of China during the period from 22 February to 28 February 2024, and identified the predictive model from the measured data and physical knowledge. Comparing with the data-driven model, the SINDy model is more interpretable. Comparing with the physics-based model, the SINDy model is more concise and accurate, with the mean absolute percentage error (MAPE) of only 5.7% from the measured values. This has important implications for greenhouse control.