Prediction of greenhouse temperature and humidity across growing seasons: Hybridization of process-based model and deep neural networks

• Proposed a hybrid approach combining process-based models with deep neural networks. • Seasonal predictions of air temperature and relative humidity dynamics with high accuracy and robustness are performed. • This study presents validation results from a continuous 80-day operation of the model, demonstrating its performance with dual data inputs: ground-level weather station measurements and forecast API data. Accurate prediction and precise management of greenhouse environments contribute to improving crop yield and quality. However, conventional physics-based models exhibit inherent recalibration constraints due to their fixed parameterization schemes, particularly when confronting nonlinear dynamics in cross-seasonal greenhouse environments. This study presents a novel hybrid mechanistic model capable of dynamic parameter calibration for multi-season applications. These dynamic parameters in the mechanistic model were adaptively predicted by deep neural networks. The greenhouse model calibration results showed that the errors for indoor air temperature and relative humidity were RMSE = 1.6104 °C, 6.9379 %, MAE = 1.0463 °C, 4.3797 %, R 2 = 0.9090, 0.8290, and PBIAS = −1.3807 %, −0.0005 %, respectively. Finally, in the absence of indoor environmental reference data, the hybrid greenhouse model with adaptive parameters was validated from September 5, 2024, to November 25, 2024. The validation involves two approaches: one using measured weather data as inputs, and the other utilizing a public weather forecast API. The validation results indicate robust model performance regardless of the input data source, including both measured weather data and weather forecast data. This study develops a robust tool for forecasting greenhouse microclimates throughout growing seasons, thereby enabling optimized environmental management.