A hybrid AI-CFD model for predicting microclimate in PCM greenhouse

As a new agricultural facility, the precise prediction and control of microclimate in a phase change material (PCM) greenhouse is of great significance to crop growth. However, the strong coupling between their dynamic thermal performance and external meteorological conditions challenges the prediction method. A novel data-physics hybrid model combining the Informer method and computational fluid dynamics (CFD) is proposed, where the Informer method is used to predict the meteorological parameters, and the PCM greenhouse microclimate is simulated by CFD. The results show that compared with long short-term memory (LSTM), gated recurrent unit (GRU), and convolutional neural network (CNN) methods, the Informer method presents the best prediction effect. The mean absolute error of outdoor temperature, wind speed, direct normal radiation, and diffuse horizontal radiation is 0.2975 °C, 0.0289 m/s, 8.0787 W/m2, and 1.0074 W/m2, respectively. It is also shown that the prediction effect of the data-physics hybrid model is quite good on three month dataset, with all the R2 of air temperature, crop canopy temperature, soil temperature, PCM temperature, and liquid fraction being greater than 0.9990. Finally, the hybrid model is compared with the direct artificial intelligence prediction methods (LSTM, GRU, CNN, and Informer) on the one-week dataset for predicting the greenhouse microclimate. The results demonstrate that the hybrid model delivers better predictive performance for the greenhouse microclimate, with the R2 of air temperature, crop canopy temperature, soil temperature, PCM temperature, and liquid fraction all exceeding 0.9600. The proposed data-physics hybrid model can accurately predict the PCM greenhouse microclimate and improve the interpretation of physical mechanisms.