Physics-Informed CNN-LSTM Framework: An Approach for Temperature Prediction

CNN-LSTM performs well in meteorological forecasting by integrating spatial feature extraction and time series modeling. However, the training data distribution easily restricts the pure data-driven model, and the lack of physical constraints will lead to a physical solution. In this work, we propose a PINN-CNN-LSTM model, which embeds atmospheric thermodynamic equation constraints on the basis of CNN-LSTM and jointly optimizes data fitting and physical consistency through the PINN loss function. Experimental results show that the proposed method is significantly better than CNN, LSTM, BiLSTM, and CNN-LSTM. The MSE, RMSE, and MAE of the PINN-CNN-LSTM model are 1.4321, 1.1967, and 0.8777, respectively. Compared with the CNN-LSTM model, it has reduced by 6.4%, 3.2%, and 2.4%.