In response to the issues of single simulation results and lack of interpretability when using data-driven models to replace traditional hydrodynamic models for rapid forecasting, an LSTM model for joint simulation of multiple factors was constructed based on a multi-task learning framework, and the fitted physical relationships among factors were incorporated into the loss function as constraint terms. Furthermore, the interpretability of the model was explored by combining the comprehensive feature evaluation results of random forest, Spearman rank correlation coefficient method, and principal component analysis (PCA). Based on a database of simulation results of 17 flood events in Tunxi District by a two-dimensional hydrodynamic numerical model, the LSTM models with and without physical constraints were trained, respectively. The results indicate that the computational speed of the LSTM model with physical constraints does not decrease significantly, which can realize multi-factor simulation within 1.s. For the overall simulation of multiple flood events, the LSTM model with physical constraints performs better, and the simulation accuracies of water level, discharge, flow velocity, and inundation area are improved by 20.38%, 17.03%, 4.02%, and 14.07%, respectively. The LSTM model with physical constraints demonstrates a stronger capability in capturing small floods and higher accuracy in simulating regular floods, but its performance still needs to be improved under extremely complex conditions.