A ConvLSTM-Based Model for Urban Flood Prediction Under Dynamic Rainfall Patterns and Exploration on Its Extrapolation Capability

Abstract Urban flooding induced by heavy rainfall is increasingly frequent, necessitating accurate and timely flood forecasting to mitigate risks. Although data-driven models have demonstrated significant potential for real-time flood prediction due to their computational efficiency, current implementations frequently neglect the critical influence of rainfall spatial heterogeneity, resulting in inaccuracies in flood prediction. Therefore, this study designed diverse rainfall scenarios featuring moving rainstorm centers and proposed a fast simulation method for urban flooding under complex rainfall conditions, utilizing the convolutional long short-term memory (ConvLSTM) model. The efficacy of the proposed method was validated across three study areas. The results indicate that the ConvLSTM model has superior performance in predicting flood inundation depth and extent, achieving an average R 2 of 0.964, outperforming two other deep learning models. Notably, this model achieved predictions within seconds based on input rainfall data, offering high computational efficiency that is hundreds of times faster than hydrological–hydrodynamic coupled models. Furthermore, we explored the model’s extrapolation capability when rainfall intensities exceed the maximum value of the training set. This research contributes insights to the advancement and refinement of rapid urban flood forecasting methodologies.