Simulation of Urban Flood Process Based on a Hybrid LSTM-SWMM Model

This study proposes a novel hybrid LSTM-SWMM model that integrates the advantages of the SWMM model and the LSTM neural network for the first time. The aim is to build an efficient and rapid model that considers the physical mechanism, in order to effectively simulate urban floods. The results indicate a good agreement between the simulated discharge process of the LSTM-SWMM model and the observed discharge process during the training and testing periods, reflecting the actual rainfall runoff process. The $${R}^{2}$$ R 2 of the LSTM-SWMM model is 0.969, while the $${R}^{2}$$ R 2 of the LSTM model is 0.954. Additionally, for a forecasting period of 1, the $$NSE$$ NSE value of the LSTM-SWMM model is 0.967, representing the highest forecasting accuracy. However, for a forecasting period of 6, the $$NSE$$ NSE value of the LSTM-SWMM model decreases to 0.939, indicating lower accuracy. As the forecasting period increases, the $$NSE$$ NSE values consistently decrease, leading to a gradual decrease in accuracy.