In this paper, a hybrid LSTM-CNN model is constructed to realize the dynamic simulation of urban flooding, which is realized by multi-layer perception of neurocognitive machine and feature recognition through convolutional computation. The model network identifies the site rain feature variables that may produce dangerous waterlogged areas by learning local spatial features for identifying the magnitude of surface generated runoff under different rainfall events and their impact on the city. The common urban flooding scenarios caused by river and heavy rainfall and other disaster-causing factors are simulated as a source of urban flooding data by fully considering the hydrological characteristics of the city and following the principles of hydrology-hydrodynamics. The normalization method was adopted to pre-process the data, the entropy method was used to determine the weights of the indicators, the effective threshold of water depth was defined, and the mean absolute error (MAE), root mean square error (RMSE), and Nash efficiency coefficient (NSE) were selected as the evaluation indicators. Taking Henan Province as the study area for empirical analysis, the LSTM-CNN hybrid model achieves a forecasting accuracy of 92.5% in the flood peak flow and peak present time, and realizes the spatial-temporal evolution analysis of the mildly, moderately, and severely affected populations in the dynamic simulation of the population risk in the whole process of urban rainstorms and floods. Based on the LSTM-CNN hybrid model can sense urban flooding in real time and provide an effective tool for risk assessment and emergency management.
