In order to improve the computational efficiency of urban waterlogging simulation and meet the requirement of timeliness of urban flooding warning, the advantage of deep learning method and simulation results of the traditional hydrological-hydrodynamic model were utilized, and the key driving factors for the development of urban rainstorm waterlogging disaster chain, including slope, elevation, and precipitation, were used as inputs. In combination with the convolutional neural network(CNN) and the long shortterm memory network (LSTM), a CNNLSTMATT model for urban waterlogging water depth prediction was constructed through introducing the attention mechanism (ATT)and by using the sparrow search algorithm for optimal selection of hyperparameters. The model was used to predict the waterlogging water depth in Dakonggang new urban area of Shenzhen City, and the results show that the CNNLSTMATT model can effectively predict the water depth of urban waterlogging caused by heavy rain. The predicted water depth with the CNNLSTMATT model in the next 30min is close to the simulation results of the hydrologicalhydrodynamic model, and the accuracy of the model decreases slightly after the prediction period is extended. Compared to the hydrologicalhydrodynamic model, the CNNLSTMATT model exhibits a simulation efficiency that is approximately 200 times higher.