Analyzing the Mitigation Effect of Urban River Channel Flood Diversion on Waterlogging Disasters Based on Deep Learning

Abstract

In recent years, urban waterlogging disasters have become increasingly prominent. Physically based urban waterlogging simulation models require considerable computational time. Therefore, rapid and accurate simulation and prediction of urban pluvial floods are important for disaster prevention and mitigation. For this purpose, we explored an urban waterlogging prediction method based on a long short-term memory neural network model that integrates an attention mechanism and a 1D convolutional neural network (1DCNN–LSTM–Attention), using the diversion of the Jinshui River in Zhengzhou, China, as a case study. In this method, the 1DCNN is responsible for extracting features from monitoring data, the LSTM is capable of learning from time-series data more effectively, and the Attention mechanism highlights the impact of features on input effectiveness. The results indicated the following: (1) The urban waterlogging rapid prediction model exhibited good accuracy. The Pearson correlation coefficient exceeded 0.95. It was 50–100 times faster than the InfoWorks ICM model. (2) Diversion pipelines can meet the design flood standard of a 200-year return period, aligning with the expected engineering objectives. (3) River channel diversion significantly reduced the extent of inundation. Under the 30-year return period rainfall scenario, the maximum inundation area decreased by 1.46 km2, approximately equivalent to 205 international standard soccer fields.