Urban Flooding Prediction Model Integrated with a Geospatial-Aware Attention Module

This paper proposes a novel, Transformer-based model for predicting urban-rain-induced flooding. The model leverages the remarkable performance of Transformers in time-series prediction tasks and innovatively integrates a geospatial-aware attention module to better capture the spatial heterogeneity and geographical features of urban flooding, thereby enhancing the model’s ability to learn spatial features, such as urban terrain. Experiments were conducted using precipitation data from June to August 2023 in Nanchang City and water-level data from urban flood monitoring stations. The results showed that the proposed method outperforms the original Transformer model for urban flooding prediction tasks while maintaining high prediction efficiency. The root mean square error of the proposed model is lower by 0.0006 compared to that of the original model, and the coefficient of determination is higher by 0.071. This indicates that the geospatial-aware attention module proposed herein can effectively improve the performance of time series models in urban flooding prediction tasks.