A dynamic urban flood risk prediction framework for real-time flood risk early warning

Short-term urban flood early warning requires not only rapid prediction of future inundation depth fields, but also timely identification of high-risk areas. To address the high computational cost of high-resolution two-dimensional hydrodynamic models and their limited ability to support real-time risk updating, this study proposes a History-Driven Source-Node Multi-Step Depth Prediction Network (HDMP-Net). The model formulates urban flood prediction as a heterogeneous mapping from overflow source terms to surface-grid responses. It integrates historical surface water depths, dynamic rainfall-overflow source terms, and source-node spatial-structural priors, and predicts continuous inundation depth fields over the next 1 h through a source-node structural attention mechanism and multi-step incremental decoding. On this basis, the predicted water depth is further used as a time-varying hazard factor and integrated with indicators including population density, building density, road density, pipeline density, elevation, and slope to establish a dynamic flood risk assessment procedure. The results show that HDMP-Net achieves high accuracy in inundation depth prediction across three independent test events, with MAE values of 0.0194-0.0225 m and RMSE values of 0.0431-0.0465 m, without evident systematic bias. The predicted results effectively preserve the consistency of inundation extent, deep-water zones, and water-depth hydrographs at key overflow points. The risk estimates derived from the predicted water depths are highly consistent with the reference risk results, with risk-level accuracy ranging from 0.8885 to 0.9650 and recall for high and very-high risk zones ranging from 0.9465 to 0.9814. In terms of computational efficiency, HDMP-Net reduces the time required to generate inundation depth fields for a complete event from approximately 149 min with the two-dimensional hydrodynamic model to 9 min, corresponding to a reduction of 93.97%. For a single short-term update, it can generate future inundation depth fields and complete risk mapping within seconds. These findings indicate that HDMP-Net provides a rapid and reliable technical basis for short-term urban inundation depth prediction, flood risk zoning, and high-risk area identification.