Integrated Real‐Time Mapping and Forecasting of Flood Inundation in Large‐Scale Digital Twins Basins Based on Physics‐Informed Graph Neural Network

This study investigated a novel approach that combined the advantages of physical mechanisms and data‐driven methods to achieve real‐time and accurate mapping and forecasting of flood inundation in the context of large‐scale digital twin basins, thereby better meeting the demands of flood control and disaster mitigation in river basins. Based on this framework, a physics‐informed graph neural network (PI‐GNN) was proposed. By incorporating the shallow water equations governing flow dynamics into the GNN as residual constraints, an encoder–decoder architecture integrating graph stacking and gated recurrent units was developed, enabling the effective embedding of physical conservation laws into the data‐driven learning process. Furthermore, a comprehensive digital twins basin platform was established by integrating multi‐source data assimilation mechanisms and three‐dimensional (3D) visualization technologies. Experimental validation in representative river basins demonstrated that the PI‐GNN model outperformed traditional methods in inundation mapping accuracy, achieving a critical success index of 0.89. In terms of forecasting performance, the root mean square error (RMSE) of water level prediction was reduced to 0.18 m for flood events with a 72‐h forecast horizon, while the end‐to‐end system response time was maintained within 89 s, satisfying real‐time operational requirements. The results indicated that the proposed approach exhibited improved physical consistency and forecasting accuracy, providing a practical and effective technical solution for enhancing flood prediction, simulation capabilities, and decision support for disaster prevention.