ILMWR: A 3-Stage Physics-Guided Machine Learning and 2D HLLC Hydrodynamic Framework for Real-Time Flood Inundation Forecasting

This paper presents ILMWR, a three-stage operational forecasting framework for real-time flood inundation. Its contribution is the design and integration of the pipeline itself – coupling a physics-based rainfall–runoff and open-channel hydraulics model, a two-dimensional depth-averaged shallow-water (Saint-Venant) solver discretized with an HLLC Riemann solver under Strang operator splitting, and a physics-guided data-driven multi-task LSTM with temporal attention – rather than any single new numerical or learning technique. The hydrological stage combines a TCVN-style runoff coefficient with the SCS Curve Number method through a sigmoid ensemble blend and recovers station-level stage via a safeguarded Newton–Raphson inversion of Manning’s equation. The hydrodynamic stage propagates inundation on a structured grid with adaptive CFL time-stepping and its numerical consistency is verified through a twin experiment using NSE, RMSE, F1, and AUC skill scores. The learning stage forecasts flood occurrence, peak discharge, and river stage jointly at 6, 12, and 24-hour horizons, with deep ensemble uncertainty quantification and an event-skill threshold selected via Youden’s J statistic.