Operational flood forecasting with adaptive dual-correction bayesian model averaging in the Huai river basin

Accurate real-time flood forecasting is critical for basin-scale flood management, yet hydraulic models often suffer from parameter and structural uncertainties, particularly during high-flow events. This study proposes an Adaptive Dual-Stage Bayesian Model Averaging (ADS-BMA) framework to improve operational flood predictions. The method integrates a HEC-RAS ensemble with a two-stage residual correction strategy, combining long short-term memory learning for dynamic errors and linear adjustment for systematic bias. Bayesian optimization is used to stabilize parameter updates and reduce manual tuning. Application to the Huai River Basin using multi-year observations demonstrates clear benefits for streamflow and water-level prediction. Compared with the standard BMA, ADS-BMA improves Nash-Sutcliffe efficiency by 8% for streamflow and 50% for water level, reduces RMSE by 14–18%, and narrows prediction intervals while maintaining good calibration. Results show that the framework effectively enhances peak-flow forecasting skill and supports more confident early-warning and decision-making in flood-prone river systems.