Multi-step ahead streamflow and uncertainty forecasting using a HyMoLAP rainfall-runoff model-based framework integrated with Bayesian neural networks in the Ouémé river basin, Benin

Abstract

Multi-step forecasting is crucial for capturing future streamflow variations and managing water resources but remains challenging due to limited accuracy of upstream flow forecasts and meteorological predictions over lead times. While data-driven methods are commonly used, this study extends the Hydrological Model based on the Least Action Principle (HyMoLAP) from daily rainfall-runoff simulation to multi-day-ahead streamflow predictions. Additionally, it integrates Bayesian Long Short-Term Memory (Bayesian LSTM), primarily to enable uncertainty quantification (UQ). Applied to the Bonou and Savè sub-catchments of the Ouémé River Basin, Benin, the HyMoLAP-based framework yields NSE values ranging from 0.997 to 0.921 at Bonou and from 0.970 to 0.799 at Savè, showing slightly higher performance than the LSTM model overall, except at Savè from the 3-day lead time onward where it becomes slightly lower, with a more pronounced difference at the 7-day horizon. Our UQ approach provides reliable prediction intervals, with a coverage probability around 90%, as nearly 90% of the observed data fall within the 90% credible intervals in both sub-catchments.