The enhanced integration of proven techniques to quantify the uncertainty of forecasting extreme flood events based on numerical weather prediction models

Skillful forecasting of reservoir inflow is one of the main prerequisites for determining reservoir operation and management policies. This research incorporates proven techniques in a novel way to develop a comprehensive framework for forecasting event-based inflow floods with sub-daily time steps (6-h intervals), considering the uncertainty of Numerical Weather Prediction (NWP) models. Accordingly, raw precipitation forecasts were extracted for six extreme flood events in the Dez River basin, Iran. A Multi-Model Ensemble (MME) system was developed using the Group Method of Data Handling (GMDH) and Weighted Average-Weighted Least Square Regression (WA-WLSR) models to post-process raw precipitation forecasts. We thereupon proposed an approach that combined the Hydrologiska Byråns Vattenbalansavdelning (HBV) hydrological model with the Long-Short Term Memory (LSTM) network (HBV-LSTM model) to enhance flood forecasting. Moreover, a comparative analysis was performed between the modeling approaches, i.e., probabilistic inflow forecasting and deterministic inflow forecasting. The results revealed that the forecasting skill of the MME model built using the WA-WLSR model was higher than that of the GMDH model. Accordingly, the highest Continuous Ranked Probability Skill Scores (CRPSS) of 0.61 and 0.67 were achieved by the GMDH and WA-WLSR models, respectively, based on a precipitation threshold of 10 mm. Additionally, both the HBV-LSTM model and the LSTM network outperformed the individual HBV model in producing inflow flood hydrographs. Based on the best flood forecasting approach, i.e., the HBV-LSTM model, the NSE exceeded 0.95, and the NRMSE remained below 0.09 for various flood events. The outcomes indicated a variability of 2–10 % in the relative peak error using the HBV-LSTM approach for different flood events. Our findings provide valuable insights for determining the key elements of reservoir operations and enhancing management strategies under flood conditions. • Development of the novel framework based on hybrid hydrological data-driven approach for short-term flood forecasting. • Post-processing ensemble precipitation forecasts without the prerequisite of extensive long-term meteorological data. • Comparative analysis of reservoir inflow flood forecasting using the conceptual HBV model, LSTM network, and the hybrid HBV-LSTM approach.