Optimizing Warnings for Hydrologic Ensemble Prediction Systems for Improved Decision-Making

Abstract Flood early warning systems (FEWSs) rely on hydrological simulations driven by numerical weather prediction (NWP) models, both of which are inherently uncertain. Ensemble prediction systems (EPSs) address these uncertainties by generating multiple future scenarios. Discharge forecasts from EPS are converted into flood warnings by applying a set of criteria whose definition is crucial for the skill of the system. While meteorological and hydrological models are under continuous development, these warning criteria lack continuous evaluation. In this paper, we use discharge simulations from the European Flood Awareness System (EFAS) to assess the skill of four NWPs—probabilistic and deterministic—and explore methods to create a grand ensemble that enhances overall skill. We examine the effects of the current warning criteria—probability threshold and persistence—optimize their values, and evaluate their effectiveness in smaller catchments. Our results indicate that probabilistic NWPs outperform deterministic models in flood warning skill and demonstrate that removing the persistence criterion from EPS enhances skill. Optimized warning criteria for a single probabilistic NWP boost EFAS skill by 6.6% in terms of f score, and a grand ensemble can further increase it by 4.1%. We identify two effective grand ensemble methods—member based and skill based—and discuss their advantages and drawbacks. The improved criteria demonstrate comparable skill in catchments half the size of those currently used. In conclusion, our study presents a methodology for evaluating and refining the skill of multimodal FEWS, stressing the critical role of tailoring the flood warning criteria to the end users. Significance Statement This study analyzes methodologies to improve the skill of the notifications issued by a state-of-the-art flood warning system. We have analyzed the skill of four meteorological models and the effects of several warning criteria. We have identified an optimal blend of meteorological models and warning criteria that boosts the system skill up to 10%. The methods developed in this research are adaptable to other flood warning systems, and the refined criteria have been implemented operationally.