Dynamic Combination of a Multi-Model Ensemble for Improved Streamflow Simulations

Accurate streamflow simulations are needed to manage water resources, evaluate flooding risks, and support agriculture and industry. Traditional ensemble approaches are usually based on meteorological ensemble but rarely consider hydrological ensemble. However, hydrological forecasts based on a single model often fail to capture the dynamic nature of hydrological systems. Addressing this gap, we present a novel dynamic combination method that adaptively leverages hydrological ensemble diversity to enhance streamflow simulations.

Using the Framework for Understanding Structural Errors (FUSE), we generated 78 hydrological models applied to 559 catchments from the CAMELS dataset across the contiguous United States. Each model was calibrated to optimize both high-flow and low-flow performance, producing a hydrological ensemble of 156 members per catchment. Our dynamic combination approach can be divided in two parts: a conceptual k-nearest neighbor algorithm to identify similar historical conditions and then model predictions at the time step of interest are weighted based on their performance for the k-nearest neighbors.

Results demonstrate that this dynamic combination approach improves upon traditional static methods, particularly in representing diverse streamflow conditions. The method captures temporal variability, reduces trade-offs among objective functions, and provides a model-agnostic framework for enhanced streamflow simulations. While the approach shows promising results, it faces limitations in its reliance on hydrological ensemble and meteorological data quality. Future work could explore machine learning integration for dynamic combination and applications to real-time forecasting and ungauged catchments.