Varying the Combination of Hydrological Models in Time and Space: Toward a More Accurate Representation of Streamflow in Large-Sample Hydrology

Accurate predictions of streamflow are needed to manage water resources, evaluate flooding risks, and support agriculture and industry. Many hydrological studies rely on a single model structure and parameter set applied uniformly across space and held fixed over time, limiting their ability to represent spatiotemporal changes in hydrological conditions. This study evaluates a model-agnostic, interpretable, computationally frugal approach to dynamically combine multiple models to improve streamflow simulation. The Framework for Understanding Structural Errors (FUSE) was used to create an ensemble of 78 hydrological models applied to 559 catchments across the contiguous United States. Each model was calibrated to maximize either high-flow or low-flow performance, yielding 156 simulations per catchment. The dynamic combination assigns weights to ensemble members that vary in space and time. It identifies past conditions similar to the current state, ranks simulations by performance under the identified comparable past conditions, and forms a combined simulation by equally weighting the m best-ranked models at each time step. The results show that the dynamic combination approach mitigates trade-offs between competing objectives (e.g., low flows vs. high flows) compared to static approaches. This method also reduces sampling uncertainty, thereby increasing confidence and robustness in streamflow simulations. Nevertheless, the dynamic combination also has limitations; for example, the current implementation cannot predict values outside the ensemble prediction envelope. Future work could integrate machine learning into the combination step to enable extrapolation. In addition, the method could be applied to forecasting and ungauged catchments.