Optimising Ensemble Streamflow Predictions with Bias-Correction and Data Assimilation Techniques

Abstract. This study evaluates the efficacy of bias-correction (BC) and data assimilation (DA) techniques in refining hydrological model predictions. Both approaches are routinely used to enhance hydrological forecasts, yet there have been no studies that have systematically compared their utility. We focus on the application of these techniques to improve operational river flow forecasts in a diverse dataset of 316 catchments in the UK, using the Ensemble Streamflow Prediction (ESP) method applied to the GR4J hydrological model. This framework is used in operational seasonal forecasting, providing a suitable testbed for method application. Assessing the impacts of these two approaches on model performance and forecast skill, we find that BC yields substantial and generalised improvements by rectifying errors post-simulation. Conversely, DA, adjusting model states at the start of the forecast period, provides more subtle enhancements, with the biggest effects seen at short lead times in catchments impacted by snow accumulation/melting processes in winter and spring, and catchments with high Base Flow Index (BFI) during summer months. The choice between BC and DA involves trade-offs, considering conceptual differences, computational demands, and uncertainty handling. Our findings emphasise the need for selective application based on specific scenarios and user requirements. This underscores the potential for developing a selective system (e.g., decision-tree) to refine forecasts effectively and deliver user-friendly hydrological predictions. While further work is required to enable implementation, this research contributes insights into the relative strengths and weaknesses of these forecast enhancement methods. These could find application in other forecasting systems, aiding the refinement of hydrological forecasts and meeting the demand for reliable information by end-users.