WMO Hydrological Status and Outlook System (HydroSOS): Approaches to blend and improve global multi-model streamflow forecasts

There is an immediate need to develop accurate and reliable global hydrological forecasts in light of the future vulnerability to hydrological hazards and water scarcity under a changing climate. As a part of the WMO Global Hydrological Status and Outlook System (HydroSOS) initiative, we investigated different approaches for blending multi-model streamflow forecasts, for developing holistic operational forecasts. The ULYSSES (mULti-model hYdrological SeaSonal prEdictionS system) delivers forecasts as a multi-model ensemble hydrological output from four state-of-the-art land surface and hydrological models appropriate for this study. The analysis was performed over 119 different catchments worldwide for the baseline period of 19812019 and, as a case study, a one year (1993) hindcast ensemble. The methodology tested blending approaches based on a performance metric based (weighted) averaging of the native multi-model streamflow, using streamflow catchment KGE as the performance metric. A simple (naive) multi-model averaging method was used to identify the added value of the weighted blended approach. Further, blending approaches were also tested for ungauged catchments by modelling the KGE metric based on catchment characteristics. The analysis included bias-correction of catchment streamflow before applying the two blending approaches (weighted and naïve). For the baseline analysis, bias-corrected weighted approaches provided the best improvement in performance for the catchments investigated. For the hindcast analysis, the performance is highly dependent on the catchment and the season. Further, modelling KGE metric for ungauged catchments showed promising results for providing weighted blended streamflow forecasts for ungauged catchments. These results indicate that there is potential to successfully implement the weighted blending approach, along with bias-correction, to improve operational streamflow forecasts. This work may contribute towards improving water resource management and hydrological hazard mitigation, especially in data-sparse regions.