Skill of seasonal flow forecasts at catchment scale: an assessment across South Korea

Recent advancements in numerical weather predictions have improved forecasting performance at longer lead times. Seasonal weather forecasts, providing predictions of weather variables for the next several months, have gained significant attention from researchers due to their potential benefits for water resources management. Many efforts have been made to generate Seasonal Flow Forecasts (SFFs) by combining seasonal weather forecasts and hydrological models. However, producing SFFs with good skill at a finer catchment scale remains challenging, hindering their practical application and adoption by water managers. Consequently, water management decisions, both in South Korea and numerous other countries, continue to rely on worst-case scenarios and the conventional Ensemble Streamflow Prediction (ESP) method.
This study investigates the potential of SFFs in South Korea at the catchment scale, examining 12 reservoir catchments of varying sizes (ranging from 59 to 6648 km2) over the last decade (2011-2020). Seasonal weather forecasts data (including precipitation, temperature and evapotranspiration) from the European Centre for Medium-Range Weather Forecasts (ECMWF system5) is used to drive the Tank model (conceptual hydrological model) for generating the flow ensemble forecasts. We assess the contribution of each weather variable to the performance of flow forecasting by isolating individual variables. In addition, we quantitatively evaluate the overall skill of SFFs, representing the probability of outperforming the benchmark (ESP), using the Continuous Ranked Probability Skill Score (CRPSS). Our results highlight that precipitation is the most important variable in determining the performance of SFFs, and temperature also plays a key role during the dry season in snow-affected catchments. Given the coarse resolution of seasonal weather forecasts, a linear scaling method to adjust the forecasts is applied, and it is found that bias correction is highly effective in enhancing the overall skill. Furthermore, bias corrected SFFs have skill with respect to ESP up to 3 months ahead, this being particularly evident during abnormally dry years. To facilitate future applications in other regions, the code developed for this analysis has been made available as an open-source Python package.