Improving 1-month forecasts in South Korea through the dynamical downscaling of machine learning based global predictions

Operational weather forecasts in South Korea currently extend up to 10 days but often fall short of adequately addressing the needs of weather-dependent sectors, such as agriculture, for longer-term meteorological predictions. However, extending forecasts beyond this timeframe remains a significant challenge. While traditional physical models have long been the foundation of weather forecasting, recent advancements in machine learning (ML) models for weather prediction have demonstrated promising forecasting skills that are comparable to, or even surpass, those of physical models. In particular, FuXi-ENS, an ML model trained on ECMWF ERA5 reanalysis data, provides global 6-hourly ensemble forecasts at a 0.25° resolution and shows great potential for one-month forecasts. To evaluate the forecasting performance of FuXi-ENS in South Korea and overcome its coarse spatial resolution, dynamical downscaling of multiple ensemble members is conducted using a regional climate model specifically tailored for Korea. For benchmarking purposes, dynamical downscaling of NOAA CFSv2 is also performed using the same regional climate model. Forecasting skill is comprehensively evaluated from both quantitative and qualitative analyses. Based on this comparative assessment, this study aims to provide valuable insights for enhancing subseasonal-to-seasonal forecasts in South Korea, offering practical benefits for various sectors reliant on extended-range forecasts.

 

Keywords: extended-range forecasting, dynamical downscaling, ML-based global predictions

 

(Acknowledgments) This study was supported by the “Research Program for Agricultural Science & Technology Development (Project No. RS-2024-00399847)”, National Institute of Agricultural Sciences, Rural Development Administration, Republic of Korea.