Data-driven forecasts of extreme weather in East Asia: feasibility of operational use

Abstract

Accurate medium-range weather forecasts are essential for disaster preparedness and resource management. They are traditionally achieved by numerical weather prediction (NWP) models, but such models are costly to run operationally. Recent advances in machine-learning-based weather prediction (MLWP) offer a promising alternative to NWP for improving forecast efficiency. This study evaluates the prediction skills of four state-ofthe-art MLWP models-FengWu, FuXi, GraphCast, and PanguWeather-and their ensemble mean (MLWPEM), initialized with operational analysis from Korea Integrated Model (KIM) as well as ERA5 reanalysis. An emphasis is placed on forecasts of extreme events over East Asia-such as heavy rainfall, typhoons, heat waves, and cold spells in 2023-up to 10 days ahead. When initialized with operational KIM analysis, MLWP forecasts exhibit performance comparable to, or better than, operational KIM forecasts in predicting 500-hPa geopotential and 2-m air temperature. Even for rainfall forecasts, MLWP models-particularly GraphCast-outperform operational KIM forecasts at short lead times (1-3 days), but their performance decreases markedly for intense rainfall events (>= 30 mm/6 h). Forecast skills are further improved when considering MLWP-EM. Case studies show that KIM-initialized MLWP-EM forecasts outperform operational KIM forecasts in predicting temperature extremes, but underestimate rainfall amounts during heavy rainfall events and typhoons. Overall forecast skills increase when MLWP models are initialized with ERA5 reanalysis, highlighting the importance of initial conditions even in MLWP. These results demonstrate the operational potential of MLWP-EM forecasts initialized with operational analysis for medium-range forecast, while underscoring persistent challenges in accurately representing extreme precipitation.