One-kilometer resolution forecasts of hourly precipitation over China using machine learning models

Abstract

Global numerical weather prediction (NWP) models often face challenges in providing the fine spatial resolution required for accurate prediction of localized phenomena and extreme precipitation events due to computational constraints and the chaotic nature of atmospheric dynamics. Downscaling models address this limitation by refining forecasts to higher resolutions for specific regions. Recently, machine learning (ML) based weather forecasting models demonstrate superior efficiency and accuracy compared to traditional NWP models. However, these ML models generally operate with a temporal resolution of 6 h and a spatial resolution of 0.25°. Furthermore, they predominantly rely on the fifth generation of the European Center for Medium-Range Weather Forecasts Reanalysis (ERA5) data, which is notorious for its precipitation biases. In this study, we utilize the High-Resolution China Meteorological Administration Land Data Assimilation System dataset, which provides more accurate precipitation data, as the target for downscaling and bias correction. This study pioneers the application of a transformer-based super-resolution model, SwinIR, to downscale and correct biases in precipitation forecasts generated by FuXi-2.0, a state-of-the-art ML weather forecasting model trained on ERA5 with a temporal resolution of 1 h. Our results demonstrate that the downscaled forecasts outperform the high-resolution forecasts from the ECMWF in terms of both accuracy and computational efficiency. However, the study also underscores the persistent challenge of underestimating high-intensity rainfall and extreme weather events, which remain critical areas for future improvement.