Evaluation of precipitation forecasting base on GraphCast over mainland China

The accurate cumulative precipitation forecasts are essential for monitoring water resources and natural disasters. Thecombination of deep learning and big data has become a new direction for precipitation forecasting. However, the currentlarge models are still lacking in-situ data verification. To accomplish this goal, the precipitation forecasting performance of astate-of-the-art model GraphCast was evaluated. Using the cumulative precipitation data from 2393 observation stations for the1-3 day period as a reference, we assessed the cumulative precipitation in mainland China region for the 1-3 day period from2020 to 2021, utilizing a high-resolution model with 0.25◦×0.25◦ grid spacing and 37 layers of parameters. The precipitation ofEuropean Centre for Medium-Range Weather Forecasts (ECMWF) was also compared. The results show that: (1) During the2020-2021 period, for the 1-day, 2-day, and 3-day cumulative precipitation forecasts, the Root Mean Square Error (RMSE)values of GraphCast were primarily between 0.46 to 9.38 mm/d, 0.44 to 9.06 mm/d, and 0.44 to 9.06 mm/d, respectively. TheMean Error (ME) values were mainly between −0.595 to 1.705 (0.01 mm). (2) As the forecast period extends, the forecastingcapability of GraphCast declines. (3) In the 1-3 day cumulative precipitation forecasts for various stations in mainland China,GraphCast demonstrates higher predictive accuracy than ECMWF. (4) Compared to ECMWF, GraphCast demonstrated thebest forecast performance in the warm-temperate humid and sub-humid north China, with the RMSE being approximately 12%higher. Our study indicates that GraphCast demonstrates significant potential and higher accuracy in precipitation forecasting.