The 24th IAHR-APD Congress (24th IAHR-APD Congress)

407 / 2024-03-14 09:18:47

Evaluation of the Pangu Weather Forecasting Model in simulating Precipitation and Atmosphere Circulation over the Yangtze River Basin, China

Pangu Weather Forecasting Model,Precipitation,Yangtze River Basin

Subtheme 1: Water-related Hazard and Risk Reduction > Topic 1: Water Hazard Under Climate Change

In order to aid management personnel, forecasters, and relevant researchers in understanding the forecasting capabilities of the Pangu Weather Forecasting model in the Yangtze River Basin and its environs, this study systematically evaluates precipitation and the atmospheric circulation over the Yangtze River Basin. Five typical precipitation cases occurring between August 22 and October 7, 2023 were selected for hindcast verification. The Pangu model, driven by ERA5 (ECMWF Reanalysis v5) data, utilized a greedy algorithm to perform forecasting with a 6-hour step for 170-hour (7 days). Pangu-generated precipitation output is compared against data from automatic weather stations.  Additionally, all instances underwent testing using ERA5 reanalysis data and GDAS/FNL analysis data, serving as upper air and ground truth data from ECMWF (European Centre for Medium-Range Weather Forecasts) and NCEP (National Centers for Environmental Prediction), respectively. Quantitative evaluation metrics including RMSE (Root Mean Square Error), ACC (Anomaly Correlation Coefficient), and PCC (Pattern Correlation Coefficient), alongside the object-oriented forecast verification SAL (Structure, Amplitude, and Location) method, are utilized to assess atmospheric circulation and precipitation forecast outcomes respectively. In comparison with the “observations”, the evaluation indicates that hindcast with Pangu model capture both the large- and regional-scale circulation features over the the Yangtze River Basin. For example, the normalized RMSE of geopotential at any pressure level are less than 0.4 within the 100-hour forecast period. It was found that for the precipitation events on August 26 and October 1 exclusively, Pangu's simulation performance in terms of heavy rainfall distribution surpasses that of ECMWF. Conversely, the reverse is observed for torrential rain. Furthermore, it was found that, to achieve an atmospheric circulation field representative of the Yangtze River Basin while conserving computational resources to the maximum extent, the model domain should encompass the region between 9°-51°N and 69-141°E, at minimum. These findings provide a foundation for determining the regional scope of artificial intelligence meteorological models in the Yangtze River Basin.