东亚区域人工智能气象大模型预报技巧评估

Evaluation of Weather Forecasts from AI Big Models over East Asia

针对人工智能气象大模型的500 hPa位势高度、2 m气温、10 m风速、降水以及热带气旋路径等,从定性和定量两个角度进行评估。结果表明:从定性角度出发,FuXi、Pangu和GraphCast 3个大模型均会响应热带异常加热,其中Pangu与GraphCast响应强度接近,FuXi响应较弱。从定量角度出发,FuXi整体展现出更高的预报能力,其最大可用预报日数超过9.75 d,Pangu和GraphCast分别为8.75 d和8.5 d。在2 m气温预报中,FuXi的时间异常相关系数为0.48~0.91,Pangu和GraphCast分别为0.43~0.91和0.38~0.83。此外,采用TS(threat score)评分对FuXi和GraphCast降水预报进行评估,FuXi在晴雨、小雨和中雨预报中更具优势,其预报技巧分别为0.22~0.41、0.15~0.24和0.06~0.22,GraphCast在大雨预报中展现更强能力。针对2019年7月29日华北暴雨和2020年8月16—17日乐山暴雨两次极端降水个例进行分析,FuXi和GraphCast均可提前8 d预报降水的空间分布,但在降水量级预报中存在偏差,随着预报时效减小,偏差也逐渐减小。在热带气旋路径预报中,Pangu展现更高精度。

Reliable medium-range weather forecasts are crucial for both science and society.Although weather predictions primarily rely on numerical weather models,the artificial intelligence(AI)weather big models have shown potential for accurate weather forecasts with reduced computational costs.However,prediction skills of big models remain uncertain,particularly in East Asia,which limits further application of weather AI models.To systematically evaluate predictive capabilities of Pangu,FuXi,and GraphCast models over East Asia,their prediction results are focusing on 500 hPa geopotential height,2 m air temperature,10 m wind speed,precipitation,and track of tropical cyclones.ECMWF reanalysis V5(ERA5)datasets are utilized to provide the initial conditions for big models,and to assess their predictive skill.Additionally,precipitation observations and China Meteorological Administration tropical cyclone datasets are utilized to access big models as well.FuXi shows the highest forecasting skills among 3 big models for 500 hPa geopotential height.The forecast from FuXi is reliable for up to 9.75 days,while the forecasts from Pangu and GraphCast are reliable for 8.75 days and 8.5 days,respectively.For 2 m air temperature forecasting,FuXi presents higher skills with an averaged temporal anomaly correlation coefficient(TCC)ranging from 0.48 to 0.91,while TCCs of Pangu and GraphCast are 0.43-0.91 and 0.38-0.83,respectively.Among 3 models,only FuXi and GraphCast provide precipitation forecasts.FuXi shows higher prediction skill compared to GraphCast in forecasting precipitation,light rain,and moderate rain;however,GraphCast has advantage in heavy rain forecast.As the lead time increases,the threat scores(TSs)of FuXi for rainfall,light rainfall and moderate rainfall are 0.22-0.41,0.15-0.24 and 0.06-0.22,respectively.The model demonstrates higher skill in the northern and southeastern regions of China.For predicting the track of cyclones,Pangu model demonstrates superior predictive skill.As the lead time increases from 6 hours to 240 hours,biases of Pangu’s prediction track increase from 17.5 km to 1850 km.The study focuses on the prediction skill of various AI big models through TCC,spatial anomaly correlation coefficient,and TS.Generally,the performance of FuXi is superior for most elements.And reasonable evaluation of AI model is helpful for the development of AI models.