我国地面降水的分级回归统计降尺度预报研究

利用TIGGE资料中欧洲中期天气预报中心(ECMWF,the European Centre for Medium-Range Weather Forecasts)、日本气象厅(JMA,the Japan Meteorological Agency)、美国国家环境预报中心(NCEP,the National Centers for Environmental Prediction)以及英国气象局(UKMO,the UK Met Office)4个中心1~7 d预报的日降水量集合预报资料,并以中国降水融合产品作为"观测值",对我国地面降水量预报进行统计降尺度处理。采用空间滑动窗口增加中雨和大雨雨量样本,建立分级雨量的回归方程,并与未分级雨量的统计降尺度预报进行对比。结果表明,对于不同模式、不同预报时效以及不同降水量级,统计降尺度的预报技巧改进程度不尽相同。统计降尺度的预报技巧依赖于模式本身的预报效果。相比雨量未分级回归,雨量分级回归的统计降尺度预报与观测值的距平相关系数更高,均方根误差更小,不同量级降水的ETS评分明显提高。对雨量分级回归统计降尺度预报结果进行二次订正,可大大减少小雨的空报。

Statistical downscaling of precipitation forecasting using categorized rainfall regression

High-resolution weather forecasting is a growing societal demand.However,the limited spatial resolution of existing models still cannot meet such a demand,so downscaling is widely applied.There are two types of downscaling:dynamical downscaling and statistical downscaling.A large computing cost is required by dynamical downscaling,and statistical downscaling is generally more acceptable because of its relative simplicity and practicability,along with its many flexible methods.More accurate forecast results can be obtained by the statistical downscaling method of establishing the function between the low-resolution raw model output and the high-resolution predicting variables.In addition,rainfall data are discontinuous and follow a non-normal distribution.So,it is important to establish a statistical downscaling model suitable for daily precipitation.Based on the ensemble forecasts of 1-7-day daily accumulated precipitation from the ECMWF,JMA,NCEP and UKMO in the TIGGE datasets,as well as an hourly merged precipitation product over China as the observed data,a forecasting study on daily precipitation over China by means of statistical downscaling was conducted.Firstly,a spatial sliding window was used to increase moderate and heavy rainfall samples.Then,the statistical downscaling technique was used to improve the precipitation forecast by constructing different regression equations based on different categories of rainfall.The results show that statistical downscaling is more effective in increasing the anomaly correlation coefficient(ACC) and the equitable threat score(ETS),and decreasing the RMSE,as compared to the bilinear interpolation method,because the observed data are added to the function to correct the statistical downscaling model.The improvement in the forecast through statistical downscaling differs among models,lead times,and rainfall levels,and depends upon the forecasting ability of the particular model.The forecasting ability of heavy rain via the statistical downscaling approach of constructing a single equation is poor-even inferior to the bilinear method.However,the forecast results after the categorized regression are more accurate than those obtained via direct regression,because the former can substantially improve the forecasting ability of different threshold values and the whole area,as reflected in the following aspects:The ACC of the categorized regression at the 168-h lead time is greater than 0.6-even larger than the ACC of direct regression at the 24-h lead time.In addition,the increasing amplitude of the ACC of the categorized regression method increases with lead time.The RMSE of the precipitation forecast increases with lead time,and the error of the categorized regression method is only 9.5 mm·d^-1 at the 168-h lead time-much smaller than the uncategorized regression method.However,forecast data with a larger ACC do not always yield a smaller RMSE,because the RMSE also depends on the magnitude of rainfall.The ETS of different threshold values of the categorized regression are larger than those produced via direct regression.The increasing amplitude of the ETS when using categorized regression decreases with the magnitude of rainfall.The ETS of less than 10 mm rainfall using categorized regression increases significantly because samples of light rain are ample.In short,the categorized rainfall regression method is a more reasonable technique for high-resolution weather forecasting.Further correction to categorized-regression downscaling forecasts of precipitation may reduce the occurrence of false alarms considerably.