ECMWF集合预报产品在广西暴雨预报中的释用  

Application of ECMWF Ensemble Forecast Products to Rainstorm Forecast in Guangxi

作　　者：赵华生[1] 黄小燕[1] 黄颖[1] Zhao Huasheng;Huang Xiaoyan;Huang Ying(Guangxi Research Institute of Meteorological Disasters Mitigation, Nanning 53002)

机构地区：[1]广西壮族自治区气象减灾研究所,南宁530022

出　　处：《应用气象学报》2018年第3期344-353,共10页Journal of Applied Meteorological Science

基　　金：国家自然科学基金项目(41575051;61562008;41765002);广西重点基金项目(2017GXNSFDA198030);广西青年基金项目(2014GXNSFBA118211)

摘　　要：基于最大相关最小冗余度算法和随机森林回归算法,该文提出一种对欧洲中期天气预报中心(ECMWF)集合预报产品进行暴雨预报的释用方法。该方法采用最大相关最小冗余度算法,对ECMWF集合预报的51个成员进行筛选,选取若干个与预报对象相关性最大、相互间冗余度最小的成员作为随机森林回归算法的输入因子。利用ECMWF集合预报降水量平均值对建模样本进行分类,使预报模型的建模样本更具有针对性。通过2012年4月-2015年12月的交叉独立样本试验预报和2016年1-9月的业务预报试验的统计结果表明:该释用方法的暴雨预报TS和ETS评分,均比采用ECMWF集合预报产品51个成员降水量预报进行插值后取平均值的释用方法分别提高了0.07和0.05以上,显示了较好的数值预报产品释用效果。Using the maximal correlation minimum redundancy algorithm and random forest regression algorithm, a rainstorm interpretation forecasting method with numerical prediction products is proposed based on the ensemble prediction system of European Center for Medium-Range Weather Forecasts（ECMWF）.The precipitation forecast of 51 members in ECMWF ensemble prediction system are interpolated to weather stations, and then, the maximum related minimum redundancy algorithm is used to filter ensemble members. Finally, several member interpolations that have the highest correlation with the predictand and the least redundancy with each other are selected as input factors of the random forest regression algorithm. Furthemore, in order to make modeling samples of the forecast model more pertinence, the modeling samples are classified using the mean rainfall value of ECMWF ensemble prediction products of 51 members. That is, when the mean precipitation using ECMWF ensemble prediction products at a certain station is relatively large and there is a possibility of precipitation above the storm level, only historical samples containing a large amount of precipitation are selected as modeling samples of the forecasting model. Therefore, the forecasting model reduces the influence of the sunny and wet weather samples on the noise of the forecasting model, so that forecasting model focuses on the training of large precipitation samples. When the mean value of the predicted ECMWF ensemble precipitation at a certain weather station is small, all samples of the weather station（including samples of sunny days and heavy precipitation） are modeled so that the training of the forecasting model can reconcile the heavy rain samples and thus as far as possible to avoid the rainstorm of weather station omissions reported. This method is applied to 89 stations in Guangxi, and a 4-year cross-independent sample test forecast for 2012-2015 is carried out. The business test forecast is carried out in 2016. In the 4-year cross-independent samp

关 键 词：最大相关最小冗余度算法 随机森林回归算法 释用 

分 类 号：P457.6[天文地球—大气科学及气象学]