基于WRF模式构建集合更新预报系统,利用Haar小波分解方法分析了北京"7·21"特大暴雨过程中三种初始扰动方案所构造集合扰动的多尺度特征,基于此探讨了混合初始扰动方法的可行性,并对比了三种扰动对误差的模拟能力。其中扰动方案一是由集合转换卡尔曼滤波方法(ETKF)对NCEP全球集合预报的分析扰动更新后得到,扰动方案二(DOWN)是直接由NCEP全球集合预报扰动插值到试验所设置的模式网格中得到,而扰动方案三(BLEND)则是将上述二者通过Barnes滤波进行尺度混合后得到。结果表明:各组扰动的能量均随时间增长,其中包含分析不确定性的ETKF扰动在预报中前期有较高的中小尺度能量,而DOWN扰动有较高的大尺度能量且能量增长速度明显快于ETKF,二者能量的大值区最终都向中尺度(64~128 km)部分发展,混合后的扰动(BLEND)能量在预报中前期增长速度最快,综合表现最优。从扰动成分来看,ETKF和DOWN中在预报前期可以快速增长的部分均集中在8~32 km的小尺度上,64~128 km部分的中小尺度的扰动信息增长缓慢,而256 km的中尺度信息则很快被耗散,这为如何选取合理的滤波波段构造多尺度混合扰动提供了依据。从降水预报结果来看,控制预报误差主要集中在降水的大值区,虚假初始扰动会导致预报初期产生虚假降水区;在暖区降水阶段,扰动对误差的模拟能力较弱,而在锋面降水阶段,扰动对误差的模拟能力明显提高,总体来看大尺度的误差较难模拟,三种方案中BLEND对误差的模拟能力最强;根据扰动-误差的相关分析同样验证了BLEND在误差模拟能力方面的优势;在降水预报TS评分方面,各组集合试验均优于控制试验,其中BLEND的效果略优于ETKF和DOWN。
A storm-scale ensemble was conducted by WRF model during Beijing"7·21"extreme precipitation event. Three initial perturbation methods is tested. The first one is produced by ETKF update and forecast cycle which contained analysis uncertainty. The second method(DOWN) is downscaled from NCEP global forecast perturbation,and the third one is produced by blending ETKF and DOWN using barnes filter with wavelength of180 km(80~ 280 km). Results show that each perturbation energy can grow with time,in which ETKF has more medium and small scale energy due to flow-dependent analysis uncertainty and DOWN has more large scale energy in early time. BLEND has the most perturbation energy during most forecast time. Energy from each perturbation all grow to medium scale(64~128 km) and the fastest growing composition are focused on small scale at early forecast hours,while the medium scale component grow slowly. These results motivate further studies on how to choose properly wavelength to construct a blending initial perturbation. When coming to the error of precipitation,spurious perturbation may lead to small spurious precipitation in early hours. Forecast perturbations for different methods all have better performance in sampling error during front precipitation than warm area precipitation.All in all,ETKF has advantage in small scale and lead time error sample and DOWN is better at large scale in the later forecast time,BLEND has both advantages of ETKF and DOWN during the whole forecast time. The threat score also show that BLEND has the best overall performance.