基于BGM的不同繁殖长度对集合预报的影响

The Effect of Different Breeding Length upon Ensemble Forecasting Based on BGM

利用T63L9全球谱模式和NCEP/NCAR再分析资料,对BGM方法中增长模的繁殖长度对集合预报效果的影响进行研究。结果表明:与控制预报相比,不同繁殖长度的集合预报都能使预报效果得到一定程度的改进,特别是第4天预报以后,改进程度随预报时效而稳步提高。三组不同繁殖长度的集合预报对控制预报的改进存在差别,分析结果表明:繁殖长度为2 d的集合预报明显效果最差,而繁殖3 d和4 d的集合预报差别并不明显。对集合Ta-lagrand分布以及离散度的初步分析表明,繁殖长度取为3 d似乎最为合理。

Ensemble of initial conditions is the most important method in ensemble prediction, and how to generate the initial perturbations is crucial. Among all the methods to create perturbations for ensemble forecast, the ＂breeding of growing modes＂ （BGM） method has gained more and more favor for its good performance and almost ＂free＂ cost of computation. The breeding method simulates how fast-growing errors are ＂bred＂ in the analysis cycle. When a breeding cycle starts, an random initial perturbation field is added upon the control analysis. After some time＇s breeding, the most non-growing or decaying perturbations are filtered and the remainders can be mainly the fast-growing modes （perturbations）. Accordingly, the perturbation＇s growth rate reaches certain value and its shape is also getting to the phase of slow change. In some sense, the perturbation in the breeding cycle reaches ＂saturation＂. This saturated status of perturbation is regarded as the estimate of the fast-growing modes in the realistic analysis error, thus is applied as the initial perturbation. From the idea of BGM method, one of the crucial questions may be when the breeding cycle ends, which is called ＂the breeding length＂ problem. In fact, the breeding length is the total time of breeding of growing modes （initial perturbations） in the ensemble prediction system. The general steps to determine the breeding length are： defining the saturated character of growing modes; investigating the increasing and saturating process of the growing modes, and defining an approximate saturated time; according to the saturated time, doing a great lot of experiments with different breeding lengths, and determining the final length by the forecast skill. The model is the global media-range forecast spectral model T63L9 and the NCEP/NCAR 6-hour reanalysis data in 1998 are used. The initial modes in the beginning of a breeding cycle are random fields with unified distribution in [ - 1, 1 ]. The value of 6-hour is chosen as the length of the integration in each cycle. Besides, ＂rescaling＂ method to make the magnitude of growing modes comparable with the initial analysis error in RMSE sense is used. Some researches show that, if an appropriate breeding scheme is adopted, the breeding of the growing modes can present a clear saturation character in both the norm （magnitude） and the form （shape） after 3-4 days of breeding. Based on these results, numerical experiments are done with breeding length of 2, 3 and 4 days. Conclusions from these experiments show that the ensemble prediction with different breeding length can all reach a certain extent improvement upon the control forecast, especially the betterments increase steadily after 4 days of forecast lead time. Results also show the differences among the three breeding lengths in the improvement upon the control forecast. It shows a better improvement in ensemble mean with a breeding length of 3 days than2 days, but the 4-day breeding seems pretty much the same as the 3-day breeding. A pilot study on the spread and the Talagrand distribution of the ensemble members with different breeding lengths is also made. It seems that the most appropriate way is to choose 3-day as the breeding length in the system.