基于TIGGE的多中心集合平均预报理论与应用研究

发布时间：2018-01-13 12:09

本文关键词：基于TIGGE的多中心集合平均预报理论与应用研究　出处：《南京大学》2016年博士论文　论文类型：学位论文

【摘要】：作为提高业务数值预报水平的有效途径,集合预报方法得到广泛应用,并已成为数值预报体系中举足轻重的重要部分。而集合平均的预报技巧,则是衡量和评价集合预报系统的重要标准之一。本文对集合平均的预报技巧进行理论和实验研究,从理论角度回答了有关集合平均的若干问题,并针对基于TIGGE的单中心和多中心集合预报系统分别进行实验分析。理论研究表明,等权集合平均的合理性,首先在于当无法根据以往的表现断定各成员的相对优劣时,它不仅能够避免选用最差的成员,而且一定优于所有单个成员的平均水平。集合平均并非总是优于它的所有成员。只有当各成员的预报技巧差距不大、成员之间不存在明显的误差正相关时,集合平均才能优于它的所有成员,取得最佳的预报效果。集合平均的预报技巧不仅取决于单个成员的预报技巧,而且更主要地取决于各成员误差之间的相互关系。随着集合成员数的增加,在新成员的预报技巧与原有成员基本相当、新成员相互之间及其与原有的集合平均之间的误差相关性与原有集合成员相互之间的误差相关性基本相当的前提下,集合平均的均方误差下降并最终达到饱和。对于给定的饱和度,集合平均的预报技巧达到饱和所需要的成员个数,取决于集合成员均方误差的平均与集合成员相互之间的误差协方差的平均之间的比值。盲目地增加新成员不能起到改进集合平均的作用。新成员自身的预报技巧固然重要,但更重要的是新成员互相之间、及其与原有集合平均之间的误差相关性。将等权集合平均理论推广到一般的非等权集合平均。结果表明,对于理想状态下的最优权重平均,当集合成员无偏、独立时,总存在具有物理意义的最优权重集合平均,使之既优于最好的单个成员,也优于等权平均。预报的均方误差可以分解为预报误差的方差和系统误差的平方两个部分之和。本质上,集合平均的目的是为了减小预报误差的方差,而偏差订正则是为了消除系统误差。将两者结合起来,可以得到更为理想的预报效果。通过将训练期设为预报期,并在此基础上构造带有偏差订正的最优权重集合平均,即可得出带有偏差订正的非等权集合平均的均方误差的下界。这在一定程度上体现了集合平均的可预报性。将上述理论应用到T213集合预报系统,15个成员的等权集合平均在绝大多数情况下优于控制预报,特别是在中期预报中集合平均的优势更为明显。对于T213集合预报系统,15个成员已基本实现95%的饱和度,意味着在不改变控制预报水平和扰动成员生成方式的前提下,即使再增加新成员,集合平均的均方误差最多只可能再降低5%。对于T213集合预报系统,等权集合平均的预报技巧已接近恒权集合平均的理论极限。同时,尽管在中期预报阶段,变权集合平均有可能取得比等权平均更高的技巧,但由于各成员预报技巧的相对排名极不稳定,变权集合平均在业务应用中仍存在较大困难,也难以取得更高的预报技巧。因此,对于T213集合预报系统,等权平均仍是集合预报的合理选择。权重的改进意义不大,只有控制预报本身的改进、以及扰动成员生成方式的改进,才是进一步改进T213集合平均的关键。对于基于TIGGE的多中心控制预报集合,结果表明,对于风、温、压、湿等气象要素,ECMWF和NCEP各有优势,而CMA的预报技巧则明显落后于其它两个中心。在绝大多数情况下,三个中心的等权集合平均优于其最好的单中心。在湿度场的预报和其它要素场的中期预报中,三个中心彼此之间的误差相关性较弱,因此集合平均的优势更为突出。尽管CMA的预报技巧明显落后于ECMWF和NCEP,但对于湿度场预报和其它要素场的8-10天预报,三个中心的平均明显优于ECMWF和NCEP两个中心的平均,说明CMA或其它预报技巧较低的成员的加入对于有效改进TIGGE多中心集合平均仍有重要意义；而这种改进的来源,则在于多中心成员之间较低的误差相关性。等权集合平均是基于TIGGE的多中心集合平均的合理方法。三个中心的等权集合平均的预报技巧,已接近带有偏差订正的变权最优权重集合平均的理论极限。
[Abstract]:As an effective way to improve the service level of the numerical prediction, ensemble prediction method has been widely used, and has become an important part of an important numerical prediction system. The ensemble average skill, is one of the important criteria to measure and evaluate the ensemble prediction system. In this paper, the theoretical and experimental research on ensemble average skills, from theory the perspective of problems related to the average collection, and according to a single center and multi center based on TIGGE ensemble prediction system are analyzed and experimental analysis. The theory research shows that reasonable average weighted aggregation, first of all is that when not according to the relative merits of past performance that members, it can not only avoid the worst members the average and is superior to all individual members. The ensemble average is not always better than that of all its members. Only when the members of the report The skills gap between members and there is no obvious positive correlation error, ensemble average ability is better than all its members, has forecast effect. The best ensemble average skill depends not only on the individual members of the forecast skill, and mainly depends on the relationship between the members of the errors. With the increase of the number of ensemble members the forecast skill of new members, and the original members of the basic premise of the new members, the correlation between the error between the original and the ensemble average and the original members of the collection of error correlation between basic phase when under the set average MSE decline and eventually reached saturation. For a given saturation, average ensemble prediction techniques the number of members required to reach saturation, the error covariance between the members of the collection depends on the mean square error of the ensemble flat The ratio between the both. Blindly add new members can not play a role. The improved ensemble mean forecast skill of the new members of their own is important, but more important is the new members to each other, and the correlation between the error of the original set. The average weight of ensemble average theory is extended to non general ensemble average. The results show that the optimal weight average under ideal condition, when the members of a collection of unbiased, independent, there is always the optimal weights with the physical meaning of ensemble averaging, which is superior to the individual members of the best, is better than the right. The average mean square error of prediction can be decomposed variance and the system error of forecasting error square two parts. In essence, the purpose is to set the average prediction error variance decreases, and the deviation is set regular in order to eliminate system error. Combining the two, can get more satisfactory The result of forecast. Through training period for the forecast period, and on the basis of constructing the optimal weights with bias correction set average, lower bounds can be obtained with non bias correction of ensemble mean square error of the average. This reflects the average predictability set to a certain extent. The application of the above theory the T213 ensemble prediction system, the 15 members of the ensemble average in most cases is better than that of the control forecast, especially in the Medium-term Forecast ensemble average of the more obvious advantages. The T213 ensemble prediction system, 15 members have basically achieved 95% saturation, means without changing the control level and disturbance prediction the members of generation, even add a new member of the ensemble average mean square error is only possible to reduce the 5%. for the T213 ensemble prediction system, such as the right set of forecast skill average is close to constant weight set The theory of limit average. At the same time, although in the Medium-term Forecast stage, variable weight set average may have higher average power than other skills, but because the members of the relative ranking of the forecast skill is very unstable, variable weight set average in business applications still exist great difficulties, it is difficult to obtain a higher forecast skill therefore, the T213 ensemble prediction system, a reasonable choice is still the average ensemble. Little improvement significance of weight, only control the prediction itself is improved, and the improved perturbation members generation, is to further improve the T213 ensemble average. The key for TIGGE based multi center control results show that the prediction set. The wind, temperature, pressure, humidity and other meteorological elements, ECMWF and NCEP have their own advantages, and the forecast skills of CMA is obviously lagged behind the other two. In most cases, three heart right set average Better than the best single center. In the humidity field prediction and other elements of market forecasting error, weak correlation between the three centers of each other, so the ensemble average advantage is more prominent. Although the CMA forecast skill is behind ECMWF and NCEP, but for the humidity field prediction and other elements of the 8-10 day forecast, the average was significantly higher than that of ECMWF and NCEP at three centers in two centers, which show that the addition of CMA or other members of the lower forecast skill for the effective improvement of TIGGE multi center ensemble average is still of significance; and the improved source is the error correlation between multi center members of lower right. The collection is a collection of TIGGE multi center average reasonable method based on average. Three of ensemble forecast skill average, is close to the optimal weights with variable weight deviation correction set theoretical limit of average.

【学位授予单位】：南京大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：P456.7

【相似文献】