在強烈非線性動力發展下，因系集分布可能嚴重偏離高斯分布，使得系集平均未必為真實大氣之最佳估計。因此本研究提出系集重新定位法(MRS)來解決這個問題，並以2011年南瑪都 (Nanmadol) 颱風個案為例，驗證系集重新定位法能否有效改善颱風的路徑預報。 本研究首先透過TIGGE (THe Observation system Research and Predictability Experiment) 的歐洲氣象中心 (EC) 全球系集預報資料，針對南瑪都颱風初期路徑的預報不確定性進行分析。結果顯示，在EC的系集成員中，預報表現較好的系集其颱風強度均較強、移動速度較快、颱風較大。而南瑪都颱風與塔拉斯颱風之間的鞍型場結構、南瑪都颱風自身外圍環流以及與塔拉斯颱風間的相互作用都是造成南瑪都颱風突然西偏的敏感區域。 接著將系集重新定位法應用於區域系集預報系統，並分析系集重新定位法對於系集預報的影響。此外，再進一步將此方法應用於系集資料同化系統上，探討能否藉由系集資料同化系統的流場相依 (flow-dependent) 特性使重新定位法的效益正回饋至同化系統中，進而改善系集整體的動力發展及後續預報能力。 結果顯示，在單純的系集預報系統實驗中，系集重新定位法可成功改善颱風路徑的系集預報，使系集離散度能合理表達預報不確定性。然而，依據動力系統非線性程度的不同，如何選取作為最佳初始場之系集成員對系集重新定位法而言相當重要。而在系集資料同化系統實驗中，因颱風同化初期洋面觀測資料不足與初始颱風結構不完整，原本的系集資料同化系統無法有效地建立合理的背景誤差結構。但透過系集重新定位法，整體系集得以合理發展，進而獲得較正確的背景誤差結構。如此不僅能改善同化結果，也將對整體預報產生正回饋。此外，本研究建議，挑選數個較好的系集成員，以其平均場當作最佳初始場，對系集的改善將更為顯著。 Ensemble mean is used to represent the best estimation of nature state and the deviation from the mean are used to approximate the forecast uncertainties. However, when dealing with the strong nonlinear dynamics, the ensemble mean and its evolution may not be as reliable as we hope. Thus, we try to solve this problem through the mean recentering scheme (MRS) with the typhoon Nanmadol (2011) case study. This research consists of two parts. In the first part, we examine the track forecast uncertainties of Typhoon Nanmadol (2011) by the EC global ensemble prediction system from TIGGE database. The result from Good Group and Poor Group comparison indicates that the TCs in Good Group are stronger, moving faster and with larger TC size. In the sensitivity region experiment, we found that the saddle field between Typhoon Nanmadol and Typhoon Talas, the outer circulation of Typhoon Nanmadol and the interaction between Typhoon Nanmadol and Typhoon Talas are the sensitivity region for Typhoon Nanmadol’s sudden recurvature. The second part is to examine the ability of mean recentering scheme with ensemble forecast system (EPS) and ensemble data assimilation system (EDAS). For the EPS, we hope MRS is able to improve the initial distribution of ensemble members so that the ensemble might keep the suitable distribution during its evolution. With the flow-dependent characteristics, we expect the EDAS could retain the improved ensemble information to ameliorate the ensemble dynamic evolution and prediction skill. The results of EPS experiment show that MRS is capable of improving the TC track forecast and its ensemble spread can represent the forecast uncertainties reasonable. However, how to select the best member is critical for MRS. In the EDAS experiment, due to the lack of observation and incomplete TC structure of initial filed, it is extremely difficult to establish an accurate background error covariance in time. Nevertheless, with the MRS, we can estimate a more reasonable background error covariance and the better information would be carried by EDAS to the next analysis time to improve entire EDAS. Also, using the top members’ average as the best member provides a better result.
