| 摘要: | 如何應用並擷取系集成員的預報結果是非常重要的。系集平均為普遍使用的統計方 法，但因雨量不連續性以及極值發生區域太局部，系集平均會將極值平滑掉導致雨量低估，本研究中使用兩個統計方式：PM與PMmod.來改善此缺陷。PM(Ebert, 2001)為使用系集平均的空間分布並經由重新分配所有系集之降雨頻率分布來改善QPF；另外，本研究自行提出新的系集QPF方法PMmod.，設計理念為一個好的系集預報理論上應可將真實的降雨極值包含在各系集成員的QPF極值中，這個方法基於PM的概念，使用系集平均之空間分布，但降雨頻率則是各成員降雨頻率之系集平均。
本研究使用兩個預報個案與2012年6月平均進行測試並分析中央氣象局(CWB)系集預報系統(WEPS)的系集平均、PM與PMmod.方法得到的QPF特性與技術得分，發現PMmod.或PM都可增加QPF技術，例如在50毫米，PM與PMmod. 皆可使官方TS進步率達11.5%，ETS進步率可達50%，但中小雨的預報技術仍是系集平均較好。在個案一的梅雨鋒面個案中，因系集模式對梅雨鋒面的速度與強度掌握有所偏差，使系集成員極值皆預報不足，因此即使使用PM仍不足預測極值。在個案二的蘇力(2013)颱風個案中，使用PMmod.或PM的前12小時QPF皆比系集平均佳，但後12小時PM的QPF明顯高估，主因為模式中的颱風移速較實際移速慢。另外在兩個個案中，無論是系集平均、PM或PMmod降雨分布皆有集中在山區的偏差或傾向，顯示WEPS對於台灣地形上的降雨偏差仍有進步的空間。
系集平均、PM或是PMmod.等後處理方式，可以有限度改善預報結果。但假使一開始模式預報就沒有掌握到實際降水的特徵，則這些後處理方式能幫助的其實非常有限，這就回溯於系集預報系統本身動力設定的改善。從前人研究可知，如果先調整原本的降水空間分布與系統誤差校正，再進行PM或PMmod.，可以取得更好的預報結果。; From ensemble members results to get useful information is an important issue for ensemble forecast system. Since rainfall distribution is not continuous, and too localized such as especially extreme heavy rainfall occur, the averaging process usually "smears" the rain rates so that the maximum rainfall is reduced and area of light rain is artificially enlarged. PM(probability-matched ensemble mean) is a new ensemble product, it has the similar spatial pattern as the simple ensemble mean, and could catch correct frequency distribution of rain rates and QPFs. Especially, we will propose a modify PM( PMmod.). It is similar as PM could display similar spatial pattern as the simple ensemble mean, its performance depends on the actual rainfall extremes should be included in the all members QPFs of a good Ensemble Prediction System (EPS). Therefore it is suitable to average all the rain rates frequency distribution from all the individual members.
From the 20120610-20120612 case (Case1), Typhoon Soulik(2013) case (Case2) and 201206 monthly experiment to analyze the characteristic and skill of the ensemble mean, PM and PMmod., we can find that they can modify the QPF, especially PM and PMmod. can make the TS progress rate up to 11.5%, ETS progress rate even up to 50% in the 50 mm(12 accumulated precipitation). But in the light rain, the ensemble mean is still the best. Nerveless, in Case1, the ensemble members don't catch the correct speed and strength of the Meiyu front, the rainfall for ensemble member distribution is generally underestimated, that is a reason why PM can't forecast the heavy rain correctly. In Case2, it has good performance for PM and PMmod results from the previous 12 hours rainfall forecast of typhoon Soulik, however in the later 12 hours rainfall forecast of the Soulik, the QPF is generally overestimate in all three statistical methods, because the simulated storm speed is slower than actual. Especiall, the spatial pattern of the ensemble mean, in both PM and PMmod concentrated in the mountains and result in bias. From this study, we can find that derive the correct distribution of rainfall result from the ensemble forecast system(WEPS) over the complex terrain in Taiwan has some progress but still need be investigated in the coming days.
If all ensemble members of forecast model can't catch the correct event signal, no matter what kind post processes method as simple ensemble mean, PM or PMmod could only modify the forecast results a little. Nevertheless, all the products like mean, previous study pointed out several include a resampling of the ensemble realizations, a rainfall pattern adjustment, and a bias-correction, could modify PM scheme to substantially reduce or eliminate the intrinsic model rainfall bias and to provide better QPFs. Expectations of the future, pattern adjustment and bias-correction both PM or PMmod processes to improve QPFs of ensemble forecast. |
