| 摘要: | 由於颱風等劇烈天氣的可預報度很低，因而發展可信的短時系集預報系統(ensemble prediction system, EPS)是相當重要的。本研究主要是利用局地分析和預報系統(Local Analysis and Prediction System, LAPS)發展時間延遲之多模式系集預報系統(time-lagged multimodel EPS)，並以此產生短時(0-6小時)的定量降水機率預報(probabilistic quantitative precipitation forecasts, PQPFs)。最終目的在於針對颱風，提供使用者更具參考價值的系集降水預報，並幫助使用者藉由系集機率預報(ensemble probabilistic forecasts, EPFs)做出最好的決策。
校驗結果顯示LAPS EPS具有良好的系集離散度-預報能力關係(spread-skill relationship)以及相當的區辨能力(discrimination ability)。因此，即便LAPS PQPFs有明顯的濕偏差，仍然可以透過線性迴歸(linear regression, LR)方法校正預報偏差而大大提升預報能力。我們針對影響校正結果的兩個重要因子進行敏感度測試，包括：(1)不同訓練樣本之測試；(2)觀測資料正確性不一致之測試。測試(1)的結果顯示：校正結果對於訓練樣本非常敏感。訓練樣本必須和驗證樣本有一致(或非常相似)的預報偏差分布，才能得到好的校正結果。測試(2)的結果顯示：由於海洋與陸地的觀測資料正確性不一致，加上絕大部份的觀測是由海洋樣本所貢獻；因此，海洋和陸地必須分區校正才能確保兩者都有較好的校正結果。
然而，在評估天氣預報系統的可用性時，不僅要考慮預報品質的良窳，同時也要考量使用者在日常生活中，以此預報資訊做決策所能得到的經濟效益。透過相對操作特徵(relative operating characteristic, ROC)面積量測的區辨能力是評估預報品質的重要指標之一，並且和預報系統的經濟價值(economic value, EV)有密切的關係。因此，本研究的另外兩個重點分別是探討ROC和EV之間一致和相關的特性，並說明使用者如何利用LAPS PQPFs做決策，以得到最大的經濟效益。比較系集機率預報和單一預報(deterministic forecasts, DFs)的ROC和EV，可以發現：系集機率預報具有較高的區辨能力，且可提供較大的經濟價值給較多的使用者；這樣的優勢隨著降水事件的強度增加而愈加顯著。關於校正的敏感度實驗顯示：預報系統的區辨能力或是所能提供的最大經濟價值(maximum EV, EVmax) 對於預報偏差並不敏感。由於中央山脈對於颱風降水的地形鎖定效應，LAPS PQPFs在山區比在平地有較佳的區辨能力和較高的可信度，並可提供較大的經濟價值給成本損失比(cost-loss ratio)較高的使用者。此外，我們發現：山區和平地使用者的經濟價值差異，主要是預報系統在兩區域有不同的區辨能力所造成。我們同時舉例說明，即便使用者的成本損失比無法被明確得知，仍然可以利用系集機率預報做出最好的決策，得到最大的經濟價值。
;Due to the low predictability in severe weather prediction such as typhoon forecasting, it is important to develop a reliable short-range ensemble prediction system (EPS).This study aims to develop the short-range (0-6h) probabilistic quantitative precipitation forecasts (PQPFs) of typhoons from time-lagged multimodel ensembles using the Local Analysis and Prediction System (LAPS). The ultimate goal is to provide valuable precipitation forecasts for typhoons based on the EPS and help the users optimize their decision making by using the ensemble probabilistic forecasts (EPFs).
The LAPS ensemble prediction system (EPS) has a good spread-skill relationship and good discriminating ability. Therefore, though it is obviously wet-biased, the forecast biases can be corrected to improve the skill of PQPFs through a linear regression (LR) calibration procedure. Sensitivity experiments for two important factors affecting calibration results are also conducted, including: (1) the experiments on different training samples, and (2) the experiments on the inconsistency of observation accuracy. The first point reveals that the calibration results are sensitive to the training samples. Calibration should be performed based on consistent forecast biases between training and validation samples The second factor indicates that the accuracy of observation is inconsistent in the sea and land areas, and samples are dominated by the ocean ones. Therefore, individual calibration for these two areas is needed to ensure better calibration results.
The measure of the usefulness of weather forecasts not only considers the forecast quality, but also the economic benefit associated with the actual use in the daily decision-making process of users. Discrimination ability, which can be assessed by the relative operating characteristic (ROC), is one of the characteristics for evaluating the quality of a forecast and is closely related to the economic value (EV) provided by the same forecast system. Therefore, the other two points in this study is to investigate the consistent and related characteristics between ROC and EV using LAPS PQPFs, and illustrate how to optimize the decision making by using the EPFs. Comparing the ROC and EV derived from EPFs and deterministic forecasts (DFs), EPFs have the advantage over DFs in respect of the discrimination and EV, and such an advantage grows with increasing precipitation intensity. Results from sensitivity experiment on calibration show that the discrimination ability and the maximum EV (EVmax) obtained from a forecast system are insensitive to forecast bias. Due to the terrain-locking effect on typhoon rainfall, LAPS PQPFs have greater discrimination ability, provide more reliable forecast information and offer higher EV to users with high cost-loss ratios in the mountain area than the plain area. Results show that the difference of EV between the mountain and plain areas is mainly attributed to the different discrimination ability in these areas. We also illustrate that even without explicitly knowing the cost-loss ratio, users can still optimize their decision-making and obtain the EVmax by using the EPFs. |
