本研究採用美國國家環境預報中心(National Centers for Environmental Prediction , NCEP)第12版全球系集預報系統(Global Ensemble Forecast System version 12, GEFS v12)之重預報資料，透過統計後處理技術-系集核密度模式輸出統計(Ensemble Kernel Density Model Output Statistics, EKDMOS)進行偏差校正與降尺度，目的在於得到臺灣地區特定測站點上具有良好預報品質與價值的1-14天極端低溫機率預報。預報評估顯示: (1)原始系集預報有離散度不足的問題，且有明顯的預報偏差；相較之下，EKDMOS有相當合理的系集離散度，並可移除絕大部分的預報偏差。(2)相較於模式原始系集預報，EKDMOS能提供可信度更高且區辨能力更佳之機率預報。(3)EKDMOS透過提升可信度與解析能力來提升原始模式的BrSS。(4)更多的使用者可採用EKDMOS預報作為決策依據，而得到高於參考原始預報之經濟效益，特別是成本/損失比接近1與0的使用者有最顯著的經濟價值提升。(5)即便EKDMOS可以顯著地改善極端溫度機率預報品質，當預報時間拉長到10天以上而逐漸喪失預報訊息度或預報訊號時，EKDMOS的預報分布仍無可避免地趨近於訓練樣本的氣候分布，這是數值天氣預報在氣象可預報度上的限制。

This study applies the Ensemble Kernel Density Model Output Statistics (EKDMOS) in the 20-year reforecasts of the National Centers for Environmental Prediction (NCEP) Global Ensemble Forecast System version 12 (GEFS v12) to perform bias correction and downscaling. The purpose is to produce 1-14-day probabilistic cold extreme forecasts with good quality and value at specific stations over Taiwan. Forecast evaluation shows that the raw ensemble forecast is under-dispersive with an obvious bias. In contrast, the EKDMOS has a much more reasonable ensemble spread with most of the bias removed. The EKDMOS has more reliable and higher discrimination than the raw ensemble forecast. The EKDMOS increases the Brier skill score (BrSS) of the raw ensemble forecast relative to the sample frequency by both improving the reliability and resolution. Users with a much wider spectrum of cost/loss ratio can obtain more benefit from the EKDMOS as compared to the raw ensemble forecast. Furthermore, users with cost-loss ratio close to 1 and 0 can benefit more than other users. Even with improved forecast quality and value, the forecast distribution of the EKDMOS unavoidably approaches the climatology of the training sample when the forecast informativeness is lost beyond 10 days.