本研究延續葉等(2016)的工作，在中央氣象局之系集預報系統的架構下，依據系集預報系統設計之 原理，並由2015年梅雨季統計校驗之結果，探討各種系集預報方法產生之定量降水預報(Quantitative Precipitation Forecast, QPF)產品特性，包括系集平均、機率擬合平均(PM, Probability-Matched Mean)、NPM (New PM)及超越機率之 QPFP (Exceeding Probability of QPF 或 QPF Percentile)產品。整體而言，所有系 集預報方法之預報技術皆隨雨量門檻提高而降低。系集平均對小（大）雨之預報技術高（低），具有對 小雨高估但大雨低估且不均勻偏離的特性。決定性預報之WRFD (DeterministicWeather Research and ForecastmgModel)在小（大）雨量的預報技術低（高）於系集平均，與觀測相比有略低估之傾向。PM在 極端降雨之校驗技術最高，但偏離指數也大；小至大雨則略有高估且偏離指數略大於1 aNFM是最中性 的定量降水預報產品，對極端降雨具有有限的預報技術，而對整體雨量僅有輕微高估，但其偏離指數最 接近1，近似於均勻無偏。QPFP之偏離指數隨機率門檻增大而減小；任意機率之QPFP隨雨量門檻增 加呈不均勻偏離，但程度小於系集平均；50%以上之QPFP對整體之降水為低估；而QPFP20與QPFP30 具有較高的預報技術，對極端降雨仍具有限的技術，雖略有過度預報，但不失為優良的指引。另外，系 集平均、PM或NPM之總降水量與系集預報系統總平均降水量相等，具有保守特性；但QPFP之總降 水量則不保守，且隨機率門檻增加而下降。 NPM是系集預報系統在理論及統計上的最佳解，具有高預報技術、高偵測率，低誤報率、低漏報 率且接近無偏的最佳的預報特性。為了防災上的需求，對於具有有限機率的高風險事件，通常會預作提醒及特別注意。因此中央氣象局之官方定量降水預報作業會傾向提供稍低機率、但高風險的事件，以致 預報稍有過度之傾向，其特性與QPFP20相似，為本研究建議另一綜合評估之重要參考。QPFP產品亦 可提供預報人員更多預報之彈性，例如針對防災之應變，可依機率提供不同情境評估，如：1.)高風險、 低機率的QPFP5或QPFP10 ； 2.)稍高風險、稍低機率的QPFP20或QPFP30 ；或3.)低風險、高機率的 QPFP50 或 QPFP70 等。

In this study, we follow the work by Yeh et al. (2016) and thoroughly examine several QPF products derived by different post processes on the ensemble members of the CWB (Central Weather Bureau) operational WRF (Weather Research and Forecasting Model) ensemble prediction system (WEPS), including ensemble mean, PM (Probability Matched Mean), NPM (modified PM by Yeh et al, 2016) and QPFPs (exceeding probability ofQPF or QPF percentile). Based on the statistical verification of 12-24 hour QPF during Mei-yu season (May and June) in 2015 for all QPF products in Taiwan area, it is found that the overall threat score (TS) and equitable threatscore (ETS) decrease with increasing rainfall threshold. The ensemble mean tends to overpredict the light rainfall while under-predict the heavy rainfall, characteristic of an inhomogeneous decreased bias with increased thresholds. Skill scores of deterministic WRF (WRFD) are slightly lower than WEPS for light rainfall, but higher for heavy rainfall. The result based on bias score also indicates that the rainfall is slightly underestimated at all rainfall thresholds as compared to the observation. PM is slightly over-predicted at most rainfall thresholds, but has a large bias at the extreme rainfall. NPM is similar to PM with less overprediction at most thresholds, especially for the extreme rainfall. Its unbiased or neutral nature makes NPM as the best solution for WEPS. QPFPs show a decreasing bias with increasing probability, which are not homogeneous as well. It also under-predicts QPFPs with probabilities up to 50% or above. QPFP20 or QPFP30 could be a good guidance thought it is overall over predict. For products derived from WEPS, it is found ensemble mean, PM and NPM conserved the domain total water amount, but QPFPs are not. In summary, NPM is the best solution for WEPS in theory and statistics, with characteristics of higher skill score, higher probability of detection, lower false alarm rate and neutrally un-bias. Another important guidance would be QPFP20 since the operation QPF tends to deliver a higher risk forecast with lower probability. Finally, QPFP is suitable for probability-based decision making. And we may provide guidance for applications of QPF products in the following three categories 1) high-risk with low probability: QPFP5 or QPFPlO; 2) higher risk with lower probability: QPFP20 or QPFP30; or 3) low risk with high probability: QPFP50 or QPFP70.