在過去研究中，多是針對颱風或梅雨系統的極端降雨發生頻率與空間特性進行研究，其中顯示出 強降雨事件的延時特性與極端降雨特性有顯著之不同。而降雨之時空特性，隨著不同的綜觀尺度天氣 系統會有顯著的差異。換句話說，短延時極端降雨事件的發生機率密度函數(PDF)，在不同的天氣系 統中將有不同特徵。我們也知道大氣數值模式對於強降雨事件的總降雨量有一定程度的預報能力，但 對具有時序列變化特性的短延時強降雨事件卻缺乏預報能力。這是因為過去欠缺對於不同天氣系統的 降雨事件發生機率分布函數的資訊，故數值模式從未針對此一問題進行校驗。因為數值模式在短延時 降雨事件預報能力的不足，所以必需透過建立統計模型的方式，結合即時觀測資料與數值模式輸出， 提高降雨預報的能力。 本研究在Su et al. (2012)的基礎上，第一步先將極端降雨的統計資料延伸至2014 年；並透過 自組織映射(SOM)技術，搭配長期之大氣數值模式再分析資料，建立初步的天氣系統分類資料庫。利 用此一資訊，進一步分析不同天氣系統下臺灣極端降雨事件的時空分佈特性。並根據統計特性所發展 的短延時極端降雨事件的統計預報模型，利用即時觀測資料推估出未來短延時極端降雨事件發生的機 率分布。此外，我們將針對作業的數值預報模式產品發展模式校驗機制，提供數值模式未來改進之參 考。本三年期計畫的最終目標，是發展全新概念之修正數值模式輸出統計預報模型（MMOS），以期有 效延長預報與防災作業上的前導時間。
Most of the previous studies focused on the spatial and temporal characteristics of extreme rainfall events for Typhoon or Mei-Yu systems. As we know, different synoptic weather systems will cause different rainfall patterns and the probability density function (PDF) of short-term extreme rainfall will be changed. Since the PDF of short rainfall frequency was never be systematic studied with any long term observational data in Taiwan. Therefore, the predictability of numerical models is usually assessed by the total rainfall amount. The numerical model output was never be evaluated by the rainfall PDF, which has the time sequence information of precipitation, but this information is very important for the disaster prevention. In order to extend the lead time of hazard warnings, we need to develop a new statistical method to combine the observational data and numerical model outputs, which can also improve the predictability of Quantitative Precipitation Forecasts (QPF). The first step of this project is using the same statistical method, which as Su et al. 2012, to extend the extreme rainfall statistics to 2014. After accumulating long period of rain gauge data, the next step is using the numerical reanalysis data to classify weather patterns and rainfall statistics into several categories via self-organizing map (SOM) technology. We will also collect and evaluate the new high time-resolution rain gauge data (10mins interval) form CWB. Our preliminary goal is to elaborate the short-duration extreme rainfall probability density function in different weather patterns via those long term observation data. We will use the PDF information to build-up a new statistical prediction model for the guidance of weather forecaster and disaster preventing agency. We will also use these PDF to develop a new evaluation method for numerical weather forecasting models which can provide some reference information for future numerical model improvements. The ultimate goal of this three-year project is to develop a modified model output statistics (MMOS) prediction system, which can extend the lead time on forecasting and hazard warnings.
