降雨時空特徵分類與降尺度分析─以宜蘭為例

Abstract

降雨模型主要分成統計與序率兩類，統計預測模型在於追求準確地預測特定時刻之單站降雨，但是通常無法考慮降雨於測站之間甚至是不同時間的相依性；序率模型則可在模擬多個測站降雨時，保留它們彼此間相依的統計特性，然其模擬雨量值易受初始狀態的影響。本研究擬發展一套結合統計模型預測能力及序率模型共變異性概念之預測方法，以求面對未來降雨發生時，能透過此預測方法的協助，進而達到有效災害防治和最佳的水資源配置。 本研究將建置以臺灣附近之大氣環流特徵來預測宜蘭縣境內每日區域降雨型態的統計降尺度模型。此模型可以在迴歸預測的同時考量測站之間雨量分布變異。此研究方法使用集群分析來分類區域降雨型態，而經驗正交函數分析用於辨識季節內強勢的大氣環流特徵，最後藉由支持向量分類模型預測區域降雨型態。 在區域降雨型態的分類上，本研究引進空間距平的概念，使得分類的結果除了有區域降雨量級的差異外，尚能分辨出不同空間趨勢的降雨；經驗正交函數分析於選定研究範圍與季節內所得到的每日大氣環流特徵，不論是七至八月或是九至十一月，皆顯示颱風是臺灣附近最主要的天氣系統；支持向量分類的預測模型的類別上趨於兩極，區域大雨及無雨被預測到的次數相對於較多，於各站的平均雨量預測上普遍是被高估的，在預測無降雨型態上，七至八月模型具有一定的能力，本研究定義無降雨型態以外的類別統稱為有降雨型態，有降雨型態的預測上，則不論是七至八月或是九至十一月的模型的表現都相當良好。

Rainfall prediction model can be generally classified into the statistical ones and stochastic ones. Statistical models emphasize an exact prediction of rainfall at a specific location and time while it lacks the ability of considering the dependence between locations. Stochastic models have the capability of reserving the dependence between locations when performing the rainfall simulation; however, the simulated rainfalls are mostly controlled by the initial values. In this thesis, a method combining the prediction power of statistical models and the concept of covariance in stochastic models is developed for providing some usable information when it comes to confronting the rainfall in the future, and this method is aimed to help disaster prevention and optimal allocation of water resources. The purpose of this study is to establish a statistical downscaling model that predicts regional rainfall patterns based on the features of atmospheric circulation. This model takes the spatial variation of rainfalls between gauges into account when performing the regression prediction. The main methods include: (1). Classifying the regional rainfall patterns using cluster analysis (2). Identifying the dominant features of atmospheric circulation in different seasons with the help of empirical orthogonal function analysis (3). Predicting the regional rainfall patterns based on the features of atmospheric circulation. After introducing the concept of spatial anomaly into the classification, the classified regional rainfall patterns not only show the differences of magnitude among every class, but separate regional rainfall with distinct spatial trend. In specified seasons and research area for atmospheric variables, the dominant features of atmospheric circulation shows that typhoon is the principal weather system around Taiwan in both Jul-Aug and Sep-Nov. The predicted regional rainfall patterns made by the prediction model of support vector classification is polarized, which means the numbers of predicted regional rainfall with heavy magnitude and no-rain conditions are rather large. The average of predicted rainfall in each gauge is basically overestimated. The prediction model of Jul-Aug has some capability of predicting no-rain pattern while both prediction models of Jul-Aug and of Sep-Nov have excellent performance of predicting raining pattern, which is defined as a pattern composed of all patterns excluding no-rain pattern.