結合型態辨識與集群分析在定量降雨預報之研究

Abstract

近年來人工智慧之研究蓬勃發展，其中以類神經網路為最具代表性的成果之一。類神經網路分為監督式與非監督式架構，其中非監督式學習之網路架構具有相當高的容錯能力，在一個分類完成的架構下，即使輸入的資料有一部分殘缺不全，仍可由殘餘的資訊辨別出該筆資料所屬之類別，進而利用辨識的結果，作為輸出，成為預報的成果。針對此一特性，本研究提出結合型態辨識法與統計學上的集群分析來對集水區內降雨之空間與時間分佈分類以及進行預報，並進一步用於雨量資料的補遺，以期建立一套具通用性的預報模式。而且不論是那幾個雨量站有缺資料，模式都能進行雨量資料的預報。 本研究基於地理環境與氣象因素的考量，假設降雨在空間上與時間上之分佈可能存有某幾種特定的類型，而提出一套以型態辨識法結合集群分析來建構降雨預報模式的程序。利用前一段時間之降雨在空間上與時間上之分佈，即可預報集水區各站下一個小時的降雨情形。無需氣象、氣候等觀測條件。經過驗證與測試，本模式用於淡水河流域之雨量預報具有相當良好之結果。當新的降雨型態被納入模式辨識時，模式也能顯示更好結果。

Recently, the research of artificial intelligence is full of vitality. The Artificial Neural Networks (ANNs) is one of the most representative of achievements. The ANNs can divide into two kinds: supervised learning and unsupervised learning. Unsupervised learning has powerful ability of holding error. A clustered construction, even the input data is imperfect, still can identify the coordinate data from the remainder data. Take the result of recognition into output, it will become the output of forecasting. In connection of the characteristic, the research brings up the pattern recognition and cluster analysis in statistics to classify the rainfall in space and time and to forecast. It is intended to build an all-purpose forecast model. Whenever any data in rain gage is missing, the model also can hold on. Based on the consideration of factors in meteorology and geography, this research assumes that rainfall exist certain pattern of space and time. A procedure of rainfall forecast of model construction is proposed. Only rainfall data of space and time in the previous time are needed to forecast the condition of rainfall next time steps. It doesn’t need any other condition in meteorology and in climate. Present proposed forecast model is tested using historical rainfall data in Danshui River basin. Reasonably good results have been observed. It indicates that the model will perform better every time when new rainfall patterns are integrated.