本研究提出一個聚類分析(Clustering Analysis)與人工智慧結合的颱風時雨量即時預測模式,簡稱CAIM(Clustering Analysis Artificial Intelligence Model),採用中央氣象局淡水氣象站於西元1967年至2006年間,台灣地區有發布颱風警報之所有歷史颱風事件小時觀測資料。預測模式分為兩資料前處理與預測模式兩階段,於資料前處理階段首先利用前饋式類神經網路(Feedforward Neural Network,簡稱FNN)將小時降雨分為「降雨」與「未降雨」兩類,歸類於「未降雨」者其預測雨量為零;歸類於「降雨」者則對其做聚類分析,並以分群後群內資料的相似性及分群後群組間資料的差異性,做為分群數的依據。預測模式階段對於各群組分別建立調適性網路模糊推論系統(Adaptive Network-Based Fuzzy Inference System,簡稱ANFIS)預測模式與自主非線性系統(Group Method of Data Handling,簡稱GMDH)預測模式預測颱風降雨。由測試的29場颱風事件得知,於較長提前預測時間(Lead Time)中,ANFIS模式表現較GMDH模式佳,提前六小時之ANFIS預測結果相關係數為0.58,均方根誤差為2.9 (mm/hr)。
This research proposes a real time typhoon hourly rainfall forecast Model that applies Clustering Analysis and Artificial Intelligence. This Model called CAIM(Clustering Analysis Artificial Intelligence Model) for short. Historical hourly data at Tamsui station during 1967 to 2006 when typhoon warnings were issued are used in this study. The forecast model is divided into two stages. Assorting hourly rainfall to “rain” and “no-rain” by FNN at first. If the rainfall is sorted out to “no-rain”, the rainfall prediction is set to zero. If the rainfall is sorted out in order to rain, Clustering Analysis is applied. Then analyzing the homogeneity within the group, and the differences between the groups to get the best clustering group numbers.Establishing the ANFIS forecasting model, and GMDH forecasting model to predict the typhoon rainfall. The results show that ANFIS has better forecast skill at long lead time then GMDH for 29 typhoons. The correlation coefficient is 0.58, and root meat square error is 2.9mm/hr.