颱風受到其氣象特徵於登陸區域地形之交互影響,使其路徑與結構改變 ,影響了降雨區域及時間分布,造成地區集中性降雨、河川水位暴漲、水庫入流量激增,若能從颱風路徑的分類瞭解其對集水區流量歷程之影響程度,進而建立相關預測模式可提供防洪策略與水庫防洪操作更多的資訊。 本研究將颱風路徑經由自組特徵映射網路(SOM)進行分類,使用分類資訊輸入倒傳遞類神經網路(BPNN)建立颱風分類模式與流量特徵模式預測未來1~3小時流量(短時距預測),並比較不分類與分類後之預測結果;再者,利用流量累積之統計曲線與預報路徑預測整場颱風之全時距入流量(長時距預測),以預報路徑進行分類後,分析整場颱風全時距流量預測結果與趨勢,討論颱風路徑與水庫集水區入流量之關係。 由短時距流量預測模式之預測結果可得知,使用路徑分類資訊進行流量預測(資料分類模式與流量特徵模式),可以在颱風流量預測上得到較好的結果,故颱風路徑確實對颱洪時期水庫入流量造成不小影響;由長時距預測結果可得知,颱風預報路徑分類與實際路徑分類相似,可獲得較好之預測結果,反之,分類差異大,則預測流量歷程與實際流量歷程差距大。 Due the effects of the regional topography and its climatic characteristics, the track and structure of the landfall typhoons would be change or destroyed that affect the temporal and spatial distribution of rainfall. That may cause regional intense rainfall, the water level suddenly rise and reservoir inflow surge. We expect to investigate the impact of typhoon track on watershed inflow hydrograph from its classification; then, to build the corresponding forecasting models for providing useful flood information. This study would apply self-organizing map (SOM) to classify the track of typhoons and use back-propagation neural networks (BPNNs) to build forecast models with typhoon classification information or inflow characteristics for predicting one- to three-step ahead reservoir inflow. Moreover, we also combine inflow accumulative curves with the forecast track of typhoon to predict the whole inflow hydrograph during typhoon landfall period (long-term forecast). After classifying the forecast path, we analyzed the forecast result of the whole inflow hydrograph to investigate the relationship between the typhoon path and watershed inflow. From the results of short-term forecast models, the path classification information is helpful to short-term inflow forecasts. The typhoon path indeed affects the amount of reservoir watershed inflow. The long-term forecast results showed that the more accurate typhoon forecast path can achieve better results. The forecast inflow would be very different from the inflow hydrograph when the classification was incorrect.