暴雨時期所產生的大量地表逕流，往往造成下游人口稠密地區嚴重的洪水威脅，對於雨量充沛且高降雨強度的台灣地區而言，集水區降雨一逕流的模擬，為重大天然災害防治工作不可或缺的基礎研究課題。近年來，隨著人工智慧 (Artifi cial Intelligence) 的長足發展，陸續有水文學者利用類神經網路 (Artificial Neural Networks) 來進行集水區降雨一逕流關係的模擬，許多案例顯示其對逕流量的推估正確性能 夠達到傳統數理模式的水準，在洪水預警等課題上有極高的應用潛力。本研究利用倒傳遞類神經網路 (Back-Propagation Network) 在大甲溪上游地區以三場颱風暴雨水文資料進行實 證研究，藉由不同輸入參數的組合比較方式進行洪水的預測與準確性的分析，初步成果能夠在 2 小時事前預測時距狀況下， 透過多個雨量測站的雨量觀測值、流量觀測值、雨量及流 量相關運算參數的輸入，達到對流量變動、洪峰流量、洪峰發生時間的預測。研究結果亦顯示：考量原始雨量及流量觀測值以外的相關輸入參數，有助於提高單場暴雨事件的水文模擬 能力，有效的輸入參數也未必需要彼此獨立：在面積較為廣闊的集水區上，使用多個降雨測站資料做為原始輸入資料的方式有其必要性，其誤差要較使用徐昇氏面積權重區域降雨量方式為小。受限於資料長度及完備性的限制，在洪水歷線整體形狀、洪峰流量、洪峰發生時間、基流量等的模擬上，部份仍存在某種程度的誤差量有待進一步尋求解決。未來亦應對有效的輸入參數遴選技術 (如：遺傳演算法 )、集水區臨前含水狀況掌握、地文因子對降雨。逕流關係的影響等課題做進一步研究。

Large volumes of runoff during storm periods makes serious flood hazard a frequently occurring problem in Taiwan, therefore, studies on rainfall-runoff simulation are important tasks for flood mitigation programmes. In recent years, hydrologists have demonstrated the great potential of using artificial neural networks for flood forecasting from rainfall-runoff studies. In this study, a back-propagation artificial neural network was used to simulate three storm events on an upstream section of Ta-chia river basin. Scenarios were designed with differnt combinations of rainfall and runoff using data from several gauges. The preliminary result shows that the model can forecast the magnitude and time of peak flow accurately when the leadtime is less than 2 hours. At the same time, the research implies that accurate modeling needs not only rainfall and runoff measurement values but also the other hydrological inputs using a network of gauges produces a more accurate prediction than rainfall inputs calculated with these original data. In particular, in larger watersheds, rainfall inputs using a network of gauges produces a more accurate prediction than area weighted techniques using Theissen polygons. Inaccuracies in estimating the rate of flow using the rising and falling limbs of the predicted flood response hydrograph occurred because of the limited availability of storm events data in this study. Further research is recommended using a genetic algorithm to optimize the selection of input variables to the neural model. Soil moisture and landuse conditions in the watershed are also suggested for consideration in the neural model in further research.