目前國內對於淹水的防災措施,主要是利用各種設計雨型所推求的淹水潛勢圖,做為災前之淹水影響範圍與最大淹水深之評估,無法在暴雨來臨時能夠有效地依即時雨量進行各地區之淹水預報分析。以淹水潛勢模擬資料為基礎,本研究提出以複合型模式建置淹水範圍預報模式;複合型模式最重要的部分是倒傳遞類神經網路(BPNN),我們利用BPNN進行保全點淹水深預測與該村之淹水面積推估。建置複合型淹水範圍預報模式可分為三大部分:(1)以降雨資訊與模式自身預測淹水深回饋為類神經網路之輸入因子,建置保全點t+1~t+3時刻之淹水預報模式;(2)以類神經網路建置保全點淹水深,推估該保全區域淹水面積之模式;(3)由DTM資料找出累積面積與高程對照表,將推估之淹水面積對應特定高程值,再搭配地理資訊即可畫出淹水範圍。在保全點預測t+1~t+3時刻之淹水預報正確率方面,10公分淹水門檻值平均超過91%、91%與90%;30公分淹水門檻值平均超過86%、87%與86%,都呈現不錯的預測效果,且修正後測試階段更使準確率提升。在推估淹水區域方面,因設計雨型為集中雨型分佈有較大之降雨量,推估淹水面積命中率相當高,其推估淹水範圍也相當符合模擬淹水範圍;而颱風暴雨事件則在淹水情況明顯時,推估淹水面積命中率比較高。 At present, flood defense strategies of the flood authorities are to estimate the flood inundation extent and maximum flood depth from the existing flood inundation potential database generated by the two-dimensional overland flow simulation model with several pre-designed rainfall patterns and scenarios. However, it cannot provide an effectively and timely flood inundation forecast due to the real-time storm rainfall. This study presents hybrid models to build the regional flood inundation forecasting model. The core part is Back-Propagation Neural Network (BPNN) that is used to forecast flood depth and to estimate the area of flood inundation. There are three parts for building the proposed hybrid models: (1) building one to three-hour ahead flood depth forecasting models of security spots due to the real-time storm rainfall, (2) building flooding area estimation models to estimate the flooding area of the security region, (3) creating the lookup tables used to transform a specific elevation into the corresponding cumulative area; then, the flood inundation map can be shown. In this study, the results show that BPNN can be successfully applied with high accuracy for one to three-hour ahead flood depth forecasting. The correct percentages of forecasting flood depths are 91%, 91% and 90% when the threshold of flood depth is 10 cm; 86%, 87% and 86% when the threshold of flood depth is 30 cm. For estimating flooding area, the proposed approach also performs well; the correct percentages are very high for the pre-designed rainfall patterns and the larger storms.