台灣每年平均約有3至4次颱風侵台，挾帶豐沛的降雨量對於水庫與河川易造成重大的衝擊，因此，颱風時期的降雨量與流量預報可提供有用的水情資訊，以做為水庫操作與河川防災之參考。本研究以石門水庫集水區為研究區域，建構類神經網路降雨修正模式進行修正多重觀測工具之定量降水估計(QPESUMS)資料，進而以修正雨量資料建置集水區降雨-逕流預報模式，探討水庫集水區之降雨與逕流關係；研究主要分成降雨及流量兩部分，降雨的部分，首先以倒傳遞類神經網路(BPNN)修正QPESUMS網格降雨量，再將集水區分成8個子集水區並以BPNN預測各子集水區未來的平均降雨量；流量的部分，探討集流時間與流量之關係，利用QPESUMS及平均降雨量預測結果建置BPNN與三角單位歷線之流量推估與預測模式。 降雨的部分，利用雨量站資訊及地文因子修正QPESUMS網格降雨量，並利用平均高程與現時刻雨量預測未來1~3小時的平均降雨；流量的部分，根據集流時間與流量之關係以BPNN及三角單位歷線建置3個模式，BPNN I使用前時刻、現時刻雨量與流量資訊以BPNN預測流量，BPNN II在預測T+3~5時刻流量使用前時刻、現時刻雨量與流量及平均降雨預測結果以BPNN預測流量，模式三使用前時刻、現時刻雨量及降雨預測結果以三角單位歷線(TUH)推估流量。 降雨的部分，由評估指標可得知QPESUMSE雨量修正有不錯的效果，由雨量預測模式之結果得知在T+1平均雨量預測之效果較好。流量的部分，由評估指標可得知BPNN I及BPNN II可有效預測流量。TUH為計算簡易之模式，對於推估流量的效果相當有限。 Typhoons hit Taiwan around three to four times a year, bringing a huge amount of rainfall, which easily cause reservoir and river a significant impact. Therefore, rainfall and flow forecasting can provide useful information to prevent reservoir operation and flood disaster during typhoon periods. This study used artificial neural networks (ANNs) to build a precipitation corrected model, precipitation forecast models and rainfall-runoff forecast models in Shihmen Reservoir watershed for investigating the relationships between rainfall and runoff within this watershed. The precipitation corrected model is used to correct Quantitative Precipitation Estimation and Segregation Using Multiple Sensor (QPESUMS) data. In this study, there are two parts: rainfall and flow. In the rainfall part, used a back-propagation neural network (BPNN) for correcting QPESUMS data; then, divided the catchment into eight sub-catchments and applied BPNN to build the precipitation forecast models to forecast one- to three-step-ahead average QPESUMS precipitation of these eight sub-catchments. In the flow part, investigate the relationship between time of concentration and flow; then, build flow forecast models by using BPNNs and triangular unit hydrograph (TUH) based on the average QPESUMS precipitation data of sub- catchments to forecast one- to five-step-ahead flow. In the rainfall part, the precipitation corrected model’s inputs are rainfall gauges precipitation data and geomorphologic factors. The precipitation forecast models’ inputs are the average elevation and precipitation of the sub-catchment. In the flow part, the forecast models’ inputs are the corresponding previous average precipitations of all sub-catchments according to the time of concentration. For the rainfall part, the results show that the precipitation corrected model can effectively correct QPESUMS data. The precipitation forecast models can obtain nice results in forecasting one-step-ahead average precipitation. For the flow part, the results show that the BPNNs outperform TUH models and can be adequately applied with high accuracy to the study of real-time flow forecasts in the study area.
