每當颱風侵襲時，水庫上游崩塌處產生大量泥砂進入河川，與水流形成渾水，沿著河道進入蓄有清水的水庫。渾水進入庫區後，因其流速與運載能力下降，使得渾水潛入庫底，沿著庫底移動，此運移渾水稱為異重流。異重流運移至大壩時，受到大壩攔阻，若不能即時排出庫區，渾水層將逐漸上升，即產生渾水潭現象。當渾水層高於取水口，便會影響供水，故預測颱風時期上游河道泥砂濃度與水庫異重流泥砂濃度，為水庫管理重要研究議題之一。 本研究應用NARX模式與R-NARX模式建置上游河道之羅浮站泥砂濃度預測模式及近大壩異重流泥砂濃度；在建置近大壩異重流泥砂推度模式前，必須先確定異重流能否到達大壩才有預測的必要，本研究依據颱風事件統計資料結果顯示，當流量超過1500cms時，異重流能持續到達大壩。 依據評估指標結果，在羅浮站的泥砂濃度預測結果顯示NARX模式成功建置精確度良好之羅浮站泥砂濃度預測模式；若在羅浮泥砂濃度無法即時取得之情況下，則R-NARX模式可使用替代NARX模式進行預測。在近大壩的泥砂濃度預測結果顯示降雨強度對於水庫庫區近大壩泥砂濃度預測結果並無影響，即異重流泥砂濃度不受降雨強度影響；NARX模式預測結果精確度高，且在訓練事件很少的情況下，模式仍具有高度之強健性；若在近大壩泥砂濃度無法即時取得之情況下，則R-NARX模式可提供近大壩泥砂濃度之趨勢。 In Taiwan, landslide frequently triggered by typhoon events can generate large quantities of sediment into a river. When the river flow enters a reservoir, the turbidity current plunge along the bottom of reservoirs as a submerged current driven by the velocity and density differences between the sediment-laden inflow and the stagnant deep clear reservoir water, called a density current. The density current reaches the dam to form a submerged muddy lake and the level of muddy lake would rise if the outlets are close. Therefore, estimating suspended sediment concentration in upstream river and near dam wall is one of important issues for reservoir management. This study applies the NARX and R-NARX models for estimating suspended sediment concentration at Lo-Fu gauging station of Shimen Reservoir upstream and at S07 gauging station of near dam wall. Before analyzing sediment concentration near dam wall, the necessary condition of density current reaching dam wall needs to be identified. According to statistical analysis of data of several typhoon events, the threshold is the reservoir inflow > 1500 cms. The results show that NARX model is successfully applied to construct estimation model at Lo-Fu gauging station to predict hourly suspended sediment concentration. If the observed sediment concentration cannot be obtained from Lo-Fu gauging station, then the R-NARX model can be instead of the NARX model for proceeding on prediction. By taking various input variables into account at S07, the results demonstrate that precipitation does not affect the prediction of sediment concentration. The performance of NARX is still nice, even though the training data is few. The R-NARX model can provide the trend of sediment concentration near the dam wall.
