機器學習法於水庫異重流到達時間及排砂效率預報之研究

Abstract

台灣水庫淤積日益嚴重，導致蓄水量逐漸減少。為了使水庫能永續經營，有效率地清淤顯得相當重要。異重流排砂為中大型水庫排除細顆粒泥砂最主要之方法，若能準確地於異重流到達壩前之時機開啟排砂設施，可有效地提高排砂效率。由於排砂設施閘門操作需要一段開啟時間，故異重流到達時間的預報對閘門操作有很大的幫助，但目前鮮少研究對於預報異重流到達時間進行研究。故本研究提出預報異重流到達時間及排砂效率之方法，以作為排砂設施操作之依據。 本研究分為三個階段，第一階段為入流量及入流泥砂濃度預報，結合多目標基因演算與類神經網路方法分別對入流量及入流泥砂濃度進行預報。本研究使用改良式自組織線性輸出映射圖，並將其與倒傳遞類神經網路模式及支援向量機兩種常用之類神經網路比較，以尋找最佳之入流量及入流泥砂濃度預報模式；為了改善長延時預報之表現，提出將前面時刻預報值當作輸入項之模式。接著，於第二階段計算異重流到達時間，本研究提出異重流到達時間模式，並基於前一階段預報之入流泥砂濃度、入流量及水庫水位，預報異重流到達時間。第三階段為出流泥砂濃度預報，此階段的預報方法與入流預報之方法相似，但因為輸入因子僅有歷史觀測之出流泥砂濃度，故使用試誤法當作因子篩選的方法，並且利用預報之出流泥砂濃度推估排砂效率，給予關閉排砂設施重要的依據。 本研究選擇台灣北部石門水庫為研究區域，並蒐集2008年至2016年9場資料完整的颱風事件，將其應用於提出模式之訓練與驗證。結果顯示：所提出之預報模式對入流量、入流泥砂濃度及出流泥砂濃度皆有良好的表現，而異重流到達時間預報與觀測資料有良好的一致性。因此，本研究所提出之方法可有效提高排砂效率並降低水資源之浪費，對於水庫永續經營提供良好的管理工具與方法。

Reservoir sedimentation is a serious problem in Taiwan. It hampers the sustainable use of reservoirs and reduces water resource availability. Various methods are applied for sediment mitigation. However, venting the turbidity currents through outlets is the most efficient way to reduce sedimentation. The accuracy of the forecasted turbidity-current arrival-time is very important for increasing the venting efficiency. However, there are few researchers studying the real-time forecasting of the turbidity-current arrival-time in reservoirs. Therefore, the novel outflow sediment concentration forecasting model and turbidity-current arrival-time model are proposed in this study. This study consists of three steps. In the first step, to obtain more effective forecasts of hourly inflow discharge and sediment concentration, novel models with better ability are desired. Integrating the multi-objective genetic algorithm (MOGAs) and the artificial neural networks (ANNs), effective hourly inflow discharge and sediment concentration forecasting models are constructed. In this study, compare the self-organizing linear output map (ISOLO) with commonly used ANNs, the back propagation networks (BPN) and the support vector machines (SVM), for finding the optimal hourly inflow discharge and sediment concentration forecasting models. To further improve the long lead time forecasting, forecasted values are added as key input to the proposed forecasting models. In the second step, a turbidity-current arrival-time model (TCATM) is proposed. The forecasted inflow discharge, sediment concentration and water level are used as input to the TCATM to forecast the turbidity-current arrival-time. In the last step, to obtain more effective forecasts of hourly outflow sediment concentration, the similar methods as in the first step are used. However, the trial and error method is used to screening the input factors. To demonstrate the effectiveness of the proposed model, an application to the Shihmen reservoir in northern Taiwan is presented. Nine typhoons from 2008 to 2016 are selected to train and test the proposed model. The results clearly confirm that the proposed forecasting model is recommended as an alternative to the existing models, and the forecasted turbidity-current arrival-time is in good agreement with the observed data. In conclusion, the proposed method can be adapted as a reference to the water quality and sedimentation management of reservoir operation.