機械學習法結合時間序列分析預報水庫出流泥砂濃度

Outflow sediment concentration forecasting using integrated machine learning approaches and time series

王乙丞(Yi-Cheng Wang)

指導教授 ： 林國峰

共同指導教授 ： 賴進松

國立臺灣大學/工學院/土木工程學研究所/碩士(2018年)

https://doi.org/10.6342/NTU201802592

摘要

颱洪往往導致大量泥砂進入水庫造成淤積，為使水庫能永續經營必須有效清除淤積。目前異重流排砂為中大型水庫主要排砂策略，若能預先知道泥砂濃度，在適當時機啟動排砂設施便能增加排砂量並減少水資源的浪費。目前現有的水庫出流泥砂濃度預報模式，在濃度轉折處和峰值會低估。因此，本研究提出庫出流泥砂濃度預報模式，可準確預報水庫出流泥砂濃度，特別是修正了濃度轉折處和峰值誤差，提供決策者操作出水工之依據，以提升排砂效率。
本研究結合自組織映射輸出圖(self-organizing feature map，SOM)、支援向量機(support vector machine，SVM)和時序列分析(autoregressive model，AR)建立水庫出流泥砂濃度預報模式，命名為 SOSVM-AR。主要架構分為三階段：分類、預報和即時修正。分類時以SOM模式分析並萃取高價值資訊的資料，經資料再處理後，以SVM預報水庫出流泥砂濃度。最後使用AR，對預報結果作即時修正，進一步增加模式準確度。
本研究選用石門水庫為研究區域，蒐集2012至2016年共六場颱風事件的入流量、出流量、入流濃度、出流濃度和時域反射法實測斷面濃度資料。經過相關係數分析篩選有效輸入項後，預報未來在t+1至t+3小時泥砂濃度，並將結果與單純使用SVM和未使用AR修正的SOSVM比較。結果顯示，在t+1至t+3時刻SOSVM-AR預報泥砂濃度尖峰值最準確，其次為SOSVM和SVM，尤其在t+3時刻最為明顯。均方根誤差、平均絕對誤差、相關係數、效率係數等四個評鑑指標指出SOSVM-AR預報結果皆優於SOSVM和SVM。未來可使用本研究提出之SOSVM-AR預報水庫出流泥砂濃度，作為決策者排砂操作的參考。

關鍵字

出流泥砂濃度預報 ； 自組織映射圖 ； 支援向量機 ； 時間序列分析 ； 自迴歸模式 ； 排砂效率

並列摘要

Reservoir sedimentation is a serious problem in Taiwan. Therefore, reducing sediment deposition in reservoirs is an essential issue. Various strategies have been used to reduce sedimentation. Venting turbidity currents through reservoir outlets can be an efficient strategy. An accurate forecasted outflow sediment concentration is necessary for accessing and increasing the venting efficiency.
In this study, an outflow sediment concentration forecasting model (SOSVM-AR), integrating self-organizing map (SOM), support vector machine (SVM), and autoregressive model (AR), is proposed to yield 1- to 3-h lead time forecasts. First, self-organizing map (SOM) is adopted to extract valuable data which has salient features. Second, the original training data and the reprocessed data are employed to train SVM. Finally, AR is used to real-time correct the forecasts.
An application to the Shihmen reservoir is presented to demonstrate the accuracy of the proposed model. Six typhoons events from 2012 to 2016 are collected to train and test the proposed model. The original SVM and the SOSVM, integrating SOM with SVM, were constructed to highlight how adding the extracted reprocessed data and real-time error correction improves the estimating performance. The results show that the proposed model outperforms over other models, especially for the peak sediment concentration. In conclusion, the proposed model can be used as a reference to reservoir sedimentation management.

並列關鍵字

Reservoir sedimentation ； Outflow sediment concentration forecasting ； Self-organizing map ； Support vector machine ； Autoregressive model

參考文獻

1. Chamoun S, De Cesare G, Schleiss AJ (2016) Experimental investigation on turbidity current venting under restrained outflow discharges. River Flow Proceedings, Saint Louis, USA

2. Akiyama J, Stefan HG (1984) Plunging flow into a reservoir-theory. J Hydraul Eng 110:484–499

3. Annandale GW (2005) Reservoir sedimentation. Elsevier Science, USA

4. Buehler J, Siegenthaler C (1986) Self-preserving solutions for turbidity currents. Acta Mechanica 63:217–233.

5. Cesare GD, Boillat J-L, Schleiss AJ (2006) Circulation in Stratified Lakes due to Flood-Induced Turbidity Currents. Journal of Environmental Engineering 132:1508–1517