最佳化整合多種機器學習法於水庫泥砂濃度之預測

Real-time forecasting of suspended sediment concentrations in reservoirs by optimally integrating multiple machine learning approaches

王柏翔(Po-Hsiang Wang)

指導教授 ： 林國峰

國立臺灣大學/工學院/土木工程學研究所/碩士(2021年)

https://doi.org/10.6342/NTU202102340

摘要

台灣地理位置特殊，位於西太平洋之颱風路徑要衝，颱風期間所造成之強降雨，挾帶大量泥沙進入水庫，進而導致台灣水庫淤積問題日益嚴重。以目前中大型水庫防淤而言，參照經濟部水利署之說明，異重流排砂為積極且經濟的方式。出流泥砂濃度預報不僅提升異重流排砂之效率，同時也減少水資源浪費，有效率地清淤使水庫能夠永續經營。本研究建立泥砂濃度預報模式，預測水庫出流泥砂濃度，以作為排砂設施操作之依據。
首先本研究應用SRH-2D數值模式針對歷史颱風事件進行資料補遺，以得到完整斷面泥沙濃度資料，再運用高效率且高準確度的機器學習及深度學習等共七種模式，包括多層感知器、支援向量機、隨機森林、深度神經網路、遞迴神經網路、長短期記憶網路以及門閘遞迴單元，預測未來3小時之水庫出流泥砂濃度；以評鑑指標評估各模式表現，最後使用交替模式發展最佳的出流泥砂濃度預報。
本研究以石門水庫為研究區域，驗證本研究提出模式之可行性及準確性，蒐集2008年至2016年共10場颱風事件的斷面濃度資料，預測未來0.5、1、1.5、2、2.5和3小時的水庫出流泥砂濃度。結果顯示預測未來3小時之水庫出流泥砂濃度，使用交替模式與簡單平均法之比較，於發電進水口出水工RMSE值從10.92×103 ppm減少至9.48×103 ppm、於石門大圳出水工從3.49×103 ppm減少至2.68×103 ppm，其值分別改進13%和23%。使用交替模式相較於優選單一模式，RMSE值的改善幅度則為14%和22%。因此提出的預報模式皆有不錯的表現，尤以預報未來1至3小時的交替模式最佳，其表現勝於優選單一模式與簡單平均法，希冀本研究提出之方法能夠作為水庫管理操作之依據。

關鍵字

機器學習 ； SRH-2D ； 深度學習 ； 出流泥砂濃度 ； 交替模式 ； 水庫管理

並列摘要

Taiwan is situated in one of the main paths of the northwestern Pacific typhoons. The torrential rain brought by typhoons carry a large amount of sediment into reservoirs. According to the statement of Taiwan Water Resource Agency, the density current venting is currently a positive and economical way to solve siltation problems. The forecasts of suspended sediment concentration (SSC) not only improve the efficiency of sediment discharge but also reduce the waste of water resources. Therefore, this study proposes a two-step switched machine learning-based approach (Switched-ML) for constructing an effective model to forecast reservoir SSC.
To compensate for a deficiency in measured SSC data, historical typhoons are modeled using the well-validated SRH-2D numerical model. Different ML algorithms are adopted in the first ML step to build multiple ML-based SSC forecasting models, including multilayer perceptron (MLP), support vector machine (SVM), random forest (RF), deep neural network (DNN), recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU). In the second step, a switched forecasting strategy is developed to optimally integrate forecasts from multiple ML-based models to provide more accurate forecasts.
The results show that the SSC forecasts obtained from the SVM and LSTM approaches are confirmed to be superior to those from other ML-based models. The Switched-ML outperforms the others, particularly for 1- to 3-h ahead forecasting. The improved accuracy of SSC forecasting resulting from the Switched-ML is beneficial. This research can be used as a reference for the effective reservoir operation and management of sedimentation during typhoons.

並列關鍵字

Suspended sediment concentration ； SRH-2D ； Machine learning ； Switched forecasting ； Reservoir operation

參考文獻

Ashida, K., Egashira, S., 1975. Basic study on turbidity currents. Proceedings of the Japan Society Civil Engineering. 237, 37-50.

Auel, C., Kantoush, S.A., Sumi, T., 2016. Positive effects of reservoir sedimentation management on reservoir life–examples from Japan. 84th Annual Meeting of ICOLD, Johannesburg, South Africa.

Bashir, A., Shehzad, M.A., Hussain, I., Rehmani, M., Bhatti, S.H., 2019. Reservoir Inflow Prediction by Ensembling Wavelet and Bootstrap Techniques to Multiple Linear Regression Model. Water Resour. Manage., 33, 5121-5136.

Bong, T.H., Son, Y.H., Kim, K.S., Kim, D.G., 2018. Modeling of Suspended Sediment Transport Using Deep Neural Networks. Journal of The Korean Society of Agricultural Engineers, 60 (4), 83-91.

Chen, J., Zhao, K., 1992. Sediment management in Nanqin Reservoir. International Journal of Sediment Research, 7 (3), 71-84.