摘要: | 洪水災害對人類社會、財產、生命影響甚巨,能夠模擬降雨逕流之機器學習是近二十年水文領域重要的研究之一,其中流量與濁度一直是其中的重要指標,兩者對於水資源的利用、水中生態環境以及公共用水都有顯著的影響,因此水庫流量以及水體濁度的預測具有重大意義。每次經歷豪大雨或者颱風的沖刷挾帶大量泥砂流入石岡壩導致濁度過高無法供水,為解決此問題並且掌握其發展過程,一個好的預測模型是必不可少的。 本論文藉由包含以k-近鄰演算法(k- Nearest Neighbors, kNN)進行遺漏值填補;以遞迴神經網路(Recurrent Neural Network, RNN)、長短期記憶(Long Short-Term Memory, LSTM)、Gated Recurrent Unit(GRU)、支援向量回歸(Support Vector Regression, SVR)、隨機森林(Random Forest, RF)等五種模型進行預測,並以隨機搜索(Random Search)、貝葉斯優化(Bayesian Optimization)進行超參數優化,提供一個能夠有效預測未來多個小時的優化策略。 本研究以一階差分方法新增資料特徵提升預測精準度,在預測未來六小時之結果中,並考量實務應用之前提下,壩前流量與濁度預測最佳之模型分別為LSTM與SVR,R-Squared值分別達到0.81與0.77,並有效預測至未來十二小時,R-Squared值分別達到0.56與0.55。在LSTM貝葉斯超參數優化之壩前流量與濁度R-Squared值分別達到0.83與0.68,在濁度預測之表現與標準化資料集、新增差分特徵方法以及隨機搜索之預測結果相比分別從0.31、0.57以及0.57增加至0.68。 ;Floods have a significant impact on human society, property, and life. Machine learning to simulate rainfall-runoff has been one of the most critical studies in the field of hydrology in the last two decades, and flow and turbidity have been among the important indicators. A good prediction model is essential to solve this problem and understand the development process. This paper uses machine learning techniques such as k-Nearest Neighbors (kNN) to filling missing values. The five models include Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Support Vector Regression (SVR), and Random Forest (RF) for prediction. Random Search and Bayesian Optimization for the hyperparameter optimization. This thesis provides an optimization strategy that can effectively predict the future for many hours. In this study, the first-order difference method was used to add new data features to improve the accuracy of prediction. The best models for predicting flow and turbidity for the next six hours were LSTM and SVR, with R-Squared values of 0.81 and 0.77, respectively, and for the next 12 hours, with R-Squared values of 0.56 and 0.55, respectively. In the LSTM hyperparameter optimization results, the Bayesian optimized flow and turbidity R-Squared values reached 0.83 and 0.68, respectively, the performance in turbidity prediction increased from 0.31, 0.57 and 0.57 to 0.68 compared to the prediction results of the standardized dataset, the first-order difference method and the random search, respectively. |