請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/8305
標題: | 混合式人工智慧於水庫入流量預報 A hybrid artificial intelligence approach to reservoir inflow forecasting |
作者: | Kuang-Chi Shih 石廣琪 |
指導教授: | 林國峰(Gwo-Fong Lin) |
關鍵字: | 水庫入流量預報,人工智慧,支援向量機,長短期記憶網路,遞迴神經網路,門閘遞迴單元,交替模式, Reservoir operation,Machine learning,Reservoir inflow forecasting,Long short term memory network,Gated recurrent unit,Switched prediction method, |
出版年 : | 2020 |
學位: | 碩士 |
摘要: | 台灣位於西太平洋颱風路徑要衝,1911年至2019年間平均每年約有3.6個颱風侵襲台灣,颱風所挾帶了豐沛的雨量,然短時間高強度的降雨往往造成水庫操作上的不確定性,因此準確的水庫入流量預報可作為相關單位於水庫洩洪中重要的依據。由於水文事件大多都為非線性關係,本研究選擇能良好處理非線性問題的人工智慧模式,運用之模式包含機器學習和深度學習共七種方法(支援向量機、隨機森林、多層感知器、深度神經網路、遞迴神經網路、長短期記憶網路及門閘遞迴單元),預報未來1至6小時之水庫入流量,並採用評鑑指標評估模式表現。因多種模式各有優缺,將七種模式再藉由交替模式,預報未來1至6小時之水庫入流量並與七種模式利用簡單平均法的預報比較。本研究結果探討四個主題: (1)比較多層感知器和深度神經網路模式探討傳統機器學習和新興深度學習應用於水文領域間之差異;(2)比較遞迴神經網路、門閘遞迴單元和長短期記憶網路模式,探討遞迴網路間差異;(3)比較所有模式,挑選出最佳之石門水庫入庫流量預報模式,與優選模式;(4)比較交替預報模式、優選模式與簡單平均法。結果顯示,短期預報(1至3小時)機器學習法與深度學習法表現皆差不多,其中深度學習法的評鑑指標較深度學習法較佳;長期預報(4至6小時),深度學習法雖評鑑指標較深度學習法較佳,但在峰值預報上較差。使用交替模式預報,表現皆比優選模式和簡單平均法表現較佳。本研究提出使用七種模式結合交替模式可供未來學研防災單位參考使用並成為水庫相關單位擬訂水庫操作策略之重要依據。 On average, Taiwan is hit by three to four typhoons each year due to located on one of the main paths of Northwestern Pacific typhoons. These typhoons bring plentiful rainfall and contribute the highest proportion to Taiwan’s water resource. However, the brief and intense rainfall often leads to difficulties in reservoir operations. Thus, accurate hourly reservoir inflow forecasting has become an essential issue to reservoir flood discharging. In this study, seven machine learning methods were employed to forecast the reservoir inflow for 1 to 6 h lead times. The machine learning models include support vector machine (SVM), random forest (RF) and multilayer perceptron (MLP). Deep learning models include deep neural network (DNN), recurrent neural network (RNN), long short term memory network (LSTM), and gated recurrent unit (GRU), respectively. The optimal hourly reservoir inflow forecasting model were determined by the cross-validation and performance measures. Because these models have the advantages and disadvantages, seven models are passed through the switch prediction method to forecast the reservoir inflow in the next 1 to 6 hours and compared with the forecast of the seven models using the ensemble means. The methodology of this study can be divided into four sections: (1) comparing the MLP and DNN models to explore the differences between conventional machine learnings and deep learnings; (2) comparing RNN, LSTM and GRU models patterns to explore the differences between using gated recurrent units or not; (3) comparing all models and determine the optimal forecasting model. (4) Comparing switched prediction method, ensemble means and the optimal model. The results indicate that the machine learnings and deep learnings performs similar in the short-term forecasting (1 hour to 3 h lead time). In the 6 h lead time, Deep learning method, DNN and GRU, have the performance measures. Proposed SPM could simulate the discharge better than the EM and the optimal model. The proposed models, which can produce accurate forecasts, is expected to be useful to support reservoir operations. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/8305 |
DOI: | 10.6342/NTU202002552 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-0608202015582600.pdf | 6.23 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。