請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/89875
標題: | 應用卷積神經網路預測區域地下水位之研究-以濁水溪沖積扇為例 Utilizing convolutional neural networks to forecast regional groundwater levels – a case study in the Jhuoshuei River basin of Taiwan |
作者: | 張佑文 Yu-Wen Chang |
指導教授: | 張斐章 Fi-John Chang |
關鍵字: | 地下水預測,類神經網路,深度學習,卷積神經網路,倒傳遞神經網路,隨機森林,台灣, Groundwater prediction,Artificial Neural Network (ANN),Deep learning,Convolutional Neural Networks,Backpropagation Neural Networks,Random Forest,Taiwan, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 由於台灣彰雲地區長期超抽地下水,造成嚴重的地層下陷。也因為氣候變化引起的降水變化加劇和極端天氣事件增多,可能導致乾旱和洪水持續時間延長,從而直接影響地下水的可用性和對地下水的依賴。在長期乾旱的情況下,含水層枯竭的風險更高,特別是在小型和淺層含水層的情況下。由於地下水的緩衝能力,缺水地區的人們將越來越依賴地下水。,使地下水成為更受重視的水資源。本研究使用深度學習的CNN-BP區域地下水位預測模式來預測台灣濁水溪流域的地下水狀況,此模式結合卷基神經網路(CNN)、倒傳遞神經網路(BPNN),以地下水位、降雨量和流量,同時預測未來第一天到第三天( T+1~T+3 )濁水溪沖積扇多個測站的地下水位。本研究使用2000/01/01~2019/12/3 25個地下水位測站、20個降雨量測站和3個流量測站以日為單位的資料,經過隨機森林(Random Forest)篩選因子後作為預測模式的輸入資料。總共將7291筆資料,切分為訓練(5844, 80%)、驗證(723, 10%)及測試(724, 10%)等三階段。
研究結果發現,本研究提出之CNN-BP模式,相較於傳統的倒傳遞神經網路(BPNN)模式,因為其能夠擷取資料時間維度特徵的特點,在預測多測站未來地下水位有更好的表現,本模式預測精準度良好,結果之R2在0.98~0.94 之間;RMSE在0.156m~0.264m之間。本研究亦發現CNN-BP模式的預測誤差與濁水溪的平均岩性有關聯,平均岩石粒徑大的地方容易低估,反之則容易高估。透過海棠颱風的案例分析,驗證本研究的CNN-BP模式可以有效預測暴雨事件下濁水溪沖積扇的潛在地下水補注,得以做為日後土地利用、分區規劃,地下水資源管理等研究之參考。 Due to the long-term over-extraction of groundwater in the Chang-Yun region of Taiwan, serious subsidence has occurred, exacerbated by extreme weather events. As a result, groundwater has become a more important water resource, leading to an increased research focus on the groundwater in the Jhuoshuei River Basin in Taiwan. This study employs a deep-learning CNN-BP model for regional groundwater level prediction. The model combines Convolutional Neural Networks (CNN) and Backpropagation Neural Networks (BPNN) to predict the groundwater levels at multiple monitoring stations in the Jhuoshuei River alluvial fan for the next one to three days (T+1 to T+3). The input data for the prediction model include groundwater levels, rainfall, and streamflow, collected from 25 groundwater level monitoring stations, 20 rainfall monitoring stations, and 3 streamflow monitoring stations, respectively, from January 1, 2000, to December 3, 2019, daily. After feature selection using Random Forest, a total of 7,291 data points were available, which were divided into training (5,844, 80%), validation (723, 10%), and testing (724, 10%) sets. The research findings reveal that the proposed CNN-BP model outperforms the traditional BPNN model by capturing temporal features of the data. It exhibits better prediction performance for future groundwater levels at multiple monitoring stations, achieving good prediction accuracy with R2 values ranging from 0.98 to 0.94 and RMSE values ranging from 0.156 m to 0.264 m. The study also found a certain correlation between the prediction errors of the CNN-BP model and the average lithology of the Jhuoshuei River, with underestimation in areas with larger average rock particle sizes and overestimation in areas with smaller particle sizes. Through verification using the Haitang typhoon data, the CNN-BP model effectively predicts the potential groundwater recharge in the Jhuoshuei River alluvial fan during heavy rainfall events. It can serve as a reference for future land use, zoning planning, and regional groundwater resource management. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/89875 |
DOI: | 10.6342/NTU202303324 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 生物環境系統工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 目前未授權公開取用 | 8.32 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。