Construct intelligent groundwater level estimation models–A case study at Zhuoshui River Basin in Taiwan

水資源匱乏係目前全球共同面臨的問題，臺灣受限於地形及降雨分佈不均，大部份雨水無法被利用，地下水因成本低廉，使其成為相當重要之可用水資源，如何有效的保育及補注地下水資源已成為重要的議題。濁水溪流域於扇頂上游地區及沖積扇扇頂為良好的地下水補注區，惟因大量抽取地下水的結果，引發嚴重地層下陷問題，為能即時掌握及預測地下水位變動趨勢，針對未來可能發生地層下陷之區域，提出因應對策，本研究蒐集與整理濁水溪流域內之主要河川流量站、地下水位站及雨量站等長期觀測資料，探討扇頂上游地區及沖積扇扇頂的降雨、河川流量、地下水間之關係，並建置地下水位推估模式，俾作為中部地區地層下陷防治之參據。 本研究採用2002至2011年地下水位、河川流量及降雨量之水文因子資料，透過相關性分析探討月平均地下水位與月平均水文因子之影響關係，並藉由人工智慧可處理複雜非線性問題之特性，應用倒傳遞類神經網路建構長時距地下水位推估模式；分析比較多種模式架構(不同輸入因子之組合)，結果顯示加入地下水位、雨量、流量及放流量之模式具有良好之推估表現(相關性大於0.8)；比較不同輸入因子組合對模式之改善率，作為地下水模擬模式之架構依據，並對建構完成的模擬模式之輸出項(地下水位)進行輸入項因子偏微分，藉由各輸入因子(雨量、流量、放流量)微分值之正負值及分佈情形，探討各輸入項於模式中與輸出項之正負向相關性；最後利用敏感度分析方法，探討雨量增加時，各站地下水位變動趨勢，進一步掌握地面水對地下水的影響機制，以有效運用於水資源管理。The shortage of water resources is a global problem. Due to the steep slopes and gradients of rivers, rapid flows and uneven spatio-temporal rainfall distributions in Taiwan, most of rainfalls flow directly into the ocean. Groundwater has become an important water resource because of its low cost and easy extraction. The upstream zone and the proximal alluvial fan of the Zhuoshui River are good natural groundwater recharge areas. However, the over extraction of groundwater occurs in the coastland of southwestern Taiwan, which results in serious land subsidence. To obtain and estimate the trend of groundwater level variations for making countermeasures in response to future possible land subsidence areas, this study establishes the relationships between rainfall, streamflow and groundwater level and constructs intelligent groundwater level estimation models for the upstream zone and the proximal alluvial fan of the Zhuoshui River basin based on long-term observed data of streamflow, groundwater level and rainfall, which can provide valuable information for the prevention and treatment of land subsidence. In this study, data of groundwater level, streamflow and rainfall recorded in the Jhuoshuei River basin during 2002-2011 were obtained from the Water Resources Agency, Taiwan. The correlation analysis is first applied to building the relationships between monthly mean groundwater level and monthly mean streamflow as well as monthly mean rainfall. Artificial neural networks (ANNs), which resemble the human thinking process and possess a great ability to handle non-linear complex systems, are implemented to configure estimation models. By taking various input combinations into account, the most suitable estimation model of groundwater level can be established by the back propagation neural network (BPNN). The results demonstrate that the constructed estimation models can suitably estimate monthly groundwater level with high correlation (larger than 0.8). For investigating the mechanism of groundwater level variation, a sensitivity analysis is then conducted on input variables of the estimation model by using the partial derivative method. Based on the distributions of the partial derivative values corresponding to each inputs (rainfall, streamflow and discharge), we establish the relationships between inputs and output (groundwater level) and identify rainfall as the most significant key factor. Then the impacts of rainfall amount on groundwater level variation can be obtained by the sensitivity analysis. The results of the proposed approach can be used as a valuable reference to water resources management and conservation.謝誌 II 摘要 IV Abstract V 目錄 VII 圖目錄 IX 表目錄 X 第一章 緒論 1 1-1 研究緣起及目的 1 1-2 研究架構 2 第二章 文獻回顧 4 2-1 地面水與地下水位變動特性分析 4 2-2 類神經網路 5 第三章 理論概述 8 3-1 類神經網路 8 3-2 倒傳遞類神經網路(BPNN) 11 3-3 相關性分析 16 3-4 敏感度分析(sensitivity analysis) 18 3-5 評估指標 19 第四章 研究案例 21 4-1 研究區域概述 21 4-2 資料蒐集與處理 24 4-2-1資料蒐集 24 4-2-2資料補遺 26 4-2-3資料校正 31 4-3 地面水與地下水之各站基本特性分析 35 第五章 結果與討論 40 5-1 長時距地下水位歷線推估模式 40 5-2 長時距地下水位歷線模擬模式 57 5-3 敏感度分析 63 第六章 結論與建議 70 6-1 結論 70 6-2 建議 71 參考文獻 73 附錄 8