人工智慧應用於區域地下水系統中砷污染推估之研究

A Study of Artificial Intelligence Techniques for the Estimation of the Arsenic Variation in the Regional Groundwater System

高力山(Li-Shan Kao)

指導教授 ： 張斐章

國立臺灣大學/生物資源暨農學院/生物環境系統工程學研究所/博士(2011年)

https://doi.org/10.6342/NTU.2011.00791

摘要

人工智慧廣泛應用於水文系統中，提升水文量預測及推估準確性，但是鮮少案例應用於地下水水質推估，而地下水水質具有污染不易察覺、變異性大、影響因子不確定、易受到周邊水域環境影響及資料取得不易等特性，一般傳統模式難以推估，另一方面，砷物質存在於地層中，已被證實是造成烏腳病主要原因，對人體健康危害相當嚴重，實有必要建立可靠地下水中砷濃度推估模式，掌握地下水中砷污染情形，因此，本研究主要目的為應用類神經網路模式推估地下水中砷濃度變化。
本研究以地下水受嚴重砷污染之台灣西部雲林縣沿海地區為研究區域，並採用水利署1992年至2005年設置於本區28座監測井之水質資料為分析對象，地下水水質採樣期間，因經費或人為因素，部份監測井停止採樣或資料缺漏，影響後續對地下水中砷污染擴散機制之瞭解，為補遺本區監測井之砷濃度資料，本研究第一部分採用倒傳遞類神經網路，建立空間模式補遺地下水中砷濃度資料，在建立模式過程中遭遇到資料過少，模式過度訓練問題，應用主成份分析、交叉驗證法及修正型目標函數加以改善，有效提升模式推估精確度；另考量在尋求最佳推估模式架構及參數過程中，由於模式架構不確定與參數過多需耗用大量優選時間，且無法獲得可最佳模式，故本研究應用遺傳演算法之強大搜尋能力優選最佳模式之架構，其中優選項目包括輸入層因子、隱藏層之神經元個數及修正型目標函數之係數，成功解決模式之架構與參數不確定問題。
本研究之第二部份，主要考量砷在地下水環境中受到多種水質因子影響，故深入探討砷與其他地下水水質間之相關性，從而利用水質因子建立地下水中砷濃度之推估模式，模式分為單一水井模式及區域模式，整體而言，以單一水井模式推估結果較佳，但是，區域模式應用範圍較廣，可展現區域地下水中濃度特性，最後，本研究將觀測與本模式推估之地下水中砷變化結果繪出地下水中砷污染潛勢圖，展現本研究區域內1992至2005年砷濃度在時間與空間變化，提供政府單位與相關研究者了解地下水中砷變化情形及傳輸機制，有效減少居民誤飲用高砷地下水之風險，達到有效管理及利用地下水之目的

關鍵字

砷 ； 地下水水質 ； 人工智慧類神經網路 ； 倒傳遞類神經網路修正目標函數 ； 主成份分析 ； 遺傳演算法

並列摘要

Artificial intelligence is extensively applied to hydrological systems and is successfully implemented in the quantitative estimation of water quality. However, artificial intelligence techniques are seldom employed in the prediction of groundwater quality. The features of the groundwater pollution include imperceptibility, complex affective factors and limited data. It is not easy to employee traditional models for estimating the water quality in groundwater systems. Arsenic (As) proves to be a main factor of black-foot disease and threatens the health of residents. Constructing a reliable model for estimating arsenic concentration in groundwater is essential. Therefore, the aim of this study is to construct an artificial neural network (ANN) model for estimating arsenic concentration in groundwater systems.
From 1992 to 2005, the government takes into account the serious arsenic pollution that occurred in the coastal area of the Yun-Lin County in Taiwan and set up 28 monitoring wells for investigating the pollution in groundwater. The collected water quality data were used when constructing models in this study. However, due to limited budget and/or human factors, some arsenic concentration data from these wells were missing, which affects the realization of the pollution in groundwater. The first subject of this study is to construct a spatial model for estimating missing data by applying ANN. During the process of model construction, inaccuracy and over-fitting commonly occur in sparse data. To overcome these problems, the principal component analysis, the cross-validation and the modified performance function are employed when constructing the model. These methods have the ability to effectively alleviate the over-fitting problem and improve model accuracy. On the other hand, searching and identifying the optimal ANN structure is quite time and labor consuming. Genetic algorithm is used to identify the effective input factors and the suitable number of neurons in hidden layer.
Another subject of this study is to build a water quality assessment model for arsenic concentration by analyzing the relationship between arsenic concentration and other water quality factors in groundwater. This subject has two scenarios: one is for the single well model; and the other is for the regional model. Results indicate that the affective factors of arsenic concentration significantly vary from the north to the south in the coastal area of the Yun-Lin County. Overall, the single well model performs better than the regional model, despite that the regional model can be extensively applied over the study area. Finally, the results of the spatial and water quality models are applied to displaying the distribution map of arsenic pollution so that groundwater managers can easily realize the temporal and spatial variation in arsenic concentration during 1992 and 2005. The information of arsenic variation can reduce the risk of drinking contaminated groundwater for local residents and effectively enhance the control and management of arsenic pollution in groundwater.

並列關鍵字

Arsenic ； Groundwater quality ； Artificial neural network (ANN) ； Back-propagation neural networks (BPNN) ； Modified performance function (MPF) ； Principal component analysis (PCA) ； Genetic algorithm (GA)

參考文獻

108. 段鏞、傅金城、鄧慰先，2007，「適應性網路模糊推論系統於降雨－洪水位預報之研究」，臺灣水利56(2)，50-70。

109. 高慧珊，2007，「以自組特徵映射網路推估蒸發量」，臺灣大學生物環境系統工程研究所碩士論文。

122. 陳永祥，2009，「演化式類神經網路於水文系統預測之研究」，國立台灣大學生物環境系統工程研究所博士論文。

1. Adams, S., Titus, R., Pietersen, K., Tredoux, G., Harris, C., 2001. Hydrochemical characteristics of aquifers near (24)Sutherland in the Western Karoo South Africa. Journal of Hydrology 241(1-2), 91-103.

2. Al Fraij, K.M., Abd El Aleem, K., Al Ajmy, H., 1999. Comparative study of potable and mineral waters available in the State of Kuwait. Desalination 123, 235–264.