應用多變量分析與倒傳遞類神經網路探討屏東平原之地下水質特性

Application of Multivariate Statistical Analysis and Back-Propagation Neural Network on Characterization of Groundwater Quality in Pingtung Champaign, Taiwan

李囿鈞(LI YOU JUN)

指導教授 ： 黃益助

國立屏東科技大學/工學院/環境工程與科學系所/碩士(2009年)

https://doi.org/10.6346/NPUST.2009.00171

摘要

環保署為能全盤掌握屏東縣境內之地下水水質現況，於1994年辦理「台灣省地下水水質監測站網實施計畫」陸續在屏東縣內設置區域性地下水監測井，截至2004年底共設置76口區域性監測井，並執行地下水質採樣工作，利用歷年地下水質監測數據作為水質監控依據。由於地下水質檢測項目眾多，為簡化運用少數因子來探討屏東平原之地下水質特性，本研究應用「多變量分析」之因子分析、群集分析，將屏東平原1998-2008年地下水質資料進行歸納、整理與分類，探討參數與水質間之相關性，再利用「倒傳遞類神經網路分析」預測未來本區水質濃度之變化趨勢。
本研究經相關性分析結果顯示豐水期與枯水期同樣具有高相關性之變數為總溶解固體、總硬度、氯鹽與硫酸鹽，但也有相異的地方，如豐水期之氨氮、亞硝酸鹽氮與鐵也具有高度相關性，枯水期則是硫酸鹽與錳，因此變數間亦存在重疊的訊息，由盒鬚圖結果顯示大部分的監測井於豐水期之水質變異性高於枯水期，但有些鄉鎮有少部分之監測項目有相反之情形發生，如枋山鄉、林邊鄉、南州鄉、屏東市、高樹鄉、新埤鄉、新園鄉、萬丹鄉與鹽埔鄉之鐵及錳、長治鄉與萬巒鄉之總硬度、崁頂鄉之亞硝酸鹽氮、氨氮、總有機碳與錳、潮州鎮之氨氮與氯鹽；比較特殊的情形為佳冬鄉所有監測項目於枯水期之水質變異性高於豐水期。盒鬚圖分析結果可用來判別屬於極端值的數值並加以刪除，以免影響後續統計分析結果。
主成分分析結果獲得屏東地區豐、枯水期之地下水質代表性因子，其中豐水期之四個代表性因子分別為：「鹽化因子、無機因子、有機污染因子、環境因子」，枯水期之三個代表性因子分別為「鹽化因子、污染因子、無機因子」，佔整體總變異量分別達86.93%與74.85%，另外利用適足性指數(Kaiser-Meyer-Oklin)發現豐、枯水期之分析數值適合進行因子分析，此外以群集分析方法將屏東地區各監測井之豐水期與枯水期依性質相似度及相異度分為四個群集，並分群探討區域內水質特性，可區分顯示群集內地下水質平均狀況，整體上發現內陸鄉鎮水質狀況優於沿海鄉鎮，且沿海區域顯示有海水入侵及鹽化情形發生。
以倒傳遞類神經網路模擬結果發現，亞硝酸鹽氮預測結果並無明顯變化，表示未來水質並無太大之變化，總有機碳之預測結果大多比實際數值高，表示未來水質可能會受有機污染物影響且兩者均符合90%信賴區間；使用判別分析針對所得到之分群結果作為實際已知的分群進行分析，其中主要取自豐水期之第二次分群結果，枯水期則是取自第三次分群結果進行判別分析，結果顯示豐、枯水期各有3口與2口井出現預測誤差，整體分群預測率分別為95.8%與97%。
本研究成果顯示，運用多變量分析及倒傳遞類神經網路可將繁瑣資料簡單化，可以較簡單的方式呈現讓使用者容易辨別，且可藉由歷年數值來模擬預測未來之水質狀況，再針對模擬結果進行評估並研擬其因應對策，可降低監測成本並發揮其最大效益且本研究成果可作為地下水資源規劃與污染整治之參考並可達到污染預警之功能。

關鍵字

多變量統計 ； 地下水質 ； 群集分析 ； 因子分析 ； 倒傳遞類神經網路

並列摘要

In order to command the groundwater quality in Pingtung Champaign, monitoring wells in shallow layer has been set up by Environment Protection Administration (EPA) since 1994 under the project named ”The integrated program on groundwater quality of the monitoring network in Taiwan area.” Until 2004, seventy six monitoring wells has been installed in Pingtung County to periodically take samples to monitor the groundwater quality. Numerous monitoring data and water quality index can be simplified by applying the multivariate statistical method to search for the factors representing the characterization as well as possible pollution sources of groundwater quality in the Pingtung Champaign. This study utilizes factor and cluster analysis to conduct the arrangement and classification of groundwater quality data from 1998 to 2008 for probing interrelationship between the water quality parameters. Then the analysis technique of the back-propagation neural network was applied to predict the variation tendency of water quality.
The results from correlation analysis showed high correlation for parameters total dissolved solids (TDS), total hardness (TH), chloride, and sulfate during both the rainfall profusion and the rainfall withered periods. But they also showed some differences such as ammonia nitrogen, nitrite, and iron in the rainfall profusion period as well as sulfate and manganese in the rainfall withered period. have same result. The Box and Whisker Plot showed variations of water quality, for most townships in study, during the rainfall profusion period were more than those during the rainfall withered period except the Jia Dong Township. However few items of water quality of townships in study did not follow the above rules such as Iron and manganese for the Fangshan, Linbian, Nanjhou, Gaoshu, Sinpi, Sinyuan, Wandan, Yanpu Township and Pingtung city as well nitrite, ammonia nitrogen, total organic carbon (TOC) for the Kanding Township and ammonia nitrogen and chloride for Chaojhou Township. The results from the Box and Whisker Plot can be used to differentiate and delete the outliner not to affect the following results of statistical analysis.
Four principal factors were obtained in Pingtung Champaign of rainfall profusion period using principal component analysis (PCA) including the salted factor, inorganic factor, organic pollutant factor, and environmental factor, nevertheless three principal factors were obtained of rainfall withered period including the salted factor, pollutant factor, inorganic factor. They can interpret 86.93% and 74.85% variances of the integrated groundwater quality characteristics, respectively. Congruence index (Kaiser-Meyer-Oklin) was utilized to evaluate the data collected from rainfall profusion and rainfall withered period feasible for factor analysis. In addition, four clusters were classified according to the similar and dissimilar characteristics of water quality obtained from monitoring wells in Pingtung Champaign. The results showed that the groundwater quality of hinterland was better than that of coastal area. Some coastal areas have already been affected by the seawater intrusion resulting in aquifer salinization.
The results from the back-propagation neural network model showed no significant variation of nitrite between the predicted and measured data, however most of the predicted TOC data was higher than measured ones. It meant that the groundwater quality may be affected by the organic pollutants in the future within 90% confidence interval. The results from differentiated analysis for data of the second cluster of the rainfall profusion period and the third cluster of rainfall withered period showed three wells of rainfall profusion period and two wells of rainfall withered period had predicted errors. The overall predicted accuracy for rainfall profusion and rainfall withered period were 95.8% and 97%, respectively.
The results of this study show application of multivariate statistical analysis and back-propagation neural network can simplify complex data, moreover, groundwater quality can be predicted and modeled via monitoring data over the years. The modeling results can be evaluated to draft corresponding strategy to reduce the monitoring cost and to enhance the cost-effective benefits. The proposed analysis methods in this research can be referred as a management plan for groundwater resources and pollution remediation to achieve the goal of early warning.

並列關鍵字

multivariate statistical analysis ； groundwater quality ； cluster analysis ； factor analysis ； Back-Propagation Neural Network

參考文獻

黃薇儒，2005，屏東平原地下水鹽化情形探討，碩士論文，國立台灣大學，環境工程學研究所，台北。

張介翰，2005，應用多變量統計方法探討區域地下水之水質特徵，碩士論文，國立台灣大學，生物環境系統工程學研究所，台北。

傅心家，1991，神經網路導論，定碁科技股份有限公司，台北。

羅華強，2001，類神經網路-MATLAB的應用，清蔚科技股份有限公司，新竹。

方世榮，2007，統計學導論，華泰文化事業股份有限公司，台北。