結合K-means法與類神經網路建立用電量推估抽水量模式-以濁水溪沖積扇為例

Application of K-means and Artificial Neural Network to Develop a Model for Groundwater Pumpage Estimation Using Electricity Records - A Case Study of Chou-Shui River Alluvial Fan

陳子裕(Tzu-Yu Chen)

指導教授 ： 徐年盛

國立臺灣大學/工學院/土木工程學研究所/碩士(2017年)

https://doi.org/10.6342/NTU201703923

摘要

本研究之目的為建立單口井之抽水量推估方法，空間上可細緻地推估每一口井之抽水量，並可加總評估任一區域內的總抽水量，該模式主要採用抽水馬達用電量推估法結合K-means聚類法與類神經網路方法，作為模式架構來推估單口井之抽水量，並將其抽水量推估結果與水平衡分析法所推估之抽水量結果進行分析比較。
　　本研究之單口井抽水量推估方法考量到不同抽水井的特性不同，應先針對不同抽水效率(抽水量/用電量)的水井進行特性分類，再分別建立各分類水井的抽水量推估模式進行抽水量估算應較為合理，然而抽水效率計算公式中的抽水量即為本研究模式所要推估的未知數，因此本研究將先建立以類神經網路方法為基礎的抽水特性分類模式，以得到抽水井的水井特性分類，再以各分類的非線性抽水量推估模式進行抽水量的計算，此為本研究用電量推估法的完整流程。
　　於抽水量推估模式的部分，本研究先以K-means聚類法針對濁水溪沖積扇2012及2014年有用電量與抽水量資料的試驗水井先進行抽水效率的分類，一共分為四類，並針對此四種分類的抽水井資料分別進行抽水量推估模式的參數檢定，本研究所用參數檢定之優選模式為一非線性規劃(Non-Linear Programing, NLP)模式，乃採用Lingo11套裝軟體進行參數求解，以得到分類之非線性抽水量推估模式參數，此外本研究亦推求不分類的抽水量推估模式參數，並以此兩種抽水量推估模式進行試驗水井的抽水量推估，比較其抽水量推估結果與實測值的差異，結果顯示先經過分類的抽水量推估模式較為準確。
　　於類神經分類模式部份，本研究以倒傳遞類神經網路方法(BPNN)、即時回饋學習演算法(RTRL)及調適性網路模糊推論系統(ANFIS)等三種方法，將馬達馬力(P)、出水管徑(D)、抽水揚程(L)當作模式輸入項，以分類類別作為輸出項，並以K-means聚類法之分類結果作為類神經網路模式輸出項的目標值，分別建立三種抽水特性分類模式，並以相關係數(Correlation Coefficient, CC)、效率係數(Coefficient of Efficiency, CE)以及均方根誤差(Root Mean Square Error, RMSE)等三項評比指標來比較此三個模式的分類結果準確度，結果顯示BPNN結果最佳，RTRL次之，ANFIS較差。此外本研究也將此三種模式之分類結果帶入分類非線性推估模式進行試驗水井之抽水量計算，並與不分類抽水量推估模式所推估之抽水量結果進行比較，結果顯示藉由此三種類神經分類模式所分類之結果計算出來的抽水量皆較不分類抽水量推估模式之抽水量推估結果好。
　　本研究將所建立之類神經分類法及分類之非線性抽水量推估模式應用在濁水溪沖積扇，其應用結果發現以用電量推估法及水平衡分析法所求得之灌溉抽水量相近，因此本研究之推估結果應屬合理範圍。推估結果可得知芳苑鄉及二林鎮在枯水期之灌溉用水抽水量較高，且在每年2~4月之灌溉用水之抽水量更大，地下水管理上應特別注意2~4月有無密集抽水之情形並加強監控是否有超抽之情形。
　　本研究所建立之用電量推估法特別適用在臺灣的農業用水井管理，本分法之優點為可細緻地推求單口井之抽水量，再加總推估各空間尺度下之抽水量，可估算單口井、一村里、一鄉鎮或一縣，在抽水量管理方面可以更加精準，且僅需各抽水井之月用電量、馬達馬力(P)、出水管徑(D)、抽水揚程(L)，便可有效率地推求各水井之抽水量，惟本方法尚仍有時間尺度上的限制，因用電量資料每月僅有一筆，故目前僅推求月抽水量，未來建議可以在時間尺度上細化(月、旬、週、日、小時)。

關鍵字

濁水溪沖積扇 ； K-means聚類法 ； 倒傳遞類神經網路 ； 即時回饋學習演算法 ； 調適性網路模糊推論系統 ； 用電量推估

並列摘要

This study is to establish a method to estimate the single-well pumpage. This method can estimate the pumpage of evey well accurately and the total pumpage in arbitrary region.The model adopts pumping motor electricity record Estimation, K-means clustering, and Artificial Neural Network to be the architecture of model and then compare and discuss the results with groundwater hydrograph analysis.
　　Single-well Pumpage Estimation considers the characteristics of different pumping wells. As a result, this study shuld classify the wells acoording to pumping efficiency ((Pumping capacity)/(Electricity consumption)) firstly, and then establish the pumpage–estimation models of every classified well. However, the pumpage of pumping efficiency formula is unknown, therefore this study adopt Artificial Neural Network as foundation to establish pumping character–classified model firstly to get the well characteristics of pumping wells. Then, this study calculates the pumpage by the classified Non-linear Pumpage-estimated model. This is the whole process of the estimation using Electricity Records.
　　In the part of pumpage–estimation models, this study utilizes the testing-well data of electricity consumption and pumpage in Chou-Shui River alluvial fan (from 2012 to 2014) to calassfy the pumping efficiency by K-means clustering into four categories, and conduct the parameter test of pumpage–estimation model by the datas of these four classied pumping wells. This study apply parameter verification optimization model that is a Non-Linear Programing model and adopt a package software -Lingo11 to get the answer and classify Non-Linear Estimated model. In addition, this study also derives the model parameter of non-classified Pumpage Estimation, and utilizes the two models to evaluate the pumpage and compare the results. The results show that the classified pumpage–estimation model is more accurate.
　　In the part of Artificial Neural Network, this study adopts Back Propagation Neural Network, Real-Time Recurrent Learning and Adaptive Network-Based Fuzzy Inference System. Then let motor horsepower, diameter of pipe and pumping head as input item and calassfied categories as output item. In addition, this study uses the result of K-means clustering as the output value of Artificial Neural Network model to establish three classified models of pumping character. Then, compare the results with non-classified model by Correlation Coefficient, Coefficient of Efficiency, and Root Mean Square Error. And also uses the results of these three models to calculate the pumpage of testing well. Then compare the results with the non-classified pumpage estimation model. The results show that the pumpage estimated by the three Artificial Neural Network methods are better than non-classified pumpage–estimation model.
　　This study applys Artificial Neural Network and classified Non-Linear pumpage Estimated models to Chou-Shui River alluvial fan and finds that the results estimated by electricity records is similar to groundwater hydrograph analysis. It means that the estimated results of this study are in reasonable range. Acoording to the results, the pumpage in Fangyuan Township and Erlin Township during dry season is higher, especially from February to April. As a result, it is important to monitor the over-pumping condition from February to April.
　　The method estimated by the electricity records is espically suitable to management of agricultural wells in Taiwan. The advantages of the method are that it can estimate the pumpage of single-well accurately and sum the total pumpage in different spaces. In other words, it can estimate the pumpage of single-well, a village, a township, and town and more accurate in pumping management. Besides, it can evaluate the pumpage of the well efficiently, and only need the pumpage every month, motor horsepower, diameter of pipe and pumping head of the pumping well. However, this method still have limit in time scale, because the datas of electricity consumption only have one record every month. As a result, this study only evaluates the pumpage monthly. In the future studys, it can refine the time scale to get more accurate results.

並列關鍵字

Chou-Shui River alluvial fan ； K-means clustering ； Back Propagation Neural Network ； Real-Time Recurrent Learning ； Adaptive Network-Based Fuzzy Inference System ； Electricity Record Estimation

參考文獻

李冠緯（2013）。利用經驗正交函數法檢定含水層水文地質參數。國立臺灣大學土木工程學研究所碩士論文。

林聖婷（2012）。濁水溪沖積扇補注量與抽水量空間分佈模式建立。國立臺灣大學土木工程學研究所碩士論文。

王韋勳（2012）。名竹盆地地下水流數值模式之建立與應用。國立臺灣大學土木工程學研究所碩士論文。

歐靚芸（2012）。結合聚類法與類神經網路發展颱風淹水預警系統。國立臺灣大學工學院土木工程學系碩士論文。

經濟部水利署（2006）。雲林科技大學水土資源及防災科技研究中心九十五年度工作推動計畫。臺北市。