利用支撐向量迴歸預測老街溪水質

人口增長與都市發展對於河川水質具顯著影響，臺灣老街溪因鄰近居住區及工業區，水質易受家庭污水與事業廢水排放影響，故河川水質問題也備受重視。本研究目的為探討支撐向量迴歸(Support Vector Regression, SVR)模型模擬老街溪水質參數的適用性。研究材料包含2008年至2019年之環保署與環檢所的老街溪水質檢測資料與中央氣象局的氣象觀測資料，篩選檢驗水質參數資料與雨量資料以納入SVR模型模擬分析。首先以敘述統計初步探勘水質數據，依據現有地面水體分類及水質標準比較老街溪13個採樣點內水質污染情形，發現生化需氧量於11個採樣點均有超過基準值的情形，銅於10個採樣點均有超過基準值的情形，氨氮與錳於全部採樣點皆有超過基準值的情形，表示水質參數中生化需氧量、氨氮、銅和錳於老街溪中有較明顯污染情形。依序以Shapiro-Wilk 常態性檢定、克-瓦二氏單因子等級變異數分析(Kruskal-Wallis test)、Dunn事後多重比較檢定(Dunn post hoc test)與相關性分析方法分析老街溪水質特性與雨量關係，發現不同測站之水質參數普遍具有顯著差異，表示老街溪水質參數濃度具有空間差異性，且各測站除最上游美都麗橋測站外其餘測站檢測之銅、錳皆與生化需氧量、化學需氧量和氨氮有較高強度之正相關性(相關係數為0.34~0.84)，推測其可能為家庭污水與事業廢水排放造成老街溪污染之主要污染水質參數，故選擇較明顯污染之水質參數生化需氧量、氨氮、銅和錳此4種水質參數為SVR模擬目標。 SVR模型模擬方面，為了解其對於老街溪水質模擬之適用性，以老街溪水質測站分組各別比較eps-SVR與nu-SVR兩種模型公式及linear與radial兩種核函數之模擬結果，每測站共計4種模型組合結果，使用交叉驗證與網格搜索方式找出最適用之模型參數組合，納入模型模擬訓練資料與驗證資料中，最後以均方根誤差、絕對平均誤差、相關係數與確定係數評估SVR模擬效能。 SVR於各測站模擬生化需氧量、氨氮、銅和錳之模擬結果普遍較為良好，均方根誤差範圍約介於0.001~7.565，相關係數數值範圍約介於0.45~0.99。僅於平鎮一號橋測站模擬氨氮與於美都麗橋測站模擬銅和錳的模擬值與實測值擬合情形不佳，以相關係數評估模擬效能均為低相關性強度，推測可能為輸入參數數據相關性強度與數量不足，導致模擬結果未達理想結果。整體而言SVR模型適用於老街溪河川水體之水質模擬，使用較少數據資料量與分析時間成本，且分析步驟亦較少依賴水文水質專業知識，建議可使用SVR模型初步探勘研究資料稀缺之研究區域，以判別此研究區域模擬適用性，可輔助後續相關分析工作，並提供未來水質模擬之相關數據參考。

關鍵字

老街溪；水質參數；相關性分析；支撐向量迴歸

並列摘要

Due to population growth and urban development, the water quality of the Lao-Jie River is affected by domestic and industrial wastewater discharges, and thus its water quality became a priority concern in Taoyuan, Taiwan. This study aims to investigate the applicability of the Support Vector Regression (SVR) model to simulate the water quality of Lao-Jie River. This study obtained water quality data from Environmental Protection Agency (EPA) and Environmental Analysis Laboratory (EAL) from 2008 to 2019, along with data from meteorological observation data with Central Weather Bureau (CWB). Firstly, descriptive statistics were conducted to evaluate the water quality parameters in 13 sampling points of Lao-Jie River according to Taiwan Surface Water Classification and Water Quality Standards. Biological oxygen demand (BOD) exceeded the benchmark value in 11 sampling points, copper (Cu) concentrations exceeded the benchmark value in 10 sampling points, ammonia nitrogen (NH3-N) and manganese (Mn) exceeded the benchmark values in all sampling points. This study observed levels of BOD, NH3-N, Cu and Mn were significantly polluted in Lao-Jie River. This study used Shapiro-Wilk normality test, Kruskal-Wallis test, Dunn post hoc test and correlation analysis to analyze the relationships between water quality parameters of Lao-Jie River and precipitation. The water quality parameters have significant differences in all sampling points, except for the upstream station, namely Meidouli Bridge. The results also show high positive correlations (correlation coefficients: 0.34 - 0.84) among levels of Cu, Mn, BOD, COD and NH3-N in most sampling points, which may resulted from domestic and industrial wastewater discharges. The major pollution water quality parameters (BOD, NH3-N, Cu and Mn) were chosen as the SVR model targets because they are main pollution contributors to the Lao-Jie River. In order to understand the applicability of the SVR model for the water quality simulation of Lao-Jie River, nu regression (nu-SVR) and epsilon regression (eps-SVR) were used, with two kernel functions of linear and radial, to train and predict data in all sampling points. Cross validation and grid search were used to identify the most suitable combinations of model parameters setting, and root mean squared error (RMSE), mean absolute error (MAE), correlation coefficient (r) and the coefficient of determination (r2) were used to evaluate the SVR simulation performance. The simulation performance of BOD, NH3-N, Cu and Mn using SVR model was good (RMSE = 0.001~7.565, r = 0.45~0.99) except for the simulated Cu and Mn in Meidouli Bridge and NH3-N in Pingjhen Number 1 Bridge were poorly fitted to the measured values and the r values indicated a low correlation. The inference may be due to low correlation and quantity of SVR input data that may result in the poorly simulation results. In general, the SVR model is suitable for the water quality modelling for the Lao-Jie River that only requires less data, less time and cost for analysis, and less reliance on hydrological and water quality expertise in the analysis procedures. It is suggested that the SVR model can be used to do the preliminarily survey for the study area where the water quality data is limited, to assist the subsequent analysis, and to provide the relevant data for future water quality modeling.