以機器學習模型預測污水處理廠放流水中的化學需氧量

Abstract

隨著人口急速成長和都市化，生活污水量持續攀升，再加上工廠排放與農業灌溉帶來的複雜污染物，污水處理廠（WWTPs）面臨更大處理量與更複雜污染物的雙重挑戰。因此，如何有效監控並預測出流水水質成為當務之急。本研究結合現場感測器數據與機器學習模型，對一工業區污水處理廠出流水COD（COD.out）進行預測。首先對感測器數據進行清洗以排除異常值，並加入時間延遲分析來捕捉污水處理過程中的停留效應，接著比較隨機森林（Random Forest, RF）、梯度提升機（Gradient Boosting Machine, GBM）與極限梯度提升（Extreme Gradient Boosting, XGB）三種模型，結果顯示RF在預測COD.out的表現最佳，平均絕對百分比誤差 (Mean Absolute Percentage Error, MAPE) 為 6.22%。此外，以夏普利加成解釋 (SHapley Additive exPlanations, SHAP) 分析各輸入參數包含進水pH（pH.in）、進水溫度（Temp.in）、氧化渠溫度（Temp.Ox.ditch）、出水pH（pH.out）、出水溫度（Temp.out）、出水懸浮固體濃度（SS.out）對模型輸出的影響程度，結果顯示，放流池、氧化渠溫度（TEMP.out 和 TEMP.Ox.ditch）以及放流池懸浮固體濃度 (SS.out) 對COD.out的影響最為顯著。當Temp.Ox.ditch與Temp.out維持在27~32°C之間、SS.out低於2.5 mg/L時，模型預測的COD.out呈下降趨勢。透過重點監測並維持這三個關鍵參數可以有效預測COD.out。

As population grows rapidly and urbanization accelerates, the volume of domestic wastewater continues to increase. Coupled with complex pollutants from industrial discharges and agricultural irrigation, wastewater treatment plants (WWTPs) face the dual challenges of higher influent loads and greater pollutant complexity. Therefore, effective monitoring and prediction of effluent water quality in the WWTPs have become a crucial task. In this study, on-site sensor data from an industrial WWTP are combined with machine learning (ML) models to forecast effluent chemical oxygen demand (COD.out). First, sensor readings are cleaned to remove outliers, and time lag analysis is incorporated to capture the retention effects occurring throughout the treatment process. Subsequently, three models, random forest (RF), gradient boosting machine (GBM) and extreme gradient boosting (XGB), were trained and compared. RF delivered the best performance in predicting COD.out, achieving a mean absolute percentage error (MAPE) of 6.22%. SHapley Additive exPlanations (SHAP) analysis was employed to evaluate the influences of each input parameter, including influent pH (pH.in), influent temperature (Temp.in), oxidation ditch temperature (Temp.Ox.ditch), effluent pH (pH.out), effluent temperature (Temp.out), and effluent suspended solids concentration (SS.out) on the model output. The results indicate that the TEMP.Ox.ditch, TEMP.out, and SS.out have the most significant influences on COD.out. When Temp.Ox.ditch and Temp.out are maintained between 27 and 32 °C and SS.out is kept below 2.5 mg/L, the model predicts a declining trend in effluent COD. By focusing on monitoring these three key parameters, the COD.out can be effectively predicted.