應用機器學習預測污水處理廠活性污泥槽溶氧濃度及關鍵因子探討

Abstract

隨著全球水資源短缺問題的加劇，污水處理與智慧化技術的應用逐漸成為水資源管理的核心議題。溶氧濃度作為污水處理廠活性污泥槽（好氧槽）中反映曝氣效率與處理效果的重要指標，其準確預測對於提升污水處理效能至關重要。本研究利用機器學習技術建立好氧槽溶氧濃度預測模型，使用線性回歸（Linear Regression, LR）、隨機森林（Random Forest, RF）、梯度提升機（Gradient Boosting Machine, GBM）、極限梯度提升（eXtreme Gradient Boosting, XGB）以及支持向量回歸（Support Vector Regression, SVR）等多種模型，比較其預測性能，並探討數據品質與進流參數經過時間位移處理（例如使用前1小時的進流數據預測未來目標值）對模型表現的影響。以台灣中部某污水處理廠（簡稱中部A廠）連續30日的水質與操作數據為基礎，將數據分為10日組（高品質數據，經頻繁校正）與30日組（常規數據）進行分析與比較。結果顯示，10日組高品質數據相較於30日組能有效提升模型預測準確性，其中隨機森林模型在結合經滯後1小時之進流特徵（對進流數據進行時間位移處理，以使用前1小時的進流數據作為輸入）與其他槽體特徵的條件下表現最佳，經超參數調整後的R平方值達到0.975，展現出優異的預測性能。此外，本研究透過SHAP（SHapley Additive exPlanations）解釋模型的特徵重要性與交互作用，發現缺氧槽MLSS、硝化液迴流量、缺氧槽ORP等特徵對模型貢獻較大，強調數據分布特性對模型性能的重要影響，並利用 SHAP 進行特徵篩選，進一步簡化模型。 本研究強調數據品質控管與感測器校正對於機器學習模型應用的必要性，證明穩定且高品質的數據能顯著提升模型預測能力。同時，本研究提出以 SHAP 方法進行特徵篩選，提供污水處理廠智慧化管理與數據優化的新方向，期望實現高效、節能的污水處理目標。

With the intensification of global water scarcity, wastewater treatment and the application of intelligent technologies have gradually become pivotal topics in water resource management. Dissolved oxygen concentration, a critical indicator of aeration efficiency and treatment effectiveness in activated sludge tanks (aerobic tanks) of wastewater treatment plants, plays a vital role in improving treatment performance. This study utilizes machine learning techniques to develop a predictive model for dissolved oxygen concentration in aerobic tanks. By employing various models, including Linear Regression (LR), Random Forest (RF), Gradient Boosting Machine (GBM), eXtreme Gradient Boosting (XGB), and Support Vector Regression (SVR), it evaluates and compares their predictive performance. Furthermore, the study examines the influence of data quality and time-lagged influent parameters (e.g., using influent data from one hour earlier) on model accuracy. Based on 30 days of continuous water quality and operational data collected from a wastewater treatment plant in central Taiwan (referred to as Central A Plant), the dataset was divided into two groups for analysis and comparison: a 10-day set with high-quality, frequently calibrated data, and a 30-day set with regular data. The results indicated that the high-quality 10-day set improved model prediction accuracy compared to the 30-day set. Among the models tested, the Random Forest model achieved the best performance by combining influent features shifted by one hour (i.e., using influent data from one hour earlier as input) with features from other tanks. After hyperparameter tuning, the model achieved an R-squared value of 0.975, demonstrating excellent predictive performance. Furthermore, SHAP（SHapley Additive exPlanations） analysis was used to interpret feature importance and interactions within the model, revealing that features such as anoxic tank MLSS, nitrate recycle flow rate, and anoxic tank ORP contributed substantially to the model, highlighting the critical impact of the range and variability of data distribution on model performance. SHAP was also utilized for feature selection, further simplifying the model while improving interpretability. This study underscores the necessity of data quality control and sensor calibration in the application of machine learning models, demonstrating that stable, high-quality data can enhance predictive accuracy. Moreover, the study introduces SHAP-based feature selection not only as a novel method for identifying key contributors to model performance but also as a practical approach to simplifying models. These findings provide valuable insights into optimizing data and model management in smart wastewater treatment systems, aiming to achieve efficient and energy-saving operations.