Machine Learning-Based Prediction and Interpretation of Aluminum and Iron Coagulant Dosages in Drinking Water Treatment

Accurate and responsive coagulant dosage control is essential for maintaining drinking water treatment performance. In this study, the multi-WTP ICR database was used to develop machine learning (ML) models for Al- and Fe-based coagulant dosage prediction. Four ML algorithms, including CatBoost, neural network (NN), random forest (RF), and support vector regression (SVR), were systematically evaluated, and SHapley Additive exPlanations (SHAP) were employed to interpret model predictions. CatBoost and NN achieved comparable predictive performance for both Al and Fe coagulant dosages, followed by RF, whereas SVR showed the lowest predictive accuracy. SHAP analysis identified TOC, UV254, alkalinity, and total hardness as the major contributing features in both models. Preliminary testing using jar tests and data from 12 WTPs reported in published studies yielded MAPE values of 5.16% and 21.09%, respectively, for the CatBoost model predicting Al-based coagulant. An online platform integrating dosage prediction and SHAP interpretation was further developed to provide an interpretable, data-driven approach for coagulant dosage estimation and support more responsive dosing decisions in drinking water treatment.