Multivariate and Machine Learning-Based Assessment of Emerging Contaminants in Urban Water Supply Systems

Abstract

This study quantified the occurrence, multi-stage removal, and controlling factors of micro-organic pollutants in two large surface-water treatment systems in Saigon (WTP1) and Dong Nai (WTP2) rivers, using experiments and predictive machine learning models across 219 samples from raw water intake to peripheral distribution points. Raw water DOC differed strongly between the two sources (median 3.32 versus 1.75 mg/L, p < 0.001). Conventional treatment removed 33.3% DOC at WTP1 and 40.0% at WTP2, 77.0% and 43.1% of UV₂₅₄, and reduced chlorophyll-a and cyanobacteria to below detection in finished water at both plants, while turbidity removal exceeded 98.8%. Correlation analysis showed that treatment stage was the dominant determinant of DOC removal (r = 0.80), whereas raw water turbidity, chlorophyll-a, and cyanobacteria were negatively associated with it (r = -0.68, -0.57, and -0.51). Principal component analysis condensed thirteen variables into five components explaining 83.5% of the total variance, with PC1 (38.6%) representing an organic and particulate loading gradient. Among three predictive models, XGBoost performed best for DOC removal efficiency (5-fold R² = 0.825, RMSE = 7.33) and for treated-water DOC concentration (R² = 0.933, RMSE = 0.23 mg/L), clearly outperforming support vector regression and linear regression. SHAP analysis identified treatment stage and electrical conductivity as the dominant predictors. Benchmarking also compared these results with the requirements of the US EPA Stage 1 D/DBPR. The results provide a transferable, surrogate-based framework for monitoring and controlling micro-organic pollutants in urban water supply systems.