Optimizing multi-barrier drinking water treatment through a data-driven process simulator based on machine learning

The growing variability and complexity of drinking water sources highlight the urgent need for data-driven tools to support sustainable and resource-efficient design of treatment systems. This study develops a machine-learning-based process simulator tailored for multi-barrier drinking water treatment. The simulator integrates global predictive models—LightGBM, XGBoost, and CatBoost—to forecast effluent quality from influent characteristics across multiple treatment configurations, including ozone-biological activated carbon (ozone –BAC) processes positioned either before or after sand filtration. Among these models, CatBoost demonstrated the highest overall predictive accuracy for COD_Mn, turbidity, pH, temperature, and residual chlorine using operational data from four full-scale water treatment plants, showing consistently strong agreement between predicted and observed values across all parameters. The simulator also achieved robust performance in bacterial classification, with a weighted average precision of 0.84. SHAP (Shapley Additive Explanations) analysis revealed distinct treatment efficiencies: placing ozone –BAC after sand filtration enhanced COD_Mn removal, while pre-sand filtration positioning favoured turbidity and bacteria removal. Real-world validation at two full-scale water treatment plants confirmed the simulator's ability to guide treatment optimization under poor influent conditions (e.g., COD_Mn > 7 mg/L), maintaining effluent quality below regulatory thresholds (<3 mg/L). This research demonstrates the potential of integrating machine-learning and process simulation to promote cleaner production, enhance treatment efficiency, and support sustainable drinking water management decisions.