Investigating carbon emissions in drinking water treatment plant through machine learning: A case study in South China

Accurate carbon emission assessment was vital for achieving low-carbon development in the water industry. This study aimed to quantify operational carbon emissions of a drinking water treatment plant (DWTP) in South China (2018–2024) and developed predictive models to identify mitigation strategies. Through IPCC emission factor methods, we calculated an average annual emission of 5263.29 t CO₂-eq and intensity of 0.1458 kg CO₂-eq·m −3 , with electricity (78.65 %, mainly pumping stations) and chemicals (21.05 %) as primary contributors. Three machine learning models, extreme gradient boosting (XGBoost), Random Forest (RF), and Artificial Neural Networks (ANN) were developed to predict emission intensity using water quantity and quality operational data. XGBoost outperformed RF and ANN in predicting the carbon emission intensity of the DWTP. Particularly for electricity and NaClO intensity, XGBoost achieved R 2 values of 0.919 and 0.776, respectively. Feature importance analysis identified water supply quantity and total bacterial colonies as critical drivers. Notably, this study compared performances of XGBoost, RF and ANN in addressing DWTP carbon prediction. Validation via cross-validation and test datasets confirmed model robustness. Our contributions included: (1) systematic identification of emission hotspots in South China DWTPs, (2) an appropriate machine learning (ML) framework enhancing prediction accuracy, and (3) actionable insights for optimizing energy and chemical use. This work introduced three ML algorithms to advance carbon prediction accuracy in water systems and provided actionable strategies for optimizing energy and chemical use, directly supporting China's dual-carbon policy.