Carbon Emission Characteristics Research of Typical Drinking Water Treatment Plants in South China Based on Machine Learning Models

Abstract

Accurate calculation of carbon emissions is essential for the water industry to achieve low-carbon development. In this study, carbon emissions from a typical drinking water treatment plant (DWTP) in South China and the carbon emissions of each segment were accurately calculated. The results showed that the average annual carbon emissions during the period from 2018 to 2024 were 5263.294 t CO2-eq, with a carbon intensity of 0.1458 kg CO2-eq·m-3. The main sources of carbon emissions from the DWTP were electricity consumption (78.65%) and chemicals consumption (21.05%), with the pumping station electricity consumption contributing the most (74.80%). Two machine learning methods, extreme gradient boosting (XGBoost) and Random Forest (RF), were employed to construct models based on the carbon emissions and water quality parameters, to compare the prediction effects of the two models, and to identify the critical factors affecting carbon emissions. The results show that XGBoost exhibits a strong predictive capacity for both electricity carbon emissions intensity (ECES) and NaClO carbon emissions intensity (NaClOS), and RF performs better in predicting NaClOS. Thus the two models can be combined to predict the main sources of carbon emissions. The feature importance analysis shows that water supply is the most important factor affecting ECES. Water quantity and total bacterial colonies were important factors affecting NaClOS. Water quantity and residual chlorine possessed relatively higher importance for polymeric aluminum chloride (PAC) carbon emissions intensity (PACS). These findings demonstrate the high potential of using machine learning to predict the carbon emissions of DWTPs.

Keywords: drinking water treatment plant, Carbon emission accounting, Machine learning model, Electricity carbon emission, Chemical carbon emissions