Prediction and Evaluation of Indirect Carbon Emission from Electrical Consumption in Multiple Full-Scale Wastewater Treatment Plants via Automated Machine Learning-Based Analysis

Abstract

The indirect carbon emission from electrical consumption of wastewater treatment plants (WWTPs) accounts for large proportions of their total carbon emissions, which deserves intensive attention. This work proposed an automated machine learning (AutoML)-based indirect carbon emission analysis (ACIA) approach and predicted the specific indirect carbon emission from electrical consumption (SEe; kg CO2/m3) successfully in nine full-scale WWTPs (W1–W9) with different treatment configurations based on the historical operational data. The stacked ensemble models generated by the AutoML accurately predicted the SEe (mean absolute error = 0.02232–0.02352, R2 = 0.65107–0.67509). Then, the variable importance and Shapley additive explanations (SHAP) summary plots qualitatively revealed that the influent volume and the types of secondary and tertiary treatment processes were the most important variables associated with SEe prediction. The interpretation results of partial dependence and individual conditional expectation further verified quantitative relationships between input variables and SEe. Also, low energy efficiency with high indirect carbon emission of WWTPs was distinguished. Compared with traditional carbon emission analysis and prediction methods, the ACIA method could accurately evaluate and predict SEe of WWTPs with different treatment scales and processes with easily available variables and reveal qualitative and quantitative relationships inside datasets simultaneously, which is a powerful tool to benefit the “carbon neutrality” of WWTPs.