Enhanced accuracy and interpretability of nitrous oxide emission prediction of wastewater treatment plants through machine learning of univariate time series: A novel approach of learning feature reconstruction

Abstract

Nitrous oxide (N₂O) is a greenhouse gas (GHG) emitted from wastewater treatment plants (WWTPs), but it is difficult to monitor by conventional methods. Machine learning (ML) have been proposed to predict N₂O emissions from WWTPs, on the other hand, univariate time series are widely used in various fields, but the existing shortcomings are difficult to solve, including the low performance and the difficulty of evaluating the interpretability of the model. In this paper, four models, random forest (RF), long short-term memory (LSTM), recurrent neural network (RNN), and support vector machine (SVM), are used to predict the 12-month N₂O monitoring data of a WWTP. Meanwhile, taking the dynamic characteristics of the original univariate time series as a multivariate time series. The results show that the R² of accuracy increases from 0.9768 to 0.9999, the R² of K-Fold Cross-Validation increases from 0.8966 to 0.9999 for the RF model with the best overall performance and SHAP analysis becomes feasible. The study offers a viable approach to enhancing the precision, resilience, and comprehensibility of the model for monitoring univariate time series in other WWTPs.