Data-driven versus conventional N<inf>2</inf>O EF quantification methods in wastewater; how can we quantify reliable annual EFs?

Abstract

Copyright © 2020 The Authors. A long-term N2O dataset from a full-scale biological process was analysed for knowledge discovery. Non-parametric, multivariate timeseries changepoint detection techniques were applied to operational variables (i.e. NH4-N loads) in the system. The majority of changepoints, could be linked with the observed changes of the N2O emissions profile. The results showed that even three-day sampling campaigns between changepoints have a high probability (>80%) to result to an emission factor (EF) quantification with ~10% error. The analysis revealed that support vector machine (SVM) classification models can be trained to detect operational behaviour of the system and the expected range of N2O emission loads. The proposed approach can be applied when long-term online sampling is not feasible (due to budget or equipment limitations) to identify N2O emissions “hotspot” periods and guide towards the identification of operational periods requiring extensive investigation of N2O pathways generation.