Applicability of machine learning in predicting N₂O emission from wastewater treatment processes: a narrative review

Nitrous oxide (N₂O) is a greenhouse gas produced during wastewater treatment. Recent advancements in computer technology have facilitated the generation and collection of large-scale multi-source datasets, rendering machine learning (ML) a powerful tool to predict N₂O emissions from these processes. In the narrative review, the emission characteristics and pathways of N₂O from wastewater treatment processes have been reviewed, while summarizing the mechanistic models, ML methods, and hybrid models used for N₂O emission prediction. In particular, the performance of different models in predicting the N₂O emission flux (accuracy), corresponding pathways and key influencing factors has been analyzed. Support vector machine (SVM), random forest (RF), and artificial neural network (ANN) algorithms were the most commonly used N₂O emission prediction models with high performance (R² > 0.90), computational speed, and interpretability. Key influencing factors identified by these models were nitrogen compounds, DO, and C/N, which was consisted with the domain knowledge. Hybrid models of mechanistic and ML algorithms (e.g., Long Short-Term Memory) were superior to the respective individual components in predicting N₂O emission flux and pathways because of the fewer data requirements and higher interpretability. However, the issues of data availability, interpretability, and transferability challenge the applicability of ML models. Hence, further studies on performance improvement strategies (e.g., generative models, interpretable ML, and transfer learning) should be conducted. Nevertheless, the studied prediction methods are important for controlling global warming.