Predicting N2O Emissions in Full-Scale Activated Sludge Systems Through Mechanistic Approach and Machine Learning: Heading Toward Generalized Model Structure Development

This study focuses on the assessments of N2O emissions during the operation of full-scale wastewater treatment plants (WWTPs). Using comprehensive measurements from two large WWTPs, simulation models were developed, employing both mechanistic (MCM) and machine learning (ML) approaches. The feasibility and limitations were explored for developing a generalized model with a transferable structure across different WWTPs. Three scenarios of input data were studied, including routine measurements, extended data set, and extended data set with a delay term. A hybrid approach was applied to support ML models with data generated by MCM. High consistency between measurements and ML model predictions was achieved, confirmed by the coefficient of determination equal to 0.71 - 0.76, 0.80 - 0.84, and 0.90 – 0.92 in the three scenarios, respectively. A key finding was that the extended data set improved the accuracy of multilayer perceptron model predictions while the number of neurons was insignificant. In the case of the K-nearest neighbors model, the number of neighbors did not affect results of simulation significantly regardless of the input data scenario. The results confirmed the potential for applying the same model structure at another WWTP but recalibration of model parameters was necessary. The ML models based on the hybrid approach can be further used to develop N2O mitigation strategies in full-scale WWTPs.