Black-box modelling of non-stationary N2O dynamics in a full-scale wastewater treatment plant

Using a data-driven approach, we present and compare linear and nonlinear methods for system identification of the potent greenhouse gas, nitrous oxide (N2O), which is produced during the biological treatment of wastewater. N2O is challenging to estimate, as the full understanding of its production process is yet to be determined. Therefore, data-driven approaches hold promise in advancing our understanding and offering solutions for model-based control, fault detection, and analysis. We present two methods for modelling the N2O in a full-scale wastewater treatment plant; the long short-term memory (LSTM) and a linear ARX model and discuss the performance of these models on real-world implementations. Results indicate that the nonlinear LSTM model has enhanced performance when compared to the linear ARX. While single-step predictions exhibit minimal mean squared error (MSE), the time-invariant models struggle to capture the production mechanisms over multi-step predictions due to the excessive need of multi-year data and non-stationarity and non-normality of the predicted variable.