Integrated Model for Understanding N₂O Emissions from Wastewater Treatment Plants: A Deep Learning Approach

Abstract

This study aims to demonstrate the application of deep learning to describe quantitatively longterm full-scale data observed from wastewater treatment plants (WWTPs) from the perspectives of process modeling, process analysis, and forecasting modeling. Approximately 750,000 measurements incorporating influent flowrate, air flowrate, temperature, ammonium, nitrate, dissolved oxygen, and nitrous oxide (N2O) during more than one year of Avedøre WWTP located in Denmark are utilized to develop deep neural network (DNN) through supervised learning for process modeling, and the optimal DNN (R 2 >0.90) is selected for further evaluation. For process analysis, global sensitivity analysis based on variance decomposition is considered to identify key parameters contributing to high N2O emission characteristics. For N2O forecasting, the proposed DNN-based model is compared with long short-term memory (LSTM), showing that the LSTMbased forecasting model performs significantly better than the DNN-based model (R 2 >0.94 and root mean squared error is reduced by 64 %). Results account for the feasibility of data-driven methods based on deep learning for quantitatively describing and understanding the rather complex N2O dynamics in WWTPs. Research into hybrid modeling concepts integrating mechanistic models of WWTPs (e.g., ASMs) with deep learning would be suggested as a further direction for the monitoring of N2O emissions from WWTPs.

Data-Driven Machine Learning in Environmental Pollution: Gains and Problems

“…A deep learning method such as RNN may play a role in this respect. The latest study used LSTM to process modeling and forecasting of N 2 O emission from WWTPs, and the obtained model showed a better performance in comparison to the previously reported DNN-based model ( R 2 > 0.94). The forecast of municipal solid waste (MSW) generation is also a time series analysis problem with high temporal variation.…”

Deep learning-based prediction of effluent quality of a constructed wetland

“…In this study, we divide data preprocessing methods into two parts: moving average and normalization. The moving average is a data smoothing method that is capable of smoothing high-frequency noise, and making the pattern more visible than original is required to ensure the stability of model performance [ 30 ]. The smoothing formula is shown in Equation (1) .…”

An Artificial Neural Network for Simulation of an Upflow Anaerobic Filter Wastewater Treatment Process

“…Instead of using sensitivity analysis, we selected the design parameters by ranking the results of simulations made using all possible permutations of three design parameters. In the context of the forecast modeling of N 2 O concentrations in wastewater treatment plants, Hwangbo et al [52] observed that an LSTM model outperformed a DNN model in terms of forecast prediction accuracy. They noted that the LSTM-based forecasting modeling depends on the response data, whereas all sensor data can be used as predictors in DNN models.…”