Nitrous oxide (N2O) emissions from wastewater treatment plants (WWTPs) exhibit significant seasonal variability, making accurate predictions with conventional biokinetic models difficult due to complex and poorly understood biochemical processes. This study addresses these challenges by exploring data-driven alternatives, using long short-term memory (LSTM) based encoder-decoder models as basis. The models were developed for future integration into a model predictive control framework, aiming to reduce N2O emissions by forecasting these over varying prediction horizons. The models were trained on 12 months and tested on 3 months of data from a full-scale WWTP in Amsterdam West, the Netherlands. The dataset encompasses seasonal peaks in N2O emissions typical for winter and spring months. The best performing model, featuring a 256-256 LSTM architecture, achieved the highest accuracy with test R2 values up to 0.98 across prediction horizons spanning 0.5 to 6.0 hours ahead. Feature importance analysis identified past N2O emissions, influent flowrate, NH4+, NOx, and dissolved oxygen (DO) in the aerobic tank as most significant inputs. The observed decreasing influence of historical N2O emissions over extended prediction horizons highlights the importance and significance of process variables for the model's performance. The model's ability to accurately forecast short-term N2O emissions and capture immediate trends highlights its potential for operational use in controlling emissions in WWTPs. Further research incorporating diverse datasets and biochemical process inputs related to microbial activities in the N2O production pathways could improve the model's accuracy for longer forecasting horizons. These findings advocate for hybridising deep learning models with biokinetic and mechanistic insights to enhance prediction accuracy and interpretability.
Read full abstract