Hybrid Modeling and Diagnosis to Reduce N2O Emissions at WRRFs

Introduction
A key challenge in monitoring N<SUB>2</SUB>O at WRRFs is managing and analyzing large amounts of data to develop mitigation insights and inform operations. This paper presents a unique hybrid modeling approach that combines data-driven models with process knowledge to predict liquid-phase N<SUB>2</SUB>O concentrations. The tool has been successfully applied to two full-scale, long-term monitoring datasets; information about the two facilities is summarized in Table 1, and Figures 1 and 2.

The focus of the project was identifying drivers for N<SUB>2</SUB>O generation using diagnostic tools and not simply fitting a model to data. The project aimed at answering the following questions:
- Why did a specific N<SUB>2</SUB>O peak happen, and which variables are associated with it?
- Which operational settings should be changed to reduce or prevent an N<SUB>2</SUB>O peak?
- Is there a model that describes the correlation between measured variables that can explain N<SUB>2</SUB>O behavior and can be used to mitigate N<SUB>2</SUB>O emissions?

Methodology
Proprietary code was developed and applied to automate the building and training of a PLS (Projection of Latent Structures) model via the NIPALS (Non-linear Iterative Partial Least Squares) algorithm. The model is integrated with diagnostic tools that objectively evaluate model performance, guide model iterations, and identify effective control handles to mitigate N<SUB>2</SUB>O emissions. A key statistic used was the percent sum of squares explained in the input (%SSX — the goodness of fit for input data) and output (%SSY — indication of model output consistency) (Figure 3). Variable Importance Plots (VIP) were used to rank the most information-rich variables and identify which variables could be removed (Figure 4). Contribution plots were used to further analyze specific events by ranking the variables with the highest contributions to the result data point (Figure 5) - this ability to examine specific events is a key advantage of the PLS modeling approach.
A stepwise and iterative procedure was applied to develop a robust and reliable hybrid model (Figure 6).
A typical iteration identifies that the model is not providing a good fit for a specific section of the data, which would trigger a discussion among domain experts and typically lead to the identification of data issues or information to add to the model inputs. This iterative procedure guided the project team in adding domain expertise in process and control. Examples are the translation of operational information (e.g., logbooks, operator interviews) into a digital format to feed into the model (i.e., removing data for events such as sensor cleaning/calibration, or unreliable data based on operator insights).

Duffin Creek WPCP Results
The performance of the first model (%SSX: 28.9) was improved with every iteration with the last model iteration explaining 72% of the data (Figure 7). Key improvements were data cleaning (based on plant information, statistical analysis, and identification of a faulty temperature sensor for the N<SUB>2</SUB>O probe (Figure 8)), and the addition of domain expertise such as influent load calculations and feeding DO control errors (Figure 9). The final model achieves a very good fit and follows diurnal and seasonal variations, with an average tank N<SUB>2</SUB>O Emission Factor (EF) of 0.7%.

The diagnostic tools identified aeration control as the main driver for N<SUB>2</SUB>O accumulation. An aeration improvement project is currently ongoing and the impact on N<SUB>2</SUB>O will be closely monitored. The second highest ranking could be linked to oxygen ingress in the pre-anoxic zone during mixer maintenance; the plant has been advised to minimize airflows into the pre-anoxic zones.

Elmira WWTP Results
Model iterations improved the goodness of fit from 43.8% to 64.9% SSX explained (Figure 11). Key improvements were achieved by cleaning data and removing variables without relevant information. The largest improvement was from removing the datasets where the performance was disrupted during MABR commissioning and MABR aeration system failure due to cold weather. Results indicate relatively low N<SUB>2</SUB>O emissions from the bioreactors downstream of the MABR with occasional spikes, with the average N<SUB>2</SUB>O EF at 0.42%.

Diagnosis of the results showed aeration as the main contributor to N<SUB>2</SUB>O accumulation and improvements have been recommended. A well-operated MABR seems to keep the downstream N<SUB>2</SUB>O emissions low, however, the MABR off-gas could contribute an additional 30% to the N<SUB>2</SUB>O emissions based on limited off-gas data. Other N<SUB>2</SUB>O peaks could be linked to wet weather events and it has been recommended to investigate controls and process optimization during and after such events.

Summary and Next Steps
The hybrid models were able to explain 65 to 70% of the measured N<SUB>2</SUB>O and accurately predicted the trends and diurnal variations. The model allows for multivariate analysis to identify key operational parameters with the highest importance on the measured N<SUB>2</SUB>O, which combined with process knowledge, enables extraction of actionable information to develop potential operational strategies to mitigate N<SUB>2</SUB>O emissions. At the Duffin Creek WPCP, recommended improvements to the aeration control system will be implemented in spring 2025 together with ammonia-based aeration control, providing the opportunity to optimize aeration energy efficiency while minimizing process N<SUB>2</SUB>O emissions; some of the mitigation insights identified by the model will be tested.