Machine Learning Based Prediction of Nitrous Oxide Emissions in Biological Wastewater Treatment using Multi-Decadal Datasets

ABSTRACT Global warming remains a critical issue, with nitrous oxide (N2O) emissions from wastewater treatment (WWT) systems contributing significantly. N2O measurements are limited, making data-driven predictions essential for mitigation. In this study, we quantify N2O emissions from BNR systems using a USEPA-endorsed protocol and developed an eXtreme Gradient Boosting (XG Boost) machine learning model for prediction and mitigation of N2O. The SHapley Additive exPlanations (SHAP) and partial dependence plot (PDP) highlighted ammonia (NH3), dissolved oxygen (DO), nitrite (NO2), and solids retention time (SRT) as key drivers. These insights enable emission reduction and process optimization, supporting resilient and sustainable wastewater treatment systems. INTRODUCTION Nitrous oxide (N2O) is a potent greenhouse gas (GHG) and ozone-depleting substance, with wastewater treatment plants (WWTPs) increasingly recognized as substantial anthropogenic sources(Ahn et al., 2010; Ravishankara et al., 2009). N2O is produced during the nitrification through nitrifier nitrification and nitrifier denitrification pathway, and in denitrification, through nitrifier denitrification and inhibition of N2O reductase(Rassamee et al., 2011). Machine learning (ML) models offer valuable insights into process performance and drivers of N2O emissions, with supervised ML approaches showing strong potential for predicting emissions in biological nutrient removal (BNR) systems(Khalil et al., 2023). However, existing models are limited by minimal datasets and lack interpretability(Khalil et al., 2023). N2O emissions are directly linked to nitrogen dynamics and serve as an indicator of BNR performance. This study integrates both historical and newly quantified data with supervised ML to capture interactions among operational conditions and N2O emissions, providing actionable insights for optimizing WWTP performance. MATERIALS AND METHODS Gaseous N2O concentrations from BNR processes were quantified using an USEPA-endorsed protocol. Measurements employed the Surface Emission Isolated Flux Chamber (SEIFC), the only EPA-approved floating body designed for measuring gaseous nitrogen fluxes from activated sludge tanks(Ahn et al., 2010). N2O concentrations were recorded at 1/min using an infrared gas filter correlation analyzer (Teledyne API, San Diego, California), while off-gas flow was simultaneously measured with a hot-wire thermo anemometer to determine fluxes in aerated and non-aerated zones. Each location was monitored continuously for 24 hours to capture diurnal variability. Data were separated into aerobic and anoxic zones, pre-processed to address missing/non-numeric values, log-transformed and min-max normalized [0,1]. The dataset was split into 75% training and 25% testing, with k-fold cross-validation for hyperparameter tuning. Extreme Gradient Boosting (XG Boost) models were developed and interpreted using Shapley Additive exPlanations (SHAP) and Partial Dependence plot (PDP) in R Studio. RESULTS AND DISCUSSION Interpretable SHAP analysis was used to uncover both the magnitude and directionality of variables influencing predicted N2O flux. Among all plant operational variables trained, the key variables driving N2O emissions are shown in Figure 1A-G. In the aerobic zone, SHAP results indicate that localized, zone-specific variables such as ammonia (NH3), dissolved oxygen (DO), and nitrite (NO2) positively influence increased N2O flux at high concentrations (Figure 1A-C). This underscores the importance of optimizing DO levels in conjunction with appropriate NH3 concentrations to sustain effective nitrification(Brotto et al., 2015). The global variable, solids retention time (SRT), shows that SRT values below 10 days negatively influence reduced N2O flux (Figure 1D). For the anoxic zone, SHAP results demonstrate that local variables like DO and NO2 positively influence N2O flux at high concentrations (Figure 1E,F), potentially linking to DO-mediated inhibition of N2O reduction to N2. A bi-phasic response of N2O flux to SRT was observed (Figure 1G), suggesting possible interactions with DO and NO2 concentrations (Ahn et al., 2010). The process optimization approach identified operating ranges for SRT as the principal process parameter, coupled with process performance response that concurrently minimizes off-gas N2O fluxes, using PDP analysis. In the aerobic zone, Figure 2A shows that NH3 concentrations < 4 mg L-1 and DO concentrations < 3 mg L-1 at SRT < 15 days resulted in less N2O flux. However, higher NH3 concentrations under both low and high DO conditions produced and emitted N2O (Figure 2A). This indicates that N2O production occurs through both nitrifier nitrification and nitrifier denitrification pathways in the aerobic zone. Thus, optimizing DO and SRT in conjunction with NH3 concentration is critical to improving process performance and reducing N2O fluxes. In the anoxic zone, the conditions that minimize N2O flux required maintaining DO concentrations < 1 mg L-1 coupled with a total process SRT < 20 days (Figure 2B). However, N2O flux increased when operating DO concentrations exceeded 1 mg L-1 under the same SRT conditions (Figure 2B). At DO concentrations > 1 mg L-1, NO2 accumulation in the anoxic zone was promoted, which in turn elevated N2O fluxes (Figure 2B)(Ahn et al., 2010; Rassamee et al., 2011). The results highlight the importance of strict DO control in anoxic zones to prevent NO2 buildup and mitigate emissions. CONCLUSION This study established ML as a powerful framework for disentangling the complex interactions governing N2O emissions from BNR systems. By integrating real-time and historical data, coupling supervised models with interpretable tools such as SHAP and PDP analyses, we quantified the relative contributions of operational and water quality parameters and identified conditions that exacerbate or suppress emissions. Results demonstrate that aerobic zone N2O fluxes are primarily driven by NH3 dynamics, with DO exerting a nonlinear influence that highlights the sensitivity of nitrification pathways to aeration control. Conversely, anoxic zone N2O fluxes were strongly modulated by DO and NO2, reflecting incomplete denitrification and inhibition of N2O reductase under suboptimal conditions. The observed biphasic response of SRT across both zones underscores its central role in shaping microbial activity and nitrogen transformation. Optimization analyses further delineated operational windows that minimize emissions while sustaining treatment performance. Aerobic zones operated at DO < 3 mg L-1 and NH3 concentration < 4 mg L-1 with SRT below 15 days, and anoxic zones maintained under strict oxygen limitation of DO < 1 mg L-1 with SRT under 20 days. These findings directly link predictive modeling outcomes to process-level strategies, providing actionable pathways for emission mitigation. Beyond prediction accuracy, the integration of ML with bioprocess fundamentals advances a mechanistic understanding of N2O dynamics, enabling data-driven optimization of wastewater treatment. The framework presented here illustrates how interpretable ML can support regulatory compliance, guide plant-wide control strategies, and ultimately reduce the climate and environmental footprint of wastewater treatment infrastructure.