(345e) Prediction of Nitrous Oxide Concentration in Wastewater Treatment Plants: Application and Benchmarking of Deep-Learning Models

Wastewater treatment plants provide sanitation and clean water. As such they are primarily regulated with respect to effluent wastewater discharge limits to receiving water bodies. However, some may contribute to global climate change in form of N2O emissions (a potent greenhouse gas) if not operated properly. Indeed the anthropogenic emissions of these gasses have increased substantially over the last 50 years rising from 27 GtCO2-eq/yr to 49 GtCO2-eq/yr [1]. While Carbon dioxide accounts for over 60%, nitrous oxide accounts for 6%. Despite the smaller share of the total GHG emissions, nitrous oxide has a global warming potential (GWP) that is 265 times higher than Carbon dioxide and can persist in the atmosphere for more than 100 years, making it a very potent GHG [1]. Nitrous oxide also disrupts the photolysis of Oxygen responsible for protecting the planet from harmful ultraviolet light from the sun by regenerating the stratospheric ozone [2]. Most other ozone-depleting substances have either been banned or heavily regulated through various regulations such as the Montreal Protocol, which leaves nitrous oxide as the most dominant ozone-depleting substance in the 21^st century [3]. Wastewater treatment systems, which are primarily designed to provide sanitation, clean water and thus prevent eutrophication, have the potential for the production and emission of nitrous oxide [4] through underlying microbial activities. Wastewater treatment plants (WWTPs) are considered major contributors to nitrous oxide emission where a 0.5% emission factor would correspond to the indirect carbon dioxide emission of the energy consumption in conventional biological nutrient removal plants [5][6]. However, reports suggest the dynamic emission factors for nitrous oxide from WWTP vary considerably from plant to plant from being negligible to significant as high as 14.6% [7]. Therefore monitoring and understanding nitrous oxide emissions become an important objective to achieve sustainable wastewater treatment technologies, especially considering that the CO2 footprint of energy becomes smaller.

Mathematical models have been applied to model the microbial activity in WWTP in the form of Activated Sludge Models (ASMs). The original ASM models have been modified and expanded in the past decades to include additional microbial conversion processes such as single-pathway and two-pathway models for N2O production [8], [9] [10] [11]. However, these efforts resulted in very complex mathematical models, which are too difficult to apply in practice (due to the high number of model parameters that need to be calibrated, lack of sufficient data (quality and quantity) to validate the model predictions, etc) [9]. With the recent developments within the field of deep-learning and the resurgence of data-driven models, it has been possible to construct models capable of capturing the high non-linearity and interaction between process variables in WWTP. Recent studies have demonstrated the feasibility of such models in describing nitrous oxide emissions[12]. Recently, a framework for plant data-driven process modeling has been proposed by combining deep-learning with Monte-Carlo simulations[13]. The framework was applied to help to understand the plant-wide process of a real-life WWTP by combining domain-knowledge, real-life data collected from a full-scale WWTP (Avedøre WWTP which treats large amounts of domestic wastewater for 350,000 population equivalent (PE)), and advanced analytics such as global sensitivity analysis [13].

In this work, we test and expand the deep learning framework to large datasets obtained from two different full-scale WWTPs that have fundamentally different design and operation characteristics. Moreover, we also use the framework to test the added value of non-traditional online data for monitoring and correlating with N2O emissions including pH and conductivity sensors. It is noted these sensors (especially conductivity) are not amenable for first-principles modeling. In the deep learning models, we investigate therefore the inclusion of various measurements located upstream such as the conductivity and pH on the quality of predicting liquid nitrous oxide concentration by evaluating the goodness of fit and conducting variance decomposition-based global sensitivity analysis to draw process insights. Based on the newly gained process insight, only impactful process variables are used to construct deep-learning-based forecasting models based on architectures such as Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU). The proposed models are trained and validated using full-scale WWTP data collected from Kalunborg Central Wastewater treatment plant (KCR) and Ornum Wastewater treatment plant (OR) and benchmarked. These models have been trained on over 50,000 data points collected over 3 months and will be used to guide the better operation of the plants. The two plants are located in the northwestern part of Zealand and have a capacity of app. 50,000 and 23,000 PE respectively.

References

[1] “IPCC Climate Change 2014: Synthesis Report.,” in Contribution of Working Groups I, II and III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, 2014.

[2] R. W. Portmann, J. S. Daniel, and A. R. Ravishankara, “Stratospheric ozone depletion due to nitrous oxide: influences of other gases,” Philos. Trans. R. Soc. B Biol. Sci., vol. 367, no. 1593, pp. 1256–1264, May 2012.

[3] A. R. Ravishankara, J. S. Daniel, and R. W. Portmann, “Nitrous Oxide (N2O): The Dominant Ozone-Depleting Substance Emitted in the 21st Century,” Science (80-. )., vol. 326, no. 5949, pp. 123–125, Oct. 2009.

[4] J. H. Ahn, S. Kim, H. Park, B. Rahm, K. Pagilla, and K. Chandran, “N 2 O Emissions from Activated Sludge Processes, 2008−2009: Results of a National Monitoring Survey in the United States,” Environ. Sci. Technol., vol. 44, no. 12, pp. 4505–4511, Jun. 2010.

[5] D. W. De Haas and H. K, “Greenhouse gas emissions from BNR plant - Do we have the right focus,” in Proceedings of EPA Workshop: Sewage Management: Risk Assessment and Triple Bottom Line, 2004, no. April.

[6] L. Ye, B.-J. Ni, Y. Law, C. Byers, and Z. Yuan, “A novel methodology to quantify nitrous oxide emissions from full-scale wastewater treatment systems with surface aerators,” Water Res., vol. 48, no. 1, pp. 257–268, Jan. 2014.

[7] M. R. J. Daelman, E. M. van Voorthuizen, U. G. J. M. van Dongen, E. I. P. Volcke, and M. C. M. van Loosdrecht, “Seasonal and diurnal variability of N 2 O emissions from a full-scale municipal wastewater treatment plant,” Sci. Total Environ., vol. 536, pp. 1–11, Dec. 2015.

[8] E. Zaborowska, X. Lu, and J. Makinia, “Strategies for mitigating nitrous oxide production and decreasing the carbon footprint of a full-scale combined nitrogen and phosphorus removal activated sludge system,” Water Res., vol. 162, pp. 53–63, Oct. 2019.

[9] G. Sin and R. Al, “Activated sludge models at the crossroad of artificial intelligence—A perspective on advancing process modeling,” npj Clean Water, vol. 4, no. 1, p. 16, Dec. 2021.

[10] K. E. Mampaey, M. Spérandio, M. C. M. van Loosdrecht, and E. I. P. Volcke, “Dynamic simulation of N2O emissions from a full-scale partial nitritation reactor,” Biochem. Eng. J., vol. 152, no. August, p. 107356, Dec. 2019.

[11] F. Schreiber, P. Wunderlin, K. M. Udert, and G. F. Wells, “Nitric oxide and nitrous oxide turnover in natural and engineered microbial communities: biological pathways, chemical reactions, and novel technologies,” Front. Microbiol., vol. 3, no. OCT, pp. 1–24, 2012.

[12] S. Hwangbo, R. Al, X. Chen, and G. Sin, “Integrated Model for Understanding N2O Emissions from Wastewater Treatment Plants: A Deep Learning Approach.,” Environ. Sci. Technol., vol. 55, no. 3, pp. 2143–2151, Feb. 2021.

[13] S. Hwangbo, R. Al, and G. Sin, “An integrated framework for plant data-driven process modeling using deep-learning with Monte-Carlo simulations,” Comput. Chem. Eng., vol. 143, p. 107071, Dec. 2020.