- 97 views
- 100 downloads
Methodologies in Mechanistic and Machine Learning Modelling of Nitrous Oxide Emissions from Wastewater Treatment Processes
-
- Author / Creator
- Khalil, Mostafa Wagih Tawfik
-
Nitrous oxide (N2O), a potent greenhouse gas, significantly contributes to the carbon footprint of wastewater treatment plants (WWTPs), with a global warming potential 300 times that of CO2 and also notorious for its ozone-depleting effect. Modelling can be a valuable alternative to quantification of N2O emissions through monitoring campaigns. Traditional mechanistic models have been limited in their application by impractical calibration processes and a lack of validation for complex treatment systems such as reactors combining biofilm and flocculent biomasses and/or highly dynamic operations. Conversely, machine learning (ML) presents a promising avenue, leveraging abundant WWTP data. However, the scarcity of comprehensive methodological ML frameworks for environmental engineering and the complexity of ML challenge their acceptance among practitioners. This thesis bridges evaluating and refining both modelling approaches to enhance predictability and highlights the areas for future research focus.
In this thesis, two-pathway nitrification and a multi-step denitrification N2O model was adapted to an Integrated Fixed-Film Activated Sludge (IFAS) system operating within a Sequencing Batch Reactor (SBR), using data from a laboratory-scale experiment. The model, characterized by a one-dimensional (1-D) biofilm, underwent a two-step calibration process informed by sensitivity and identifiability analyses. While it achieved good alignment with experimental data, revealing the model's predictive capacity, its application to different operational conditions in the validation data illustrated limitations in generalization. Further, through extensive simulations (over 1500), the influence of dissolved oxygen (DO), temperature, and ammonia levels was explored. These simulations highlighted the critical role of temperature in setting optimal DO levels, crucial for balancing N2O emission reduction with enhanced ammonia removal efficiency.
Parallel to the mechanistic approach, this thesis pioneers a comprehensive ML methodology for N2O emissions modeling using a long-term dataset from a full-scale WWTP. Recognizing the need for an online monitoring tool that also supports decision making, the proposed approach emphasizes not just model accuracy, but it also considers model complexity, computational speed, and interpretability. Various algorithms, including k-Nearest Neighbors (kNN), decision trees, Deep Neural Networks (DNN), and ensemble learning models such as extreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), and random forest were evaluated. A novel adjustment of a parametric multivariate outlier removal method aligned with data distributions, minimizing data loss. An effective feature selection strategy optimized the balance between data acquisition, model performance, and complexity—cutting feature count by 40% without compromising accuracy. Additionally, an integrated method combining feature selection with hyperparameter optimization (HPO) was introduced, leveraging a Genetic Algorithm (GA), specifically NSGA-II, against the Nelder-Mead algorithm to navigate the intricate, nonlinear data landscape. This comparison underscored effectiveness of GAs in streamlining model complexity and enhancing performance, paving the way for the development of interpretable, computationally efficient ML tools especially for real-time applications.
This thesis reveals that while the applied N2O mechanistic model offers acceptable predictions within the operational schemes used for calibration, its applicability to varied settings is limited. Despite the complexity of calibration, mechanistic models emerge as indispensable tools for scenario analysis, enhancing design and planning with precise "what if" explorations. In this context, the study underscores the critical influence of temperature in guiding optimal DO setpoints for effective ammonia removal and emission mitigation. On the ML front, models like kNN and AdaBoost not only demonstrated high accuracy in long-term emission prediction but also challenged the assumed necessity for deep learning, offering simpler, yet effective alternatives.
The developed holistic modelling framework aids the application of ML models as practical tools for ongoing process monitoring, owing to their accuracy, lower complexity, and adaptability.
This thesis contributes to the field by developing methodologies of employing mechanistic and ML models for prediction and mitigation of N2O emissions. The thesis offers a methodological framework for N2O emission modelling using ML, with insights that hold potential for broader application in the field of wastewater treatment, equipping practitioners with a robust toolkit for addressing environmental challenges. -
- Graduation date
- Fall 2024
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.