Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


Atie Amirgol-Dissertation.pdf (7.46 MB)

ETD Abstract Container

Abstract Header

Evaluating Data-Driven Optimization Options for Dissolved Organic Carbon Treatment by Coagulation and Powdered Activated Carbon

Amirgol, Atie
http://rave.ohiolink.edu/etdc/view?acc_num=akron1627734124517122

Abstract Details

2021, Doctor of Philosophy, University of Akron, Civil Engineering.
Optimizing dissolved organic carbon (DOC) removal during drinking water treatment is important for many reasons. Specifically, DOC impacts the efficiency of many treatment processes and the DOC remaining after treatment creates disinfectant demand and is a precursor for the formation of disinfection by-products (DBPs). Variable disinfectant demands make it challenging for water utilities to maintain adequate residuals for protecting the public against waterborne diseases and DBPs are known carcinogens, so reducing their concentration in public drinking water supplies is a valid endeavor. Effective DOC removal generally involves multiple treatment processes and chemicals, creating an opportunity to use operational data to investigate optimization opportunities. In the last ten years, artificial neural networks (ANN) have been used for multi-objective optimization of water treatment processes with the goal of minimizing the cost of treatment while providing “high” quality drinking water to the consumer. ANNs can satisfy the need for general applicability without requiring extensive experimental efforts. Despite documented and realized advantages, ANNs come with a few drawbacks which may deter potential users. For instance, they perform as a “black box” (i.e. lack functional form) and do require some level of modeling sophistication. The purpose of this research was to evaluate optimization of dissolved organic carbon (DOC) removal by coagulation and powdered activated carbon (PAC) applying ANN models to operational data (more than three years) for the Akron Water Treatment Plant (WTP) in Akron, Ohio. To achieve this purpose, ANN models were built to predict UV254 removal through the conventional treatment train at Akron WTP which consists of pre-oxidation, PAC adsorption, coagulation, filtration and disinfection. ANN models were evaluated and compared based on their descriptive statistics (mean squared error, coefficient of correlation and ρ). Model results were generally good and the coefficient of correlation for the final models were 0.869-0.876. Variance-based global sensitivity analysis of all ANN models was performed using Sobol sensitivity analysis to gain insight on uncertainty of the model output. To practice feasibility of applying more user-friendly alternatives to ANN models, linear models were developed but the only comparable linear models were achieved by incorporating the sensitivity analysis results of the ANN models for selecting the first order interaction terms. To compare ANNs and their correlated MLR models with published/accepted and experimentally derived models, Edwards coagulation model for DOC removal and, adsorption isotherm and kinetic models for PAC adsorption were considered. Edwards coagulation model was calibrated for operational data from Akron WTP for period of February 2016 to October 2020 based on days where only coagulation process was applied for removal of organics and PAC was not used, with coefficient of correlation 0.9. To evaluate removal of DOC through adsorption by PAC, three PACs were tested on water from Akron WTP for removal of two fluorescence components with humic fluorophore nature (C1 and C2) and UV absorbance at 254 (UV254) at multiple doses (5, 10, 15, 20 and 40 mg/L) and contact times (5, 10, 15, 30, 60, 120 and 240 minutes). Adsorption capacity and kinetics were studied calibrating Freundlich isotherm and, pseudo-first and pseudo second-order models, respectively. Pseudo-second order performed better than pseudo-first order where for C1 and C2, coefficient of correlation and ρ ranged between 0.87-0.99 and 0.018-0.121 respectively, and for UV254, coefficient of correlation and ρ ranged between 0.56-0.98 and 0.013-0.146 respectively. Dose-based power regression models were developed with coefficient of correlation varying between 0.88-0.99. Comparability of models for simulation of selected parameters was addressed in two steps. First, a combination model consisting of ANN that only considered coagulation process and the dose-based PAC adsorption model which was adjusted based on the initial variability of the effective parameters was evaluated. A model was introduced that can predict removal of UV254 based on the PAC dose applied, initial concentration of UV254 and the contact time. Second, calibrated Edwards model and adjusted dose-based PAC adsorption model were combined to predict UV254 removal through coagulation and PAC adsorption. Results of comparison of these two sets with an ANN based on coefficient of correlation and ρ showed that a stand-alone ANN model performs better than combination of published/accepted and experimentally derived models. For optimizing settled UV254 and the associated costs, a bi-objective optimization problem was defined. In this problem, final ANN simulation models along with their corresponding linear model and a linear model with no interactions were optimized by applying NSGAII genetic algorithms. Models were selected as the objective functions for optimizing settled UV254 and the cost function was defined based on the chemical dose settings: pre-oxidants like chlorine dioxide and potassium permanganate, PAC and coagulant. To fairly compare the optimization solutions, the hypervolume indicator was used; Pareto optimal fronts for the optimization models were developed, a reference point was selected slightly greater than the maximum possible value of the objectives on any given day and the area covered between the Pareto optimal front and the reference point was calculated for each model. To have a comprehensive evaluation of the hypervolumes, solutions were developed for 207 days. Comparison of the distribution median for the hypervolumes confirmed that MLR models provide more optimization opportunity than corresponding ANN waters. This research showed that MLR models in which the interaction terms are selected based on variance-based sensitivity analysis of ANN models provide more optimization opportunity compared to the ANN models when multiple objectives are addressed in water treatment.
Christopher Miller (Advisor)
David Roke (Committee Member)
Qindan Huang (Committee Member)
Ping Yi (Committee Member)
Richard Einsporn (Committee Member)
217 p.
Civil Engineering; Environmental Engineering
Powdered Activated Carbon; Machine Learning; DOC Treatment; Artificial Neural Networks

Recommended Citations

Citations

  • Amirgol, A. (2021). Evaluating Data-Driven Optimization Options for Dissolved Organic Carbon Treatment by Coagulation and Powdered Activated Carbon [Doctoral dissertation, University of Akron]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=akron1627734124517122

    APA Style (7th edition)

  • Amirgol, Atie. Evaluating Data-Driven Optimization Options for Dissolved Organic Carbon Treatment by Coagulation and Powdered Activated Carbon. 2021. University of Akron, Doctoral dissertation. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=akron1627734124517122.

    MLA Style (8th edition)

  • Amirgol, Atie. "Evaluating Data-Driven Optimization Options for Dissolved Organic Carbon Treatment by Coagulation and Powdered Activated Carbon." Doctoral dissertation, University of Akron, 2021. http://rave.ohiolink.edu/etdc/view?acc_num=akron1627734124517122

    Chicago Manual of Style (17th edition)

akron1627734124517122
54
© 2021, all rights reserved.
This open access ETD is published by University of Akron and OhioLINK.