A Comprehensive Study of Drinking Water Coagulation with Aluminum Sulfate
1 online resource (223 pages) : PDF
University of North Carolina at Charlotte
Coagulation is usually the first and most important step in conventional drinking water treatment processes. The efficiency of all downstream processes is directly dependent on the effectiveness of the coagulation stage. Coagulants such as aluminum sulfate (alum) have been used for treating drinking water for over a century now. Since the early 1900s, researchers have been studying coagulants in hopes of understanding the mechanisms by which they help remove contaminants from water. Despite the significant contributions and breakthroughs by many researchers, we still rely on a trial-and-error process (jar testing) to optimize coagulation. Accurately modeling the coagulation process has been nearly impossible because water is a chemically complex medium that varies spatially and temporally. There are also many competing and interacting factors that influence how coagulants interact with contaminants and the resulting overall treatment efficiency. This research aimed to develop an accurate computer model for coagulation with aluminum sulfate with practical, real-world applications. The study identified and addressed five primary challenges related to the study and modeling of coagulation. The five challenges were as follows: (1) independently control water quality parameters, (2) isolate the effects of coagulation dose and pH, (3) standardize the jar test procedure, (4) identify performance metrics that are scalable and independent of jar test mixing parameters, and (5) establish effective optimization strategies. A design of experiments approach was used to create 16 synthetic waters based on four water quality factors (dissolved organic carbon (DOC), specific ultraviolet absorbance (SUVA), alkalinity, and turbidity) at two discrete levels. An extensive dataset was built by measuring the performance at 1,632 unique combinations of water quality and coagulation conditions where all relevant coagulation factors – water quality, coagulation conditions, and mixing parameters – were tightly controlled. The measured performance metrics were settled and filtered water turbidity, DOC, UV254, zeta potential, and total chemical costs. Efforts to predict turbidity removals using simple regression models were unsuccessful. The regression models considered ranged from linear regression models with varying complexity to more advanced regression models based on machine learning such as support vector machines and gaussian process regression. When tested on a new water (i.e., one that was not used to train the model), the root-mean-square error (RMSE) of the predicted turbidity ranged between 28 – 37%, while the R2 values ranged between 0.41 – 0.64. It was apparent that regression models could not model all the complex underlying non-linear behaviors of the coagulation process. On the other hand, a deep neural network (DNN) trained on the same dataset produced acceptable results. The RMSE and R2 values of the trained neural network (on the test dataset) were 9.6% and 0.88, respectively. The user input parameters were only DOC, SUVA, alkalinity, and turbidity of the raw water. More importantly, the trained neural could generate a contour plot (a process that requires 17 jar tests to produce experimentally) that correctly predicted the size and shape of the effective coagulation boundaries with acceptable accuracy. Similarly, the model correctly predicted the behavior of the coagulation process to changes in water quality conditions (e.g., an increase in SUVA). The trained neural network could also predict full-scale filtered turbidity removals within ±1.4% at 11 different drinking water treatment plants under cold and warm water conditions. It should be emphasized that these results were obtained simply by providing raw water quality parameters. The model was only trained to predict filtered water turbidity removals using synthetic waters. It can be concluded that deep neural networks are ideally suited for modeling complex drinking water treatment problems such as coagulation, a task in the past considered to be virtually impossible.
CoagulationFlocculationJar TestMachine LearningMixingOptimization
Keen, OlyaSun, MeiPoler, JordanXu, Terry
Thesis (Ph.D.)--University of North Carolina at Charlotte, 2021.
This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s). For additional information, see http://rightsstatements.org/page/InC/1.0/.
Copyright is held by the author unless otherwise indicated.