Investigation of sorbent characteristics used in low temperature dry flue gas desulfurization
Loading...
Date
2022-04-22
Authors
Makomere, Robert Someo
Journal Title
Journal ISSN
Volume Title
Publisher
Vaal University of Technology
Abstract
Coal is an efficient raw material for generating electric power. However, coal-fired power stations generate a great deal of emissions in the form of exhaust gases from the combustion of coal. The exhaust gases known as flue gas contain a significant amount of greenhouse gases (CO2, SOX, HF, Hg and NOX) that have led to detrimental effects on the environment. Sulphur oxides (SO2 and SO3) present in the exhaust gas stream facilitate acid rain formation while ozone layer depletion is accelerated by the presence of excess carbon dioxide (CO2). Flue gas desulphurization (FGD) is a post-combustion technology used to mitigate specifically SOX emissions from coal power plants. This is per pollution regulations set by air quality regulators for instance the Environmental Protection Agency (EPA). Dry FGD (DFGD) is a current subset of FGD systems that can be easily retrofitted to SO2-generating units at lower capital expenditures compared to wet and semi-dry FGD.
In this study, sorbent utilization was tested using a modified nano Ca(OH)2-diatomite sorbent. The reaction occurred in a simulated packed bed reactor consisting of sorbent balls. Cellulose nanocrystals were employed as the precipitating support for Ca(OH)2 formation while diatomite material was used as the siliceous additive. The final sorbent manifested superior scrubbing (𝑌1) and conversion (𝑌2) efficiencies compared to the commercial grade Ca(OH)2. The highest scrubbing and conversion responses were achieved when a diatomite ratio of 0.25 was used.
This research also explored experimental modeling using Artificial Neural Networks (ANN), a deep learning MATLAB fitting tool that can be trained to estimate final responses. The learning network using Levenberg-Marquardt (LM) algorithm was statistically compared to the Response surface methodology (RSM) technique to assess performance reliability. ANN was more precise in mapping the predicted responses with the actual experimental values achieving higher R2 values (𝑌1=0.993 and 𝑌1=0.9986) as opposed to those from RSM (𝑌1=0.9753 and 𝑌2=0.9771). Error analysis using RMSE and MSE justified the superior efficacy of the ANN model.
The final part of this study evaluated the algorithms present in ANN which can present more acute predicted metadata. Levenberg-Marquardt (LM) and Bayesian Regularization (BR) were chosen and investigated based on their iterations (epochs), validity tests, R2 values, RMSE and MSE. Although the BR algorithm took much more computing time, the predicted and actual values were correlated more effectively compared to the LM training tool. However, both algorithms were reliable in forecasting the sulfation and reagent utilization responses, and hence can be used to model DFGD.
Description
M. Tech. (Department of Chemical Engineering, Faculty of Engineering and Technology), Vaal University of Technology.
Keywords
Artificial neural network, Bayesian Regularization, Dry FGD, Levenberg-Marquardt, Response surface methodology, Sorbent conversion and sulphation efficiency