Prediction of mass transfer coefficients of air-stripping packed towers for volatile organic compound removal.

Abstract

Removal of volatile organic carbons (VOCs) from contaminated ground water using air stripping towers is a best available technology. The design of packed towers necessitates an estimate of the overall mass transfer coefficient, KLa, which is usually computed using parametric models. These models have been found to have several drawbacks and limitations. The main thrusts of this thesis are a critical analysis of the literature and existing data, the compilation of a very large, high quality database, and the development of an alternative methods for predicting KLa. The thesis describes packed tower air stripping technology, including design considerations, design procedures, and potential operation problems. It presents methods that are used to develop mass transfer coefficient. A new database that is both comprehensive and representative of current air stripping applications has been assembled. An evaluation of the well-accepted Onda correlations concluded that the latter systematically overestimates KLa for packings that are typically used in full-scale applications. A large part of the observed difference between the experimental and Onda predictions is due to a lack of fit of the Onda model. Thus, the prediction quality of the correlations still can be improved. Also, experimental KLa values obtained in this study varied with packing depth. Based on the above and using the Onda model form, a new and improved Onda model was developed. The new correlations addressed several of the shortcomings of the Onda model. The predictions by the improved correlations were superior to those of the Onda model. To overcome the chronic limitations of the existing methodology, this thesis proposes a new approach based on neural network (NN) technology. The NN solution can be divided in three steps: (i) architecture selection, (ii) weights optimization, or training, and (iii) validation of the solution. The KLa predictions by the NN were superior to those of both the Onda and the improved Onda models. The main achievement of the neural network model is not only the lower error of its estimates through a better fit of the experimental data, but it is also its ability to describe better the relationships between mass transfer and the operating variables. The NN model successfully simulated the sharp increase in KLa at high gas loading rates that both the Onda and improved Onda models were unable to simulate. Moreover, the NN model was able to simulate the high non-linearity of the process, particularly the effect of packing depth, and liquid and gas loading rates. (Abstract shortened by UMI.)