Prediction of CO2 Breakthrough Time in Packed Bed Adsorption Column Using Artificial Neural Network

Abstract

Carbon dioxide is a component that constitutes a significant fraction of natural and biogas after methane and needs to be removed to increase its energy content for efficient combustion. A packed bed adsorption column is often utilized for separating gas constituents, and its dynamic performances are determined for successful cyclic adsorption design. Experimental procedure and phenomenological modeling are often used to determine or predict non-linear packed bed dynamic behavior. A simplistic surrogate model is a powerful tool for generalizing non-linear problems. In this work, a simple artificial neural network (ANN) was applied to predict the breakthrough time of CO2 in a packed bed adsorption column. The phenomenological packed bed column mathematical model was validated prior to dataset generation for ANN model training. The ANN architecture with a 4-10-1 structure is used. Several data pre-processing techniques and activation functions were analyzed to improve the learning ability of the ANN. The results show that the data pre-processing transformation techniques could increase the learning ability of the ANN model. The log transformation normalization (LTN) with elu activation function in the hidden layer is the best choice and is used to train the ANN. The trained ANN model can predict the breakthrough time with 99.99% accuracy between the actual and ANN model values. This demonstrates that the trained ANN model can be used to predict the breakthrough time of CO2 with excellent accuracy.