Using a neural network to predict deviations in equilibrium model of CO2 capture by absorption with potassium carbonate

The hot potassium carbonate (HPC) process aims to remove CO2 present in the synthesis gas. This removal is performed in the absorption process, where reaction of CO2 with K2CO3 solution occurs. This reaction is slow and H3BO3 can be used to speed up the reaction. Two approaches can be used to simulate this process: equilibrium and rate-based. In general, the equilibrium model does not correctly predict the absorption process, and the use of the rate-based model is more recommended. However, implementing the rate-based model is complex, as it demands greater number of adjustment parameters and differential equations. An alternative to using the equilibrium model and increasing its representativeness is to calculate the Murphree efficiency of components present in the process. In this context, this work aims to propose a methodology based on Artificial Neural Networks (ANNs) for the calculation of these efficiencies using two commercial software simultaneously: Aspen Plus and MATLAB. An ethylene oxide industrial plant was simulated in order to evaluate the limitations and perspectives of both models and the effect of including Murphree efficiency calculations in the equilibrium model. Simulation results were compared with plant data and predicted that the simplest equilibrium-based models for the absorber can lead to deviation of up to 20% in the prediction of the CO2 capture rate, while the model corrected with the Murphree efficiency, calculated from the neural networks proposed in this article, reduce this error to less than 5% in all operational conditions under evaluation.

Automated summary

The hot potassium carbonate (HPC) process is used to remove CO2 from synthesis gas through absorption with K2CO3 solution. The reaction is slow, and H3BO3 can be used to accelerate it. Equilibrium and rate-based models can be used to simulate the process, with the rate-based model being more recommended for accuracy. However, implementing the rate-based model is complex, while using the Murphree efficiency and Neural Networks to calculate the equilibrium model can reduce prediction errors to less than 5%.