Application of artificial neural networks for simulation of experimental CO2 absorption data in a packed column

https://doi.org/10.1016/j.jngse.2011.05.001Get rights and content

Abstract

The generalization performances of the Back Propagation Multi-Layer Perceptron (BPMLP) and the Radial Basis Function (RBF) neural networks were compared together by resorting to several sets of experimental data collected from a pilot scale packed absorption column. The experimental data were obtained from an 11 cm diameter packed tower filled with 1.8 m ¼ inch ceramic Rashig rings. The column was used for separation of carbon dioxide from air using various concentrations and flow rates of Di-Ethanol Amine (DEA) and Methyl Di-Ethanol Amine (MDEA) solutions. Two in-house efficient algorithms were employed for optimal training of both neural networks. The simulation results indicated that the RBF networks can perform more adequately than the MLP networks for filtering the noise (measurement errors) and capturing the true underlying trend which is essential for a reliable generalization performance.

Highlights

► Extensive data are presented for packed absorption of CO2 from air by various solvents. ► Conventional software’s provided inadequate concentration profiles across packed bed. ► RBF networks generalized better than MLP networks due to their solid theoretical foundations.

Introduction

Separation of carbon dioxide from air and various industrial gases is essential from both operational and environmental views. For example, CO2 must be separated from natural gas to increase its heating value or carbon dioxide is usually extracted from various flue gases in beverage industry. To reduce global warming, CO2 should be also removed from industrial flue gases before exhausting them to atmosphere.

In many practical applications, the natural gas contains around five percent carbon dioxide (e.g. Iranian Mozduran sour natural gas contains 6.5% CO2). On the other hand, flue gases resulting from complete combustion of almost pure natural gases (e.g. Iranian Mozduran sweetened natural gas contains 98.4% CH4) usually contain around 9% CO2 when natural gas is burnt with 5% or more excess air. The experimental CO2 concentrations (1–5%) were selected to cover approximately the above specifications while meeting the CO2 flow-meter restrictions.

Although several adsorption and membrane processes are recently used for CO2 separation purposes (Xu et al., 2005, Fauth et al., 2005, Gray et al., 2005, Park et al., 1997, Datta and Sen, 2006) absorption processes are still more popular in this area (Vaidya and Kenig, 2007, Yeh et al., 2001, Huttenhuis et al., 2007). Alkanolamines (such as DEA or MDEA) are usually used for efficient separation of carbon dioxide from various industrial gases. Packed towers usually provide higher mass transfer areas and lower pressure drops when compared to tray towers.

Lin and Shyu (1999) investigated the absorption of carbon dioxide from nitrogen using MEA and MDEA solutions in a packed column under various operating conditions. A two parameter theoretical model was presented for describing the CO2 absorption behavior. The proposed model was validated using test data. They concluded that “an increase in the absorption load due to increased inlet CO2 concentration or gas flow rate leads to a much shorter breakthrough time. However, an increase in the amine concentration significantly enhances the CO2 absorption”.

Sultan et al. (2002) presented a theoretical model to investigate the effect of various operational parameters on the performance of a regeneration packed column. The experimental data were then correlated to estimate the water evaporation rate from desiccant (CaCl2) at various operating conditions. They concluded that “the water evaporation rate increases with increase of air and solution inlet parameters, namely, flow rate and temperature”.

Brettschneider et al. (2004) used a non-equilibrium heat and mass transfer model to describe the chemical absorption of ammonia, carbon dioxide and hydrogen sulfide in an aqueous solution containing sodium hydroxide, MEA and MDEA. The chemical reaction and its influence on mass transfer in the electrolyte system were accounted by enhancement factors. The calculation of the hydraulic parameters was based on standard correlations. The predictions of the mass transfer model were validated using experimental data.

Sharma et al. (2004) employed back propagation artificial neural networks to investigate the fault diagnosis in an ammonia–water packed distillation column. The network was reported to perform satisfactorily on detection of the designated faults. The relative importance of various input variables on the output parameters was calculated by partitioning the connecting weights. The simulation results indicated that “bottoms temperature, overhead composition and overhead temperature are not much affected by the disturbances in feed rate, feed composition and vapor rate in the given range”.

Liua et al. (2006) proposed a complex computational mass transfer model (CMT) for modeling of chemical absorption processes in packed columns. The model was able to consider heat effect for prediction of the concentration and temperature profiles as well as the velocity distributions. The presented model coupled the computational fluid dynamics (CFD) with computational heat transfer (CHT). The model was successfully validated using borrowed experimental data collected from a 0.1 m ID and 7 m height pilot scale tower randomly packed with ½ inch ceramic Berl saddles. The column was used for chemical absorption of CO2 from air by aqueous monoethanolamine (MEA) solution at total pressure of 103.15 kpa. Other sets of literature data collected from an industrial-scale packed column (1.9 m ID and 26.6 m height) randomly filled with 2 inch stainless steel Pall rings were also used for validation purposes. Chemical absorption of CO2 from natural gas was conducted using aqueous MEA solution. They argued that “the common viewpoint of assuming constant turbulent mass transfer diffusivity (Dt) throughout the entire column is questionable, even for the small size packed columns”, since Dt varies along both axial and radial directions.

In this article, the generalization performances of Multi Layer Perceptron (MLP) and Radial Basis Function (RBF) artificial neural networks are compared together using a pilot-scale packed absorption column. The training data were collected by conducting several experiments on absorption of CO2 from air using various DEA and MDEA solutions with different concentrations. Two different points will be emphasized in this article: a) Superior performance of RBF networks (when equipped with proper regularization level) to MLP networks and conventional software’s for empirical modeling of the absorption process. b) Providing experimental data for absorption of CO2 using a packed tower, which is quite limited and is essential for better analysis of the entire process.

Finally, a brief overview of neural networks and the two in-house algorithms for training RBF and MLP networks have been presented in Appendix A A brief overview of neural networks, Appendix B Training of the MLP network, Appendix C Training of the RBF network.

Section snippets

Experimental setup

A pilot scale packed column was used to collect the required experimental data. Fig. 1 represents the schematic diagram of the packed tower and the corresponding auxiliary equipments. Two separate glass columns (ID = 4.5 inches), one mounted over the other, and each one of them packed with 90 cm of ¼ inches ceramic Rashig rings were used. A liquid re-distributor was assembled between the two packed sections. Sampling points were fitted at both ends of the column for collection of pressure drop

Experimental data

After calibration of the Hempl gas analyzer apparatus, various measurements of inlet and outlet gas concentrations were performed. The following operating variables were varied during the experiments:

  • Type and solvent (DEA, MDEA and pure water),

  • Gas and liquid flow rates,

  • Concentrations of both solvents and gas streams.

The small temperature fluctuations during each experiment were ignored and the average temperature was calculated using the initial and final conditions. The barometric pressure was

HYSYS and Aspen simulation results

The pilot scale process illustrated in Fig. 1 was simulated using HYSYS and Aspen technical software. Various built in thermodynamic packages were tried during the simulations. Fig. 5 shows two predicted simulated concentration profiles across the packed section computed by HYSYS and Aspen. Table 1 provides a comparison between the measured outlet concentrations and the nearest simulation results using the most appropriate thermodynamic package for each case. Although both softwares failed to

Neural network simulation results

The entire collection of experimental data presented in Fig. 3, Fig. 4 for CO2 outlet concentrations and percentage of absorbed CO2 were used to train both MLP and RBF networks.

By definition, the RBF regularization network employs the same number of neurons as the data points. For better comparison of both networks performances, the number neurons of MLP networks were kept equal to the number of training exemplars.

The regularization network was completely self-sufficient and did not require any

Conclusion

The experimental data collected in this article was used as an influential tool to improve our understandings of the packed absorption processes especially the absorption of CO2 from air by various alkanolamine solutions. The simulation results presented here show that both HYSYS and Aspen software may perform inadequately for prediction of CO2 concentrations profiles across the packed column. This failure emphasizes the complex behavior of the process.

As an alternative approach, the

Acknowledgment

The authors wish to acknowledge the valuable contribution of Khangiran gas refinery officials for providing DEA and MDEA solutions and Ferdowsi university of Mashhad for the financial support. We also appreciate the kind assistance of Mr. Mirmehrabi.

References (24)

Cited by (0)

View full text