ABSTRACT
Removing CO2 as an acidic-potential component from different gaseous flows is a main topic in different industries producing green-house gases, especially in natural gas sweetening units. A group of well-known absorbents for CO2 are the amine solutions. The common amine compounds consisting di-ethanolamine (DEA), methyl-di-ethanolamine (MDEA), and their mixture in aqueous solution have been investigated in this study. The effort was to develop new models for estimation of CO2 loading capacity of the presented amine solutions using genetic programing (GP) and stochastic gradient boosting (SGB) trees as two advanced and novel machine learning approaches in this area. A total of 175 sets of experimental data of CO2 absorption including independent variables (temperature, CO2 partial pressure, concentrations of DEA and MDEA in water) and objective function (CO2 loading capacity) were collected from literature and fed to the mentioned algorithms (GP and SGB) as input dataset. Then, each algorithm was run over the dataset, separately and two new models were created. Finally, strict statistical evaluations were implemented to assess the estimating capability of the new models. The statistical parameters including correlation coefficients (R2 SGB = 0.99848 and R2 GP = 0.99087), root-mean-square deviations (RMSDSGB = 0.00903 mol/mol and RMSDGP = 0.02244 mol/mol) and average absolute relative deviations (AARDSGB = 0.95628% and AARDGP = 8.71909%) show that the utilized powerful algorithms have enhanced the applicability of the new developed models providing good` estimations in operational processes. Final results show superiority and more accuracy of the new SGB model for confident predictions in amine process.
Nomenclature
AARD % | = | Average absolute relative deviation |
AMP | = | 2-amino, 2-methyl-1-propanol |
ANFIS | = | Adaptive neuro-fuzzy inference system |
ANN | = | Artificial neural network |
ARD% | = | Absolute relative deviation |
CDEA | = | Concentration of DEA |
CMDEA | = | Concentration of MDEA |
DEA | = | Di-ethanol amine |
DIPA | = | Di-isopropanol amine |
E | = | Extract phase mass (or mol) rate |
G | = | Gas feed mass (or mol) rate |
GB | = | Gradient boosting |
GP | = | Genetic programming |
GRN | = | Generalized regression neural networks |
ICA | = | Imperialist competitive algorithm |
LSSVM | = | Least squares support vector machine |
MDEA | = | Methyl, di-ethanol amine |
MEA | = | Mono-ethanol amine |
M | = | Molar concentration (mol/L) |
n | = | Number of samples in the dataset |
PCO2 | = | Carbon dioxide gas partial pressure |
PSO | = | Particle swarm optimization |
PZ | = | Piperazine |
R | = | Raffinate phase mass (or mol) rate |
R2 | = | Squared correlation coefficient |
RMSD | = | Root-mean-square deviation |
S | = | Solvent phase mass (or mol) rate |
SGB | = | Stochastic gradient boosting |
SVM | = | Support vector machine |
T | = | Temperature |
TEA | = | Tri-ethanol amine |
x1 | = | Mass (or mol) fraction of CO2 in solvent |
x2 | = | Mass (or mol) fraction of CO2 in raffinate phase |
y1 | = | Mass (or mol) fraction of CO2 in gas feed |
y2 | = | Mass (or mol) fraction of CO2 in extract phase |
= | Predicted dependent variable | |
= | Experimental dependent variable | |
= | Average of experimental dependent variable | |
α | = | CO2 loading capacity |
Supplementary material
Supplemental data for this article can be accessed here.