Comparative analysis of artificial neural network (ANN) models: CO2 loading in MDEA and blended MDEA/PZ solvents

Abstract

CO2 loading plays a vital role in gauging a solvent's capability to absorb CO2 from combustion gases and natural gas flow. Yet, the customary approaches for measuring CO2 loading hinge on laborious and costly laboratory experiments. Therefore, researchers employed artificial neural network (ANN) approaches to estimate CO2 loading using various datasets, although case-specific methods and approaches restricted their attempts. This study employed two datasets with 303 (CO2 absorption via MDEA) and 361 (CO2 absorption via MDEA/PZ) acceptable experimental data to create ANN models to estimate CO2 loading in aqueous amine solvents. The CO2 loading is evaluated using a wide range of input parameters, including temperature, CO2 partial pressure, Methyldiethanolamine (MDEA) solvent, and Piperazine (PZ) solvent for cases A and B, respectively. Twelve distinct training algorithms, four distinct activation functions, and three hidden layers with neuron variations created the models. ANN models trained with a blend of PZ and MDEA solvents generated more accurate predictions compared to models trained with solely MDEA solvent. The sigmoidal activation functions (Tansig and Logsig) generated more precise predictions than the linear activation functions (Poslin and Purelin). In addition, “Levenberg-Marquardt (lm)” with the highest coefficient of determination (R2) was selected as the best training algorithm among the twelve training functions. The best average coefficients of determination were determined to be 0.964 for the tansig activation function with [25 18 10] number of neurons in case A and 0.967 for the logsig activation function with [30 20 12] number of neurons in case B. The statistical analysis (R2, MSE, RMSE, AAD, AARD, and run time) confirmed that the network provided accurate estimates of CO2 loading for both datasets. The prediction potential of the network is then further indicated by model generalization inside the training data set.