Direct neural network modeling for separation of linear and branched paraffins by adsorption process for gasoline octane number improvement

Abstract

An artificial neural network (ANN) approach was used to develop a new predictive model for the calculation of hydrocarbons breakthrough curves in separation of linear and branched paraffins by adsorption process. Three-layer ANN architecture was trained using an experimental database and the concentration at t time over initial concentration (C/Co) was calculated as output variable. Experimental temperature (T), times of adsorption (t), octane number (ON) and the density (ρ) of the hydrocarbons were considered as main input variables for the model. For the ANN optimization process, the Levenberg–Marquardt (LM) learning algorithm, the hyperbolic tangent sigmoid transfer-function and the linear transfer-function were applied. The best fitting training data set was acquired with an ANN architecture composed by 22 neurons in the hidden layer (4-22-1), which made possible to predict the C/Co with a satisfactory efficiency (R2>0.96). A suitable accuracy of the ANN model was achieved with a mean percentage error (MPE) of ∼5%. All the C/Co predicted with the ANN model were statistically analyzed and compared with the “true” C/Co experimental data reported in the experiments carried out in the lab. With all these results, we suggest that the ANN model could be used as a tool for the reliable prediction of the breakthrough curves obtained during the separation of linear and branched paraffins by adsorption processes.