Modelado y simulación de un sistema de destilación binaria benceno-tolueno para la predicción de la composición de los productos mediante redes neuronales artificiales.

Abstract

The purpose of this research was to model and simulate a benzene-toluene binary distillation system for the prediction of product composition through artificial neural networks. For the desing of the Artificial Neural Network (RNA), data obtained from the simulation in the DWSIM free access software with operating conditions proposed in a medium-impact scientific article published by the Department of Chemical Engineering of the Instituto Nacional de Tecnología de Calcuta was used. By pre-processing data, outliers were corrected by applyng the Grubbs test, therefore, a database with 100 process variables was used as inputs for modelling in Matlab. Of which 15% was used for validation, 15% for testing and 70% for RNA training. The Levenberg-Marquardt was used for model training. The modelling architecture of the FeedForward Backpropagation type had 2 input layers, 18 hidden layers, and 2 output layers. It is concluded that the desing of the artificial neural network has an optimal prediction performance since, statistically, it presents a correlation coefficient equal to 0.94076 and an average square error equal to 0.0049875. Besides, the registered values of –P for the molar fraction of benzene in the distillate and the residue as actual and predicted products are 0.9030 and 0.7926 respectively, indicating that there are no significant differences between the analyzed data. However, the implementation of a soft sensor is recommended for vector weights that influence RNA training.