Evaluación del uso de redes neuronales artificiales en la predicción de resultados de un proceso de isomerización para su uso como herramienta didáctica en la materia de simulación de procesos

Abstract

The objective of this work is based on the development and evaluation of a neural network capable of predicting the molar concentration of isobutane and isopentane in an isomerization process with a correlation coefficient as close to 1. For the design of the network, the system is first modeled through an open code simulation software, DWSIM; once simulated, a base of 115 pairs of data is obtained from it by varying the mole fractions in the feed, reflux relationship and conversion percentage; a statistical analysis is applied to this group of data to verify the existence of atypical data and to corroborate if the mean of the data is acceptable. The Matlab software is used to select the architecture and the network logarithm that best suits the data set. Finally, a statistical analysis was applied to compare the values obtained by the simulator to those predicted by the neural network. As final result, it was verified that the best fitting network is the Bayesian Regularization algorithm with a network of twenty hidden neurons, six input neurons and four output neurons, giving a mean square error of 1.73e-04 and a total correlation coefficient of 0.9990. Thus concluding that the network is able to predict the mole fraction of isobutane and isopentane in the isomerization process with the same accuracy as a simulator would do, since there is no significant difference between the means of the real data and those predicted by the network. It is recommended to use different thermodynamic methods and specific ranges for the input variables to obtain results suitable to the reality of the isomerization system.