Desarrollo de una RNA para la predicción de la concentración del ácido fórmico producido en una columna de destilación reactiva de pared divisora

Abstract

The present work dealt with the creation of an artificial neural network (ANN) capable of predicting the molar concentration of formic acid (FA) produced in a reactive split-wall distillation column (R- DWC); DWSIM was used for the simulation of the process, the validation of the results, and a sensitivity analysis with the purpose of recognizing the operating conditions of the R-DWC that considerably influence the composition of FA; the identified variables were used in the creation of a database whose inputs are: molar fraction of methyl format (MF), molar composition of methanol (MeoH), temperature of the streams, operating pressure, heat removed in the condensers, and as outputs the molar composition of FA and MeoH. The design, training and external validation of the network was performed in MATLAB. The design consisted of inputs, 30 hidden networks and 2 outputs, which is trained by the Bayesian Regularization (BR) training algorithm achieving a mean square error (MSE) of 4.2102E-04 and a linear regression coefficient (R) of 0.9815. The predictive capacity of the network was evaluated by calculating the average MSE of the predicted data both in training and validation, reaching percentages of 8% and 7.28%; these data were supported by an ANOVA statistical analysis, obtaining a confidence level of 95°/o. Finally, with the data obtained, an optimization study was carried out in order to determine the operating variables that would allow reaching the maximum possible FA mole fraction. The improved configuration allowed reaching a FA concentration of 0.708 from the concentration of 0.646 of the fed MF; the process simulation presented a maximum percentage error of 17.6%. The design of a distillation column that allows the separation of the produced AF/1120 azeotropicmixture is recommended.