Predicción de la composición de Carbonato de Dimetilo mediante redes neuronales artificiales a partir de la simulación de su proceso de síntesis

Abstract

The work objective was to predict the composition of dimethyl carbonate using artificial neural networks (ANN) from the simulation of the reactive distillation and pressure oscillation distillation process of methanol and propylene carbonate in DWSIM. The network design was carried out in Matlab with five variables in the input layer, 80 neurons in the hidden layer, two variables in the output layer, and the Bayesian regularization testing algorithm. 120 simulation data were used to test and validate the network. The feed temperature, the reflux and pressure relationship in the reactive distillation column, as well as the reflux and pressure relationship in the distillation column by pressure oscillation have been selected as input variables; the mass flow of propylene glycol and dimethyl carbonate are the output variables. A correlation coefficient of 0.928 and a mean square error of 0.0056 show the good performance of the network during its testing with 80 neurons and the Bayesian regularization algorithm. A p-value greater than 0.05 allows to validate the prediction model with 95% confidence, confirming that there is no statistically significant difference between the real and predicted data by the artificial neural network. These results indicate that the prediction model using ANN is efficient in predicting the dimethyl carbonate synthesis. It is recommended to evaluate other parameters that influence the performance of the artificial neural network, such as the activation functions and the quantity of hidden layers on the learning rate.