Desarrollo de una RNA para la predicción de la destilación azeotrópica por oscilación de presión para la separación de isopropanol

Abstract

The main objective of this study is to perform a simulation and validation of a pressure swing distillation system, widely used in the chemical and pharmaceutical industry to separate the azeotropic isopropanol-water mixture. The central objective consisted of the development of an Artificial Neural Network (ANN) with the ability to predict the molar concentrations of isopropanol and diisopropyl ether in the distillate outlet streams and the bottom of the three columns. To develop the ANN, a database from the DWSIM simulator was used. This database consists of 400 records, with six input variables including feed temperature, IPA feed mole fraction, DIPE mole fraction at the inlet and the pressures of the three distillation columns (Col-1, Col-2, Col-3). Four output variables comprising the concentrations of IPA and DIPE in the distillate and residue were also obtained. In the process of development, training, validation and testing of the ANN, MATBLAB software was used in conjunction with the Neural Net Fitting application. The network architecture was based on the Bayesian Regularization algorithm, with six inputs and four outputs, and was configured with a hidden layer consisting of 50 neurons. The ANN achieved a (MSE)Mean Squared Error of 0.00075905 and a linear regression coefficient R of 0.99188, indicating good performance. However, to assess the reliability of the network, the percentage error between ANN predictions and the actual values obtained in DWSIM was calculated. This analysis included internal and external validation. In addition, a statistical program was used to perform ANOVA to support the results obtained. Finally, a process improvement was performed with the purpose of determining the operating conditions that allowed the separation of isopropanol-water obtaining a high purity IPA and DIPE.