Simulación de una unidad de destilación de crudo (CDU) en DWSIM

Abstract

The objective of the work presented was to create an artificial neural network (ANN) which is capable of providing accurate data based on a prediction of the behavior of a crude distillation unit (CDU) by changing the specifications of the crude and thus obtaining the composition of the final products of interest at different values of temperature (T) and other variables that are important when operating a crude fractionation tower, therefore recognizing that the temperature at which there is %VAP is the one that affects the process by analyzing the TBP distillation curves. The design, training and validation of the network was performed in MATLAB collecting data in the open-source program DWSIM validating the simulation results with high impact publications using Aspen HYSYS being this one of the most used at industrial level by the number of tools and data that this program gives the user for a successful application. The network design was carried out with 325 data for training and 25 data for validation, its architecture has 10 inputs, 20 hidden neurons and 5 outputs using the Bayesian Regularization (BR) algorithm for training, obtaining a Mean Squared Error (MSE) of 2.6904 E-04 and a Linear Regression coefficient (R) of 0.9971. The predictive capacity of the network was analyzed by comparing the average percentage errors in the external validation, obtaining 0.66, 2.23, 0.57, 0.68 and 0.48 for LN, HN, LD, HD and RESIDUE respectively, and these values were supported by a statistical analysis using the ANOVA method, determining that there is no statistically significant difference at the 95% confidence level. It is recommended to use this work for future studies on RESIDUE and separation in a Crude Distillation Unit (CDU) under vacuum.