Desarrollo de una RNA para la predicción de la concentración de alcohol furfurílico a partir de la simulación de su proceso industrial

Abstract

The purpose of this integration work was to predict the concentration of furfuryl alcohol by means of an Artificial Neural Network (ANN), for this purpose, an industrial process was simulated in the free software DWSIM, which was used for a sensitivity analysis of the design variables of the process. The variables that affected the final concentration of furfuryl alcohol were used as inputs for the ANN. The database consists of 200 data with four inputs: pressure and temperature at the upper stage in the distillation column, catalyst loading and reactor volume, and their corresponding three outputs: mole fraction of furfuryl alcohol at the bottom of the column, mole fraction of H2O and the percentage conversion of furfural in the reactor. The algorithms applied in the training of the network were three: Levenberg Marquardt, Bayesian Regularization and Scaled Conjugate Gradient. The three algorithms were compared and the one with the best performance was the Levenberg Marquardt algorithm. The network was developed in MATLAB with 10 hidden neurons in its architecture and trained with the Levenberg Marquardt algorithm. A mean squared error (MSE) of 0.0007854 and a regression coefficient of 0.9905 were achieved. The network was validated by means of a comparative statistical analysis obtaining a 95 % reliability. A final molar concentration of furfuryl alcohol of 0.9903 was achieved in the simulation. Therefore, the feasibility of the industrial process for obtaining furfuryl alcohol was demonstrated. It is recommended to apply a detailed study on the secondary products generated in the furfuryl alcohol production process and to extend the learning range of the network with different inputs and outputs for the generation of a new ANN trained with another algorithm to become a didactic tool.