Predição do equilíbrio de fases entre biodiesel, álcoois e CO2 em estado supercrítico por meio de redes neurais artificiais

Abstract

Currently, the search for renewable fuels is becoming increasingly important for the production of energy in Brazil and in the world. Biodiesel, a promising substitute for mineral diesel, is a renewable resource and minimizes greenhouse gas damage. However, the biodiesel production process is dependent on the transesterification reaction, a heterogeneous reaction that has its yield compromised due to the presence of two liquid phases. This drawback can be solved by adding supercritical carbon dioxide (CO2) to the system, since above the critical point the separation between phases of a fluid disappears. CO2 can still aid in the separation and purification process of the final product, in addition to dispensing solvents that are often toxic and difficult to remove, it is a solvent that is easy to recover. However, for a detailed study on the applications of supercritical CO2, it is necessary to define the behaviors between this and the other components involved in biodiesel production. In this way, modeling the behavior of these systems is very important. The development of an artificial neural network (ANN) model for the modeling of CO2 + Biodiesel + Methanol and CO2 + Biodiesel + Ethanol was the focus of this study. The ANNs method is a technique widely applied to engineering problems. This tool has the ability to model very complex problems, due to its ability to learn from experimental data. It was aimed to predict the transition pressures present in both systems, by means of an ANN with 5 input variables. In order to achieve the best result for both systems, 588 ANN structures were elaborated. Three different classes of ANNs were tested, the Feed-Forward network, the Cascade-Forward network and the Elman network. The number of neurons, the number of layers and the activation functions applied to the first layer of the neural network were studied. Among the networks developed, the Feed-Forward network and the Elman network demonstrated the best performance representing the CO2 + Biodiesel + Methanol and CO2 + Biodiesel + Ethanol system, in this order. The Feed-Forward network used 3 layers with 10 and 8 neurons in their hidden layers, in addition to making use of the tansig function in its first layer and the purelin function in the second layer and in the output layer. The Elman network presented 2 layers with 10 neurons in the hidden layer, which made use of the logsig function. These structures were selected by evaluating the lowest validation ean square error (MSE) and their results were compared with the Peng-Robinson models using the van der Waals and Wong-Sandler quadratic mixing rules. The values of average deviation (AD) and root mean square square deviation (RMSD) committed by both ANNs were significantly lower than the values exhibited by the Peng-Robinson models in the literature.