Modelagem e simulação termodinâmica de sistemas de equilíbrio entre dióxido de carbono, metanol/etanol e glicerol em condições supercríticas a partir de redes neurais artificiais

Abstract

The search for fossil fuel substitution options has been growing year after year. In this context, biodiesel gains space on the world stage, being an alternative to renewable fuel. In the transesterification reaction responsible for the production of the compound, glycerol is also formed, the main byproduct of the reaction and an important factor for the economic viability and profit maximization of the process. One of the factors that influence the transesterification reaction is the presence of two phases in the system, which occur due to the difference in polarity between the components, which makes mass transfer difficult. Thus, for the reaction to occur in a viable way, it is necessary to use catalysts (which may be heterogeneous, homogeneous or enzymatic) or a supercritical state, which changes the physicalchemical properties of the medium and causes only a liquid phase to occur. Although the methods for producing biodiesel have been known for a long time, the modeling of the system's behavior is still highly complex, with great difficulty in prediction using classical thermodynamic methods, such as state reactions. In view of this, new methods appear as a possible option to describe the system, such as Artificial Neural Networks (ANNs). These networks are computational systems with structures based on the human brain, capable of learning from experimental data. Therefore, this work aims to prescribe the systems Glycerol + Methanol + CO2 and Glycerol + Ethanol + CO2, through Artificial Neural Networks. For this, tests were developed related to data separation in training, validation and testing, different training algorithms, training parameters, activation functions and number of neurons in the hidden layers, in order to find the best network configuration to describe the systems. From this, 516 network architectures for the system were elaborated. For the Glycerol + CO2 + Methanol system, the best performing network reached a validation MSE of 0.1238 and R2 a validation of 0.9910. This network used CascadeForward architecture, with three hidden layers of 5, 3 and 3 neurons, using the softmax activation function in the first hidden layer and purelin in the others. For the Glycerol + CO2 + Ethanol system, the best performing network achieved a validation MSE of 0.0896 and a validation R2 of 0.9982. The network used Elman's architecture, with two hidden layers and 10 and 9 neurons, with a softmax activation function in the first hidden layer and purelin in the second. With the configurations of the optimal networks achieved, these were compared to works present in the literature of systems similar to those studied, in order to understand the quality of the results obtained. With the comparison, it is possible to notice that the method with RNAs is in fact very effective for the prediction of the system, presenting much lower error values in relation to the systems present in the literature.