Inferência de frações de butano no propano e de frações de propano no butano em uma coluna depropanizadora utilizando rede neuronal

Abstract

This paper presents the studies carried out for the construction of a neural network, in order to calculate quality parameters (specifications) of oil derivatives. These quality parameters are dependent on process parameters, such as flows, pressures and temperatures, but the analytical expression of the function expressing this dependence is not known. However, this dependence is registered in the data collected throughout the years, during the operation of the production chain the refinery. This database allows the neural networks to be used to estimate, in real time, the quality parameters of interest as a function of the process parameters. The architecture of the neural network used was the multilayer perceptron. The chosen specifications were the percentage of heavy fractions in the top production and the light fractions in the bottom production of a distillation column. With the chosen architecture, several topologies were tested and the performance in function of the mean square error, in order to find the most appropriate topology. Although it has been proven that a neural network is able to approximate any continuous function, finding the appropriate architecture and topology for the network is a great challenge