Desenvolvimento de analisadores virtuais e sua aplicação na predição do ponto final de ebulição da nafta de craqueamento

Abstract

Soft sensors can be successfully used in control and optimization applications. They can be seen as an efficient low cost alternative for online monitoring of quality indicators of process streams. The goal of this work is to develop a set of computational routines based on mathematical and statistical techniques in order to improve the building of these soft sensors. The application here developed presents a modular structure which allows the user to access independent tools for data pre processing, variable selection and mathematical system modeling. The application was used to predict final boiling point (FBP) of light naphtha in a fluid catalytic cracking unit of Petrobrás refinery. Welooked for the simplest model with the smallest number of variables. This case study was designed to test the difficulty of prediction of the FBP in case of hydraulic problems in the main fractional distillation column. Because of the large range of modeled operations performed by the soft sensor it was not possible to reach values close to the reproducibility obtained in laboratory tests. On the other hand the model is capable of signaling the presence of favorable conditions for the loss of specification of FBP. Also, it was possible to perceive that the FBP is sensible to three variables: main column top temperature, which influences the thermodynamic equilibrium of this section; raiser top temperature, which influences the amount of light hydrocarbons production; and the differential pressure in the top section of the main fractional distillation column, which signals hydraulically unfavorable conditions for controllingthe FBP. Amongst the models that were obtained the multilayer Perceptron with Bayesian regulation stands out due to its more powerful generalization capabilities.