Busca heurística e inferência de parâmetros cinéticos de reações de hidrotratamento de óleo diesel a partir de dados experimentais escassos

Abstract

Currently, the amount of contaminants (such as sulfur, nitrogen and aromatics) in diesel oil is strictly controlled, due to their impact on pollutant emissions. The most important process used to meet fuel specifications regarding contaminants is the catalytic hydrogenation, specifically hydrotreatment (HDT). Optimization of the HDT process can benefit from reliable kinetic models of the hydrodesulfurization (HDS) and hydrodenitrogenation (HDN) reactions, among others. Modeling the HDS and HDN reactions is a complex task, as these occur in triphasic media and it is difficult to determine the composition of diesel oil. Fenomenological models demand a huge experimental effort on generation of data to estimate transport parameters or the utilization of empirical correlations, frequently developed under conditions dissimilar to those of an HDT process. Additionally, utilization of detailed kinetic models is unfeasible, due to the complexity of the reactional mixture. Consequently, pseudo-homogeneous models are largely employed, as they are simpler than heterogeneous models. The equations used to described HDS and HDN kinetics are based on power-law or Langmuir-Hinshelwood terms, Lumping of sulfur or nitrogen compounds into a single pseudo-component does not allow for extrapolation of a model adjusted to a given feed to other feeds, independently of the type of model (homo- or heterogeneous) and equation of reaction rate (power-law or Langmuir-Hinshelwood) employed. However, it is interesting to use a single heteroatomic pseudo-component in the models, as a reduced demand for experimental data enhances the potential for application of the models, given that the model-feed specificity is overcome. In the present work, generalization of pseudo-homogeneous models employing power-law based kinetic rates was accomplished through the inference of apparent kinetics parameters using seven macro-properties of the feeds, which were: density at 20oC; viscosity at 20oC; total sulfur concentration; total and basic nitrogen concentrations; total aromatics content; and temperature of 90% vaporization of the simulated distillation curve. Conventional artificial neural networks (feedforward multilayer perceptrons) performed the inference. Experimental data are scarce, mainly due to the high cost of experimentation, inducing the development of a methodology for dataset enlargement, based on statistical variability of the kinetic parameters and macro-properties. Dataset enlargement made possible the development of independent neural networks for prediction of each kinetic parameter, correctly capturing correlations among macroproperties and the parameters. Two different groups of neural networks were developed, one with higher complexity networks (expressed by a larger number of hidden neurons) and another with simpler ones (possessing fewer hidden neurons). Both were capable of capturing the aforementioned correlations, even though the simpler networks filter out the majority of the disturbances applied to the inputs.