Aplicação de métodos de calibração multivariada na previsão de propriedades de combustíveis fósseis e biocombustíveis usando dados da espectroscopia por infravermelho, médio e próximo, e espectroscopia Raman

Abstract

With the development and increase in the use of renewable fuels, there is also need to guarantee its production and distribution with the specified quality. In this way, it became necessary to study effective methodologies to determine physicochemical properties in a fast and non-destructible way, allowing its application to the quality control on / in / at-line monitoring of fuels in an industrial environment. In literature, near and middle infrared spectroscopy have been exhaustively coupled with chemometric tools to determine the properties of diesel-biodiesel blends. However, the inclusion of Raman spectroscopy application can make testing easier and contribute to building of low-cost instruments for quality control of fuels and biofuels. In this scenario, the present study aims to present models for prediction of the properties of fuels (diesel) and biofuels (biodiesel), pure and its blends, based on spectroscopic techniques (NIR, mid-IR, and Raman) coupled with chemometric tools, which include the use of variable selection techniques combined with PLS and the application of random forests (RF) and support vector machine (SVM) for nonlinear properties.In the first part of the study (case 1), the refractive index and density of 148 biodiesel samples were determined using spectroscopic data. In the second part of the study (case 2), the cold filter plugging point (CFPP) and the kinematic viscosity at 40˚C from 333 samples of biodiesel, diesel, diesel-biodiesel and blends diesel-biodiesel-ethanol were determined using spectroscopic data. In both cases, 80% of samples were used for the calibration set, and the remaining 20% was used as an external validation set. The results showed that PLS model with mid-IR spectra with intervals selected by UVE algorithm provided the best prediction for the refractive index with five latent variables, root-mean-square of error (RMSEP) equal to 1.0∙10-4 and prediction coefficient (R2pred) equal to 0.95. The best prediction model for density was obtained by RF method with mid-IR spectra, with RMSEP equal to 1.16∙10-4 g.cm-3 and R2pred equal to 0.98. For CFPP prediction model, the best model was obtained using SVM with mid-IR spectra and intervals selected by the RC algorithm, with RMSEP equal to 0.7˚C and R2pred equal to 0.94. The best prediction model for kinematic viscosity at 40˚C was obtained by RF method with mid-IR spectra, with RMSEP equal to 0.0184 mm2.s-1 and R2pred equal to 0.98. The results in this work showed that mid-IR spectra were the most adequate in predicting the studied properties. The figure of merit sum of Wilcoxon test probability (SWTP) developed in this study was indispensable to choose the best model.