Using Spectroscopy and Support Vector Regression to Predict Gasoline Characteristics: A Comparison of<sup>1</sup>H NMR and NIR
Ana L. Leal, Artur M. S. Silva, Jorge C. Ribeiro, F.G. Martins
Abstract
The applicability of two alternative spectroscopic techniques (i.e., 1H NMR and NIR) for the quantitative characterization of gasoline was compared in this work. The chemometric approach followed to build the regression models was support vector regression, and two distinct kernel functions were tested: Gaussian and linear. Additionally, a significance test was performed on test set predictions to determine if the difference between the estimations of 1H NMR and NIR-based models is statistically significant. According to the performance indexes of the developed models, NIR spectroscopy is preferable over 1H NMR for the prediction of most gasoline physical–chemical properties. Still, for most of the cases, it was also demonstrated that the estimations resulting from both spectroscopic techniques are not significantly different from each other. The accuracy level attained with the support vector regression models is adequate and enables the replacement of the standard methods of analysis for at least 10 different gasoline quality parameters.