Aplicação de espectroscopia no infravermelho próximo (NIR) e médio (MIR) associada a métodos quimiométricos, para avaliação de parâmetros físico-químicos em frações de petróleo

Abstract

Petroleum products are, in general, highly complex and a considerable effort is needed to characterize their chemical and physical properties, though sometimes the results of several analyses are urgent and this is compromised by the way the analyses are carried out. Thus, chemometrics associated with molecular spectroscopy (particularly NIR and MIR) has good potential as a tool in analytical chemistry, creating alternative methods to characterize and evaluate physical and chemical properties of petroleum and its derivates with high precision, reliability and rapidity. To improve the predictor performance, appropriate procedures are being used to select spectral regions associated with the property of interest. In face of their applicabilities, this work proposes the use of chemometric tools, with variable selection (Interval Partial Least Square, iPLS and Sinergism Interval Partial Least Square, siPLS; Elimination of Uninformative Variables, UVE and Genetic Algorithm, GA), associated with mid infrared (MIR) and near infrared (NIR) spectroscopies to determine the following properties in petroleum fractions: API gravity, Cetane index, Refractive index (at 20°C), Sulfur content (%m/m), Smoke point (mm), Aniline point (°C), Freezing point (°C), Plugging point (°C), Cloud point (°C) and Pour point (°C), enabling the evaluation of performance of the obtained models, as well as the techniques used in variable selection. This evaluation was performed by determination and analyses of the following requirements: coefficient of determination (R2), several calculated errors for the calibration and prediction set. The models were also subjected to statistical tests (α=0,05), and the figures of merit were calculated. The best models to predict API gravity and cloud point were created by applying iPLS to the MIR data, whereas for prediction of sulfur content, refractive index, and smoke and aniline points the models were created by applying iPLS to NIR data. In this context, the best model to predict the pour point was the GA. Finally, to predict freezing point, none of the variable selection methods improved the predictive capability when comparing to the model created using only PLS in MIR data. Thus, the conclusion is that a better performance was obtained for the models created from MIR data. Regarding efficiency of variable selection methods, the iPLS and siPLS methods resulted in a best performance.