Multivariate regression models obtained from near-infrared spectroscopy data for prediction of the physical properties of biodiesel and its blends

Multivariate calibration based on Partial Least Squares (PLS), Random Forest (RF) and Support Vector Machine (SVM) methods combined with variable selections tools were used to model the relation between the near-infrared spectroscopy data of biodiesel fuel to its physical-chemical properties. The cold filter plugging point (CFPP) and a kinematic viscosity at 40 degrees C of the biodiesel samples and its blends were evaluated using spectroscopic data obtained with a near-infrared reflectance accessory (NIRA/NIR-FT-IR). Therefore, one hundred forty-nine blends were prepared using biodiesel from different sources, such as canola, corn, sunflower, and soybean. Furthermore, biodiesel samples purchased from the Brazil South Region were added to the study. One hundred samples were used for the calibration set, whereas the remaining samples were used as an external validation set. The results showed that the SVM model with baseline correction + mean centering preprocessing gave the best prediction for the CFPP, with a root-mean-square of error (RMSEP) equal to 0.9 degrees C. Among the models presented, the best result for predicting the kinematic viscosity at 40 degrees C was obtained by the PLS regression method using an interval selected by UVE with baseline correction+ derivative preprocessing, with the RMSEP equal to 0.0133 mm(2).s(-1). The results in this work showed that the proposed methodologies were ade-quated in predicting the biodiesel fuel properties. The figure of merit Sum of Wilcoxon Test Probability (SWTP) presented in this study was necessary for the conclusion of the best model.