Estimation of gasoline properties by¹H NMR spectroscopy with repeated double cross‐validated partial least squares models

Abstract

Commercial gasoline must satisfy several product specifications before trading.In the present work, repeated double cross validation using partial least squares regression was applied to create reliable prediction models for 13 physicochemical parameters (eg, density, vapour pressure, evaporate at 70 C, evaporate at 100 C, evaporate at 150 C, final boiling point, research octane number, motor octane number, aromatic content, olefinic content, benzene content, oxygen content, and methyl tert-butyl ether content) of gasoline produced in Matosinhos' refinery. The input variables for the regression are the 1 H NMR spectral intensities of a total of 448 samples, which were recorded using a pico-Spin NMR spectrometer operating at 80 MHz. The output variables are the corresponding property values, which were also measured according to ISO standard methods. A spectral feature elimination before multivariate analysis was done to remove noise and speed up the chemometric analysis. The optimum complexity of each model was achieved by repeated double crossvalidation strategy, consisting of 100 repetitions of two nested cross-validation loops. Quantitative partial least squares yielded accurate predictions of 11 of 13 properties within the reproducibility of ISO standards. The methodology presented in this work has been proven effective in property estimation and enables a significant reduction in the total time of gasoline quality control.

Using Spectroscopy and Support Vector Regression to Predict Gasoline Characteristics: A Comparison of1H NMR and NIR

“…A previous study using 1 H NMR in combination with PLS yielded the following SEP values: 0.62 for RON, 0.53 for MON, 0.56% v/v for olefins, 1.60% v/v for aromatics, 0.053% v/v for benzene, 0.11% m/m for oxygen, 0.89% v/v for MTBE, 2.21% for E70, 1.54% for E100, 0.79% for E150, 3.58 °C for FBP, 2.72 kg/m 3 for density, and 4.93 kPa for vapor pressure . Comparing those results with the ones given in this work, one can confirm that, except for MTBE, SVM-based models outperform PLS models for the estimation of gasoline properties using 1 H NMR.…”

“…There are numerous investigations reporting the successful application of both technologies in predicting some physical–chemical properties of gasoline. Among the most recent studies, it is possible to find applications of portable spectrometers or novel chemometric approaches that reportedly give more accurate predictions than the most established ones. − However, few studies have been published on the comparison of the 1 H NMR spectroscopy and NIR vibrational spectroscopy as potential techniques for properties estimation in the petrochemical industry, − and none of them includes the simultaneous estimation of the amount of gasoline physical–chemical properties analyzed in this work. In the current study we aim at comparing the performance of multivariate regression models developed using either NIR or 1 H NMR spectral data to predict 13 different gasoline parameters.…”