Random forest as a potential multivariate method for near-infrared (NIR) spectroscopic analysis of complex mixture samples: Gasoline and naphtha

Abstract

Random forest (RF) has been demonstrated as a potential multivariate method for near-infrared (NIR) spectroscopic analysis of petroleum-driven products, highly complex mixtures of diverse hydrocarbons. For the study, a NIR dataset of gasoline samples and two separate NIR datasets of naphtha samples were prepared. These samples were carefully prepared over a long period to maximize compositional variation in each dataset. Partial least squares (PLS), the most widely adopted method in multivariate analysis, and RF were used to determine research octane numbers (RONs) of gasoline samples, and total paraffin, total naphthene and total aromatic concentrations of naphtha samples. The resulting accuracies of quantitative analysis for these samples were generally improved when RF was used. In addition, chance for overfitting of a model, which would occur occasionally in PLS modeling, was substantially lessened or possibly eliminated by the use of RF. On the contrary, in the case of RF, a calibration dataset composed of samples with narrow interval in property or concentration variation was required to improve the accuracy. Consequently, RF could be a useful multivariate method to analyze NIR as well as other spectroscopic data acquired from petroleum refining products, when properly utilized.