A model for the prediction of research octane number (RON) and motor octane number (MON) of hydrocarbon mixtures and gasoline-ethanol blends has been developed based on infrared spectroscopy data of pure components. Infrared spectra for 61 neat hydrocarbon species were used to generate spectra of 148 hydrocarbon blends by averaging the spectra of their pure components on a molar basis. The spectra of 38 FACE (Fuels for Advanced Combustion Engines) gasoline blends were calculated using PIONA (Paraffin, Isoparaffin, Olefin, Naphthene, and Aromatic) class averages of the pure components. The study sheds light on the significance of dimensional reduction of spectra and shows how it can be used to extract scores with linear correlations to the following important features: molecular weight, paraffinic CH3 groups, paraffinic CH2 groups, paraffinic CH groups, olefinic-CH=CH2 groups, naphthenic CH-CH2 groups, aromatic C-CH groups, ethanolic OH groups, and branching index. Both scores and features can be used as input to predict octane numbers through nonlinear regression. Artificial Neural Network (ANN) was found to be the optimal method where the mean absolute error on a randomly selected test set was within the experimental uncertainty of RON, MON, and octane sensitivity.

A model for the prediction of the derived cetane number (DCN) and carbon/hydrogen ratio (C/H) of hydrocarbon mixtures, diesel fuels, and diesel−gasoline blends has been developed on the basis of infrared (IR) spectroscopy data of pure components. IR spectra of 65 neat hydrocarbon species were used to generate spectra of 127 hydrocarbon blends by averaging the spectra of their pure components on a molar basis. The spectra of 44 real fuels were calculated using n-paraffin, isoparaffin, olefin, naphthene, aromatic, and oxygenate (PIONA-O) class averages of pure components. It is shown that this strategy retains knowledge of C/H, an important indicator of the chemical structure. Three methods were compared to assess the prediction of DCN and C/H ratio from the assembled IR spectra, i.e., partial least squares regression (PLSR), support vector machine (SVM), and artificial neural network (ANN). It was found that ANNs gave the best performance with DCN prediction errors of ±1.1 on average and C/H prediction errors of ∼0.8%. Lasso-regularized linear models were also used to find simple combinations of wavenumbers that yield acceptable estimations.

No abstract

No abstract

No abstract