The aim of this study is to develop an alternative analysis for the characterisation of heavy oil products. Predictive chemometric models have been developed by mid-infrared (MIR) and near infrared (NIR) spectroscopies. This work is mainly concerned with the predictive model optimisation of saturate, aromatic, resin and asphalten contents (SARA). A simultaneous optimisation procedure of spectral data pre-processing methods and variable selection by genetic algorithms was evaluated. This approach led to the best NIR predictions and showed the potential interpretation of the selected variables. A comparative study of MIR and NIR spectroscopies for the development of heavy oil property predictive model was also performed. Results have shown that NIR spectroscopy is globally better for our application. It has been shown that spectroscopic data fusion can improve predictive power of models. It seems however necessary that both spectroscopies, when considered separately, correspond to similar predictive power in order to expect an improvement when combining MIR and NIR. The interpretation potential of multiblock has been confirmed for the identification of MIR and NIR specific information. Finally, models developed for the prediction of density, contents of SARA, Conradson carbon, hydrogen, sulphur and nitrogen were found satisfactory for an application at laboratory.
