In order to fulfil specifications, avoid quality give-aways, improve process safety and save on costs, refineries need to have on-line methods for the control of their processes. On-line analysers allow product streams characteristic variables to be controlled, providing a great amount of data in almost real time.¹ Currently, the spectroscopic analysers [near infrared (NIR), mid infrared, nuclear magnetic resonanc and ultraviolet] are the most widely used. These analysers allow the value of the property of interest to be inferred from the measurement of a spectrum. Due to the great quantity of variables that constitute a spectrum, mathematical tools are required to extract the information related to the properties of interest. Repsol YPF has developed, and is in process of implantating, several NIR systems in its different laboratories, pilot plants and industrial complexes, which include from gasoline or diesel blending systems control to some petrochemical product manufacture controls. In the development of these systems, different hardware and software have been compared. In this work, the results obtained with three different chemometric methods for predicting critical gasoline properties for blending control are analysed and compared—partial least squares, a topological method and artificial neural networks.

Keywords: chemometric, NIR, PLS, topological method, artificial neural networks, gasoline, blending, control