Near Infrared Spectroscopy Combined with Support Vector Machines as a Process Analytical Chemistry Tool at Petroleum Refineries

Abstract

Introduction Process control and product quality control in petroleum refineries are often performed by means of on-line analysers which provide real-time information to a central system. This allows the adjustment of process variables if process disturbances or set point tendencies arise that differ from desired values. Analytical methods based on near infrared (NIR) spectroscopy combined with chemometric methods have been used as a tool for process analytical chemistry (PAC). This approach can provide useful information about sample composition at petroleum refineries because the various petroleum fractions and their physico-chemical characteristics are associated with different proportions of n-alkanes, iso-alkanes, cycloalkanes and aromatic compounds present in the fractions. Near infrared spectroscopy is able to distinguish between different contents of C–H bonds of methyl and methylene groups present in these compounds together with N–H, S–H and O–H bonds present in minor quantities. The partial least squares (PLS) algorithm is the one most commonly used in multivariate calibration model development. PLS is a linear method that can model some degree of non-linearity using an appropriate number of latent variables. However, certain applications of the support vector machines (SVM) algorithm have been shown to produce results better than those obtained by PLS. Feedstocks for petroleum refineries are often from different geographic regions and possess quite different characteristics. Likewise, petroleum fractions used in the production of derivatives may vary their characteristic properties over a wide range. It is difficult to include in a calibration model all the variability which may occur in practice in the refining process by means of experimental design and, moreover, complex mixtures such as petroleum fractions may not have a linear relationship with the parameter under study. For these reasons, it becomes necessary to use analytical methods that can provide non-linear calibration models with low prediction errors even for samples which have not been included in the calibration analytical range. One such method uses the SVM algorithm; it is based on statistical learning theory and can provide non-linear calibration models with high generalisation performance. These features make SVM an attractive option for the development of calibration models more effective for on-line analysis of petroleum, petroleum fractions and petroleum derivatives at petroleum refineries. The SVM algorithm was originally developed to solve pattern recognition problems and was subsequently extended to handle regression problems. Support vector machines for classification or support vector classification (SVC) simultaneously minimises the empirical error of classification and maximises the separation margin between classes using an optimal separation hyperplane, leading to a unique solution. One of the major features of SVC models is that they can operate in a kernelinduced feature space allowing non-linear modeling while good generalisation performance can be obtained even with relatively small datasets. These characteristics mean that SVC can provide a better classification performance than linear classification algorithms such as soft independent modeling of class analogy (SIMCA). To deal with regression problems, a modification of the classification algorithm was performed; in support vector regression (SVR), the use of the e-insensitive loss function that limits regression errors and penalises deviations beyond the adjusted limit during model development, helps ensure good performance. Key steps in SVC and SVR model development are the parametric optimisation and appropriate choice of the kernel function. This work reports the performance of SVM applied to NIR spectroscopy data for (1) development of regression and classification models for on-line analysis of petroleum fraction streams and diesel oil at a refinery and (2) compares the SVM predicted results with corresponding data produced by SIMCA and PLS methods.