Multivariate Calibration of Sulfur in Sour Crude Oils via Near-Infrared Spectra

Abstract

Regular supplies of crude oils during a nominal two-year period were used to develop and test partial least squares (PLS) multivariate models for determining sulfur based on near-infrared (NIR) spectra. Several models built automatically by chemometrics software were selected for testing in a series of cross validations and via large, independent prediction data sets. Three “annual” multivariate models were created in this manner, comprising the spectra acquired in 2017, 2018, and 2019, with each of them tested with the remaining two sets used for independent predictions. One of those three models performed best in terms of combined calibration and prediction errors, with the other two being suboptimal either because of the higher calibration errors or because of the limited predictive abilities. This is followed by creating three biannual models with the respective remaining set used for the prediction. Updating the models in this manner largely proved beneficial primarily due to alleviating the issues seen in the suboptimal models. The overall prediction error based on the analysis of about 1300 industrially relevant samples and across two years of acquisitions steadily indicates the prediction error of about 0.1 wt % for crude oils with the average content of ∼3 wt % of sulfur.