Variation in accuracy of industrially relevant NIR-based multivariate models for sulfur in crude oil and spectral trends during an extended period of time

Abstract

This study presents a detailed analysis of commercial crude oil partial least square (PLS) predictions of Sulfur based on near-infrared (NIR) spectra over a period of six months, covering hundreds of samples from very diverse crudes. A significant trend in the accuracy of Sulfur predictions is observed when the entire six-month data set is split in approximately monthly subsets. Using the wealth of samples most advantageously, we selected several individual crudes, three heavy and one light, and ran principal component analysis (PCA) of their respective spectra. It is found that for the heavy crudes there is a trend in the spectra that largely overlaps with the observed trend in the Sulfur predictions for those crudes. The one light crude analyzed shows only a weak trend in the raw spectra, which also agrees with the PLS predictions remaining largely unchanged through the inspected six-month time interval for that crude. All the trends identified in PC scores are clearly illustrated in the spectral differences among the pre-treated spectra. The spectral trends are found to partially translate into the trends in predictions of distillation fractions.