Predicting feedstock and percent composition for blends of biodiesel with conventional diesel using chemometrics and gas chromatography–mass spectrometry

The two main goals of the analytical method described herein were to (1) use principal component analysis (PCA), hierarchical clustering (HCA) and K-nearest neighbors (KNN) to determine the feedstock source of blends of biodiesel and conventional diesel (feedstocks were two sources of soy, two strains of jatropha, and a local feedstock) and (2) use a partial least squares (PLS) model built specifically for each feedstock to determine the percent composition of the blend. The chemometric models were built using training sets composed of total ion current chromatograms from gas chromatography–quadrupole mass spectrometry (GC–qMS) using a polar column. The models were used to semi-automatically determine feedstock and blend percent composition of independent test set samples. The PLS predictions for jatropha blends had RMSEC = 0.6, RMSECV = 1.2, and RMSEP = 1.4. The PLS predictions for soy blends had RMSEC = 0.5, RMSECV = 0.8, and RMSEP = 1.2. The average relative error in predicted test set sample compositions was 5% for jatropha blends and 4% for soy blends.

► Biodiesels from three different feedstocks were blended with conventional diesels.
► The blends were separated by GC–qMS with a polar column as well as a nonpolar column.
► Feature selection, scaling, PCA, HCA, and KNN were used to determine the feedstock.
► Blend percent composition was determined using a PLS model built for the feedstock.