Variable selection and modeling from NIR spectra data: A case study of diesel quality prediction using LASSO and Regression Tree

Abstract

The objective of this research is to design a model for predicting diesel fuel parameters from the data obtained from near infrared spectroscopic analysis of the fuel. Due to the complexity and the sheer number of peaks obtained in the spectral data, only those wavelengths that have a significant impact on the parameters are filtered out. Four types of variable selection techniques (LASSO, correlation coefficient, Mallow's Cp criterion, Relative sensitivity ratio) were applied on the NIR spectra data. Following variable selection, two models based on ridge regression and regression tree were developed. The models were used to successfully predict six diesel fuel parameters: cetane number, boiling point, freezing point, total aromatic content, viscosity and density from NIR spectra data. Variable selection by LASSO followed by regression tree modelling produced the best prediction accuracy.