Optimization of NIR spectroscopy based PLSR models for critical properties of vegetable oils used in biodiesel production

Abstract

Near infrared spectroscopy associated with multivariate analysis has been applied to monitor vegetable oils quality. In this work, twelve physical and chemical parameters of vegetable oils, relevant for quality monitoring of biodiesel raw materials, are quantified using NIRS: iodine value, water content, acid number, phosphorous content, saponification, flash point and fatty acids (myristic acid, palmitic acid, stearic acid, oleic acid, linoleic acid and α-linolenic acid). Two different procedures to select the most adequate spectral range for each property were proposed and applied to optimize partial least squares regression (PLSR) models on NIRS data. Models were developed resourcing to cross-validation error (RMSECV) and tested on an independent prediction set (RMSEP). Variable Window Size Method (VWSM) is based on the sequential screening of all possible single contiguous spectral regions. The Multiple Window Combination Method (MWCM) consists of testing every combination of defined spectral regions. Both approaches produced approximately similar quality models for most modelled properties in terms of cross-validation indicators, despite MWCM in general yielded lower errors when applied to an unseen dataset. Optimal models produced RMSEP of 0.03 ppm, 0.51 g I2 100 g-1 and 0.11% (w/w), respectively for water content, iodine value and α-linolenic acid. These last two parameters are especially relevant, given the need to comply with the EN 14214 and the first one because transesterification process yield. Application of VWSM and MCWM yielded superior models in terms of accuracy when compared with literature reported models for the same parameters, while increasing models interpretability.