Near-Infrared (NIR) Spectroscopy for Biodiesel Analysis: Fractional Composition, Iodine Value, and Cold Filter Plugging Point from One Vibrational Spectrum

An effective calibration model of biodiesel fuel properties prediction, based on near-infrared (NIR) spectroscopy data and an artificial neural network (ANN), was built. Biodiesel samples were derived from multiple sources and prepared using multiple experimental parameters. Four different fuel properties, including fractional composition, were accurately predicted. The root-mean-square errors of prediction (RMSEPs) on an independent sample sets for the end boiling point (50% v/v), the end boiling point (90% v/v), the iodide value, and the cold filter plugging point were 1.73 degrees C, 1.78 degrees C, 0.90 g I-2/100 g, and 0.77 degrees C, respectively. Multiple linear regression (MLR), principal component regression (PCR), partial least-squares (projection to latent structures, PLS) regression, (kernel) polynomial and spline versions of partial least-squares regression (Poly-PLS and Spline-PLS), and ANNs were compared for the prediction of biodiesel properties. Data preprocessing techniques and calibration model parameters were independently optimized for each case. The ANN approach was superior to the linear (MLR, PCR, and PLS) and quasi-nonlinear (Poly-PLS and Spline-PLS) calibration methods. The ANN approach was a factor of 7.5 +/- 1.9 more efficient than MLR and a factor of 2.6 +/- 0.9 more efficient than PLS (according to RMSEP ratios). We confirmed that biodiesel is a highly nonlinear object. Nine data pretreatment (preprocessing) methods (mean centering, mean scattering correction, standard normal variate, Savitzky-Golay derivatives, range scaling, etc.) were tested. The first/second-order Savitzky-Golay derivative, followed by Mean Centering plus Orthogonal Signal Correction, was found to be effective for biodiesel NIR data preprocessing.