Comparative study between Partial Least Squares and Rational function Ridge Regression models for the prediction of moisture content of woodchip samples using a handheld spectrophotometer

Abstract

AbstractThe use of woodchip for energy use is expected to increase in the next years because of the European targets for mitigating climate change and reducing greenhouse gas emissions. The technical standard EN ISO 17225‐4 determines the woodchip quality classes based on different chemical–physical parameters. Among them, moisture content is one of the most important, and its real‐time monitoring could improve the product quality, increase combustion efficiency, and consequently provide a potential decrease of pollutant emissions. Although the lab procedure to determine moisture content is quite simple, it is too long compared to the real needs of power plant operators. A rapid and simple solution may be represented by near‐infrared spectroscopy coupled with chemometric techniques. In detail, two regression methods, that is, Partial Least Squares (PLS) and Rational function Ridge Regression (RRR), have been compared in order to develop a good model for the prediction of moisture content of woodchip samples as arrived at the lab or directly in the power plant. In addition, the prediction performance has been studied as a function of the number of the spectral measurements of the sample. The results showed that handheld instruments could give reliable results by taking enough replicated spectra. According to the estimated measurement uncertainty of moisture measurements, both methods suffer from lack of fit, and the performance could be better if all unknown sources of errors could be eliminated. In general, RRR performs significantly better than PLS, although, as with many nonlinear methods, the risk for outliers in predictions with RRR is higher than with PLS. This could be overpassed performing predictions with both methods, and in cases of big differences, PLS prediction could be chosen.