Key points are not available for this paper at this time.
• NIRS enables fast prediction of RDF moisture content with good accuracy. • Simple regression at selected wavelengths yields precise moisture estimates. • MLP and ridge regression show robust feature selection for NIRS data. • Dry spectra are predicted from wet spectra using autoencoder and random forest. • Random forest achieves the best accuracy in dry spectrum reconstruction. Refuse derived fuel (RDF) sourced from industrial and municipal waste contains primarily plastics, paper/cardboard and textiles. Since RDF has high heating values and low prices, it is used as alternative fuel to coal in cement production, offering CO 2 reductions. However, the cement industry currently lacks real-time quality control of RDF, resulting in unnoticed reductions in heating value caused by elevated moisture content. Moisture content determination by drying takes up to 4 h, which does not allow for quick reactions in operation of the cement kiln. This study addresses the challenge by applying near-infrared spectroscopy (NIRS) to predict the moisture content of paper and cardboard particles, which retains most moisture in RDF. The particles were moisturized by different methods with NIRS measurements at each stage. For the prediction, different approaches ranging from linear regression of selected wavelength absorptions to multiple machine learning regression techniques like ridge, lasso and multilayer perceptron are compared in accuracy and feature representation. The multilayer perceptron showed accurate results and robust feature selection, although also simple linear regression at selected wavelength gave adequate accuracy. In a second step, moisture content and spectra of wet particles were used to predict the spectra of dry particles. These spectra are essential to correctly identify particle types or predict heating values, since training data consists of dry particle data. While linear models are not reaching acceptable results, both an autoencoder and a random forest regressor yielded spectra with correct shapes, which is sufficient for classification with derivative spectra.
Fischer et al. (Sat,) studied this question.