Moisture Prediction of Biomass Fuel Based on Near-Infrared Spectroscopy and Deep Learning Algorithm

The moisture content of biomass (biomass MC) is related to the costs of power plants and affects the efficiency of the boiler. Online monitoring of the biomass MC will be beneficial for combustion control and biomass fuel management. In this study, the dimensions of biomass crushing and the height of the near-infrared (NIR) spectrum acquisition on a power plant conveyor belt were simulated in an experimental setup. Multiple types of biomass samples from different regions were prepared under different moisture environments and scanned to obtain the NIR spectra of the biomass. Partial least squares (PLS), support vector regression (SVR), and backpropagation neural network (BPNN) were used to evaluate MC prediction models based on NIR data and deep feature data extracted by the deep autoencoder (DAE) and the supervised deep autoencoder (SDAE). The BPNN model performs best in modeling based on NIR data. The supervised deep autoencoder with the backpropagation fusion neural network (SDAE-BPFNN) model based on a deep learning algorithm achieved optimal performance in deep feature data modeling. The coefficient of the root-mean-square error of prediction (RMSEP) of SDAE-BPFNN was 2.51% wb, which is a relative reduction of 15.48% compared to the RMSEP of the BPNN based on NIR data. Nonlinear feature extraction of spectral data, retaining the information on variables related to MC, high accuracy, and strong robustness are the advantages of the proposed model based on deep learning algorithms.