Machine learning prediction of the conversion of lignocellulosic biomass during hydrothermal carbonization
Navid Kardani, Mojtaba Hedayati Marzbali, Kalpit Shah, Annan Zhou
Abstract
The elevated conditions needed for hydrothermal carbonization of biomass require a special high-pressure reactor which makes it expensive and time-consuming. The soft computing approaches as proposed here can predict the conversion of any feedstock based on the composition and operating conditions without the need for any kinetic modeling. In this study, Extreme Gradient Boosting method (XGBoost), Multilayer Perceptron Artificial Neural Network (MLPANN) and Support Vector Machine (SVM) were trained in python programming language using the data available from the literature for hydrothermal carbonization of different biomass. Statistically, XGBoost showed a higher accuracy among all studied approaches with R2 of 0.999 and 0.964 for training and testing data, respectively. The conversion was sensitive to temperature, time, lignin, moisture content, cellulose and hemicellulose, respectively, for the range of conditions applied. It was also revealed that none of the parameters were negligible, however operating conditions were more influential followed by lignin content. This proposed approach can be extended to include liquefaction and gasification processes, where the distribution of products can be estimated for any lignocellulosic biomass.