Forecast of glucose production from biomass wet torrefaction using statistical approach along with multivariate adaptive regression splines, neural network and decision tree

Abstract

Artificial intelligence (AI) has become the future trend for prediction after the data is provided to machine learning. This study uses data analysis to optimize the experiment, find the best-operating conditions, and obtain the maximum glucose concentration for bioethanol production where wet torrefaction (WT) is used to perform biomass pretreatment. Forty-nine (49) sets of data are split into training and test data in the ratio of 7:4. Glucose concentrations from five different feedstocks are trained and predicted using a neural network (NN) and multivariate adaptive regression splines (MARS), followed by a decision tree (DT) to predict the classification of the materials. The predicted NN results are better than MARS, so the NN training is used for the glucose prediction along with the Box-Behnken design (BBD) experiment. The BBD experiment is performed with the parameters of temperature (170, 175, and 180 °C), reaction time (10, 20, and 30 min), and sulfuric acid concentration (0, 0.01, and 0.02 M) for the WT of sorghum distillery residue. By adding the BBD experimental data in NN training, the fit quality of the model is improved to 99.78 %. The NN model predicts that the highest glucose concentration occurring at the optimal conditions (i.e., 173 °C, 10.5 min, and 0.02 M sulfuric acid) is 15.216 g/L with a relative error of 5.55 % between the prediction and experiment. These resuts indicate that NN is an appropriate approach to predicting glucose production from biomass WT for bioethanol production. Additionally, the analysis of variance (ANOVA) evaluation shows that the order of the vital parameter for glucose concentration is sulfuric acid, followed by reaction time and temperature.