A comparative analysis of biomass torrefaction severity index prediction from machine learning

• Published in: Applied energy 2022-10, Vol.324, p.119689, Article 119689
• Main Authors: Chen, Wei-Hsin, Aniza, Ria, Arpia, Arjay A., Lo, Hsiu-Ju, Hoang, Anh Tuan, Goodarzi, Vahabodin, Gao, Jianbing
• Format: Article
• Language: English
• Subjects: artificial intelligence; Artificial neural network (ANN); biofuels; biomass; computers; energy; feedstocks; Machine learning; Multivariate adaptive regression splines (MARS); neural networks; prediction; temperature; torrefaction; Torrefaction and biochar; Torrefaction severity index (TSI)
• ISSN: 0306-2619; 1872-9118
• DOI: 10.1016/j.apenergy.2022.119689

[Display omitted] •Machine learning is developed to predict the torrefaction severity index of torrefied biomass.•MARS model identifies the temperature as the most influential factor on TSI value.•Two hidden layers with 85 neurons provide a better performance of the ANN model.•The best-fit quality (R2) values of MARS and ANN are 0.9851 and 0.9784, respectively.•Three-dimensional TSI profiles from MARS and ANN are almost equivalent. Machine learning (ML) is one type of artificial intelligence (AI) commonly used for computer programming. Multivariate adaptive regression splines (MARS) and artificial neural networks (ANN) are two common and popular tools in AI that allow the user to analyze the pattern of complex data. The torrefaction severity index (TSI) is an index to define torrefied biomass quality at different torrefaction conditions. In this study, MARS and ANN models are applied to predict TSI. The considered input parameters in predictions using MARS and ANN approaches comprise feedstock type, temperature, and duration. The MARS model indicates that temperature is the most influential factor on TSI, followed by duration and feedstock type. In contrast, the ANN model reveals that the feedstock type is a dominant factor, and temperature and duration are not important. The performance of the ANN model is evaluated in three different combinations of numbers of hidden layers and neurons. It shows 2 hidden layers along with 85 neurons giving the best performance. The highest R2 values in MARS and ANN are 0.9851 and 0.9784, respectively. The relative root means square error analysis shows that both MARS and ANN have good fit quality with the relative errors of 1.49% and 2.16%, respectively. Overall, the comparison reflects that MARS is a more suitable model for predicting solid biofuel’s TSI. The general observation suggests that the ANN lacks sensitivity to the input parameter. Nevertheless, ANN performance may be improved by adjusting the number of hidden layers and neurons.