Application and comparison of Artificial Intelligent (AI) models on pyrolysis of oil palm biomass

Biomass is becoming an increasingly popular renewable resource due to its wider availability, carbon neutral nature and can be substitute to fossil fuels. In Malaysia, palm oil is a major commercial commodity and palm oil milling industrial process generates huge amounts of biomass in form of oil palm shell (OPS), empty fruit bunches (EFB), and oil palm fibers (OPF). These oil palm biomasses have potential to be converted into fuels and chemicals. Pyrolysis is one of the thermochemical process that is capable of converting biomass to solid (biochar), liquid (bio-oil) and gas products. There are numerous kinetic and reaction models that describe the pyrolysis behavior. However, they are often complicated and multiscale and therefore, not yet fully known to the researchers. To overcome this problem, in this study, two different types of AI models (Artificial Neural Network (ANN) and Decision Trees (DT)) was applied, adapted and compared to predict the pyrolysis behavior of OPS and EFB biomass. ANN is a machine learning algorithm based on biological neurons of the brain, while DT follows a flow-chart similar to a tree structure, depending on the decision rules conditioned by data features. Both AI algorithms are capable of modelling complex and non-linear relationships. This ability helps to predict the pyrolysis behavior of biomass. For this study, experimental data from thermogravimetric analyzer (TGA) machine were gathered under different conditions (heating rate) and given as input to the ANN and DT models for training and validation. Both models demonstrated high values (0.9992 for ANN and 0.9995 for DT) of coefficient of determination (R2) for validation results which signals that the modelled data closely followed the experimental data. While the mean square error (MSE) for validation results was in the range of 2.793 for ANN and 2.979 for DT. In future, AI models can be used to investigate the pyrolysis behavior of biomass without relying heavily on experiments in the TGA machines.