Abstract
Biohydrogen has emerged as a promising renewable energy carrier with the potential to reduce dependence on fossil fuels and mitigate the associated environmental impacts. However, accurately predicting biohydrogen yield from biomass gasification remains inherently challenging due to the complex and nonlinear interactions among feedstock properties, operating conditions, and gasification kinetics. This study aims to address this gap by developing artificial neural network (ANN) models to predict biohydrogen yield from the gasification of sugarcane bagasse (SCB). Gasification offers a flexible thermochemical pathway capable of producing hydrogen-rich syngas from lignocellulosic biomass, while SCB presents an abundant agricultural residue with strong potential for sustainable waste valorisation. The study presents a data-driven approach to predicting biohydrogen yield from SCB gasification using ANN models. Although machine learning (ML) techniques have increasingly been applied to biohydrogen prediction problems, their use in modelling heterogeneous, literature-derived datasets specific to SCB gasification remains limited.
A literature-derived dataset was analysed, and correlation analysis was used to examine the influence of key process variables on biohydrogen yield. The correlation analysis revealed that temperature, gasification agent, air-to-biomass ratio, reactor type, and steam input were the most significant contributors. Supervised regression modelling was employed to compare the performance of the ANN against other ML algorithms, including linear regression (LR), Support Vector Regression (SVR), and Random Forest (RF). The models were evaluated using standard performance metrics, including the coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). To optimise the ANN model, four training algorithms were evaluated: Levenberg-Marquardt (L-M), Bayesian Regularisation (BR), Resilient Backpropagation (RP), and Scaled Conjugate Gradient (SCG).
The results showed that the ANN outperformed the other ML models, achieving the highest predictive accuracy with an R² of 0.742 and the lowest error metrics across RMSE (6.94), MAE (5.39), and MAPE (24.3%). SVR showed competitive performance, whereas LR and RF were less effective in capturing the nonlinear dynamics of biohydrogen production in this study. Further optimisation, conducted by evaluating four ANN training algorithms: L-M, BR, RP, and SCG, showed that the ANN model trained with the L-M algorithm achieved the highest overall predictive...