Revisiting data-driven approaches for methane and hydrogen production prediction with emphasis on explainability, applicability, and industrial implications

Abstract

This study aims to enhance the accuracy, explainability, and industrial applicability of methane (CH₄) and hydrogen (H₂) yield predictions in supercritical water gasification (SCWG), a promising method for biofuel production. Traditional approaches often lack precision and scalability, necessitating advanced data-driven solutions. We revisit machine learning (ML) models, focusing on hyperparameter optimisation (Grid Search, Random Search, Bayesian Optimisation), model explainability (LIME, SHAP, permutation importance), and industrial implications. Using SCWG data from Yimin lignite, six ML models and two ensembles were evaluated with MAE, RMSE, and R² metrics under 5-fold cross-validation. Random Forest outperformed others, achieving R² values of 0.995 for CH₄ and 1.00 for H₂, with low errors. Explainability analyses identified temperature and CE as the most critical factors, while optimisation techniques showed minimal performance differences, supporting simpler optimisation techniques like Random search in resource-constrained settings. Williams plots confirmed robust applicability, with over 99.9 % of data points within acceptable ranges (±2 residuals and low leverage). These findings enhance SCWG process efficiency, reduce costs, and minimise environmental impacts, advancing sustainable energy solutions.