Beyond Elemental Composition: Leveraging Fuel Classification Systems for Robust Machine Learning-Based Higher Heating Value Prediction

Accurate prediction of higher heating value (HHV) is essential for the design, optimization, and performance assessment of energy conversion systems involving biomass, fossil fuels, and waste-derived materials. Although machine learning-based approaches have demonstrated strong predictive capability using ultimate and proximate analysis variables, most existing studies implicitly assume that fuels constitute a homogeneous population once numerical descriptors are provided, thereby overlooking the potential value of fuel classification information. In this study, fuel classification is reframed as an explicit and quantifiable information source for HHV modeling through a unified machine learning framework that integrates three independent classification systems, ECN Phyllis, NTA 8003, and the data-driven HOM Classification System, alongside numerical compositional features using one-hot encoding. A dataset of 929 solid fuel samples was used to evaluate multiple regression models under a consistent five-fold cross-validation protocol with structured hyperparameter optimization. To move beyond aggregate performance metrics, a comprehensive analysis framework combining correlation screening, permutation-based feature importance, ΔRMSE evaluation, ablation experiments, and SHAP-based interpretability analysis was employed to quantify the contribution of individual features and classification systems. The results demonstrate that incorporating fuel classification information leads to consistent improvements in HHV prediction accuracy compared to numerical-only baselines. Among the examined systems, the HOM Classification System provides the strongest and most robust contribution, supplying non-redundant predictive information beyond elemental composition and supporting stable performance across heterogeneous fuel categories. Overall, the findings establish fuel classification, particularly the HOM system, as an active predictor rather than descriptive metadata and offer a more generalizable, interpretable, and robust framework for HHV prediction in biomass and waste-to-energy applications.