A graph-based direct association strategy with fuzzy interpretability for linking individual coal properties in coal blending schemes to coke quality

Existing coke quality prediction methods mainly rely on a two-stage prediction strategy, where blended coal properties are first estimated from individual coal properties and then used for coke quality prediction. This process inevitably introduces information loss and accumulated prediction errors, which limit the prediction accuracy and reliability of these models in practical coal blending applications. To address this challenge, this study proposes, for the first time, a graph-based direct association strategy linking individual coal properties in coal blending schemes to coke quality. Each coal blending scheme is transformed into a graph structure, where individual coal samples are regarded as graph nodes and their corresponding coal properties together with blending ratios are used as node features. Different graph neural network models are further investigated to directly learn the relationship between individual coal properties and coke quality indicators, namely Coke Reactivity Index (CRI) and Coke Strength after Reaction (CSR). A total of 306 coal blending schemes are used for model development and evaluation. Experimental results show that the proposed one-stage graph-based framework outperforms traditional two-stage prediction methods. Among different graph neural network models, the graph attention network achieves the best overall prediction performance, obtaining mean absolute errors of 2.06 and 2.60 for CRI and CSR prediction, respectively. The t-SNE visualization further demonstrates that the proposed one-stage framework can learn more discriminative feature representations than the traditional two-stage framework. In addition, the proposed framework demonstrates good robustness under blending conditions with different numbers of coal types and can provide interpretable analysis through feature importance and node importance visualization. Furthermore, fuzzy interpretability is introduced at the feature level to characterize the gradual contribution levels of different coal properties.