Coke quality prediction in the presence of data heterogeneity: a domain-knowledge-guided clustered modeling framework

Accurate prediction of coke quality is crucial for ensuring the stability and efficiency of blast furnace ironmaking processes. Traditional prediction methods typically treat complex production data as a single homogeneous dataset for direct modeling, which struggles to effectively address the inherent data heterogeneity arising from fluctuations in raw material composition, variations in process parameters, and diversity in operating conditions. This limitation often results in insufficient model generalizability and prediction accuracy that fails to meet practical industrial requirements. To address this challenge, this study proposes a novel domain knowledge-guided clustering-based modeling paradigm. The modeling process is systematically deconstructed into three hierarchical stages: semantic label generation, high-precision classifier construction, and differentiated regression prediction. By deeply integrating process mechanisms with data-driven methodologies, an integrated prediction framework is developed that accommodates multiple operating conditions and performance indicators. Experimental validation based on real industrial data demonstrates that, in predicting coke cold strength (DI 15150) and thermal properties (CSR, CRI), the proposed method significantly outperforms traditional single-model approaches in both prediction accuracy and stability, with prediction errors strictly confined within engineering acceptable limits. This study not only confirms the effectiveness of the clustering-based modeling approach for coke quality prediction but also provides a transferable solution framework for complex industrial process prediction problems characterized by similar data heterogeneity.