<p>Precise control of slag composition is critical for steel quality in ladle furnace refining, but challenges like complex nonlinear reactions, multi-steel grade data heterogeneity, and experience-dependent control limit accuracy. This study proposes a data-driven prediction method integrating steel grade information granulation, operating condition clustering, and multi-output machine learning. First, data granulation is performed on the steel grade process parameters to achieve unification of dimensions; subsequently, the performance indicators of five clustering algorithms are compared, and the optimal clustering algorithm is selected. Finally, one-hot encoding is applied to the steel grade clustering results of this algorithm, and the encoded results are fused with 25 key process features. The fused feature dataset is used as input to train five regression prediction models, while the Bayesian optimization algorithm is adopted to complete the hyperparameter tuning process of the models. Validation results based on 1780 industrial heats indicate that during the clustering phase, the DBSCAN algorithm exhibits the optimal performance in mining process correlations with a silhouette coefficient of 0.987; during the regression prediction phase, the XGBoost model performs best, with <i>R</i>² values for CaO, SiO₂, and Al₂O₃ reaching 0.756, 0.985, and 0.947, respectively, and over 90 pct hit rate within 5 pct relative error for main components. This method addresses the gap of multi-component synchronous prediction, reduces data heterogeneity, enhances reliability, and provides real-time guidance for adjusting slag-making agents and alloys, cutting production costs and supporting intelligent ladle furnace refining control.</p>