<p>Accurate prediction of slag compositions in converter is essential for optimizing steelmaking efficiency and improving resource utilization. A novel stacking deep network was established to enhance the accuracy of slag compositions prediction in converter. The established network was designed to incorporate deep learning as part of a collection of multiple base models, combined with tree-based models. This network introduced more complex decision boundaries to the axis-aligned geometry, thereby improving model diversity. This network differs from existing ensemble methods by using a first layer to integrate multiple base learners for diverse features capture, while the second layer’s skip connection mechanism reuses input features to address covariate shift and improve model expressiveness and stability. In addition, this network introduced an innovative stacking layer design, replacing the standard linear regression model with more sophisticated tree-based models and deep learning networks. This enhancement allowed the stacking layer not only to weight the outputs of the base models but also to learn deeper patterns and feature interactions. To improve the network’s generalization, K-fold bagging was employed to minimize the risk of overfitting. Feature importance analysis further underscores the established network’s ability to identify the key factors influencing slag compositions, offering valuable insights for optimizing the steelmaking process. Experimental results demonstrated that the established network significantly outperforms existing methods in predicting slag compositions in converter, with mean squared errors of 4.33, 2.80, 3.39, and 0.49 for CaO, SiO₂, MgO and FeO, respectively, highlighting its robust generalization capability. The established network approach offers an effective solution for predicting slag compositions in converters, thereby overcoming the limitation of delayed slag detection.</p>