Abstract: Based on the actual production data of 260 t converter, prediction models of the main components of endpoint slag,CaO, SiO₂, TFe and MgO, were established through four algorithms of machine learning algorithm,XGBoost (eXtreme Gradient Boosting), elastic regression, linear regression and AdaBoost (Adaptive Boosting). By optimizing the parameters, the determination coefficients R² of XGBoost endpoint slag composition prediction model is all above 0.8.The slag splashing time model was modeled using five algorithms: SVR (Support Vector Regression), LGBM (Light Gradient Boosting Machine), GBDT (Gradient Boosting Decision Tree), RF (Random Forest) and XGBoost, respectively. Then, the integrated slag splashing time model was obtained by integrating SVR, XGBoost and GBDT. Stacking integrated slag splashing time model improves the prediction effect of each single model, and the prediction hit rate is 89.95 % in the error range of ±20 s.

Key words: converter, slag splashing, endpoint slag composition, slag splashing time, prediction model, machine learning