Abstract: RH refining constitutes a critical procedure in steelmaking processes. The end-point carbon content serves as a vital indicator affecting steel properties. Accurate prediction of end-point carbon content in RH refining is of great significance for optimizing process control and guaranteeing finished product quality. To precisely predict the end-point carbon content, models are established by random forest (RF), linear regression (LR), support vector regression (SVR) and extreme gradient boosting (XGBoost) algorithms combined with recursive feature elimination (RFE). In the selection of feature variables for model training, different from the conventional heatmap method, a dynamic range strategy is proposed. A reasonable interval for the preset number of features is defined, each model within this interval is evaluated through grid search of model hyperparameters, and the model with optimal performance is selected to avoid limitations brought by a single feature quantity. Verification results based on 102 sets of on-site production data demonstrate that the model trained by the SVR algorithm with 18 feature variables screened via RFE achieves the optimal performance in predicting end-point carbon content during RH refining. The mean absolute error and mean squared error reach 1.427×10^-6 and 6.626×10^-10, respectively. A prediction hit rate of 93.5% is obtained under the precision requirement of ±5×10^-6.