<p>The dephosphorization capacity of converter slag has played a vital role in controlling endpoint phosphorus content and achieving efficient steel refining. However, its prediction remains challenging due to the nonlinear influence of hot metal composition, slagging regime, and process fluctuations. In this study, a feature-weighted Stacking ensemble learning model, based on machine learning algorithms including LightGBM, XGBoost, SVM, and KNN, is proposed to predict slag dephosphorization capacity, using slag basicity, total iron content (T.Fe), and P₂O₅ as key indicators. The model employs diverse base learners and SHAP-based feature weighting, trained and cross-validated on real converter steelmaking data to enhance both interpretability and predictive performance. Monte Carlo simulations were conducted to evaluate uncertainty propagation under T.Fe, basicity, and combined perturbations, revealing that basicity is the most influential factor affecting P₂O₅ prediction. The final dephosphorization model attains an MAE of 5.34, RMSE of 7.69, and <i>R</i>² of 89.3 pct, demonstrating high accuracy and robustness. The optimal process parameters are: slag basicity 2.8 to 3.59, T.Fe 18-24 pct, and hot metal silicon content 0.22 to 0.46 pct. This method offers an effective tool for the intelligent control of dephosphorization in converter steelmaking.</p>