Prediction of Electric Arc Furnace End‐Point Temperature via Fusion of Physical Mechanism and Stacking Ensemble

Accurate endpoint temperature prediction in electric arc furnace (EAF) steelmaking is critical for energy efficiency and product quality; however, it remains fundamentally challenged by severe nonlinearity and dynamic variations. Traditional machine learning models lack interpretability and robustness against industrial noise, whereas purely mechanistic models fail to capture nonlinear deviations. To address these limitations, we propose Phy‐Stacking, a novel mechanism‐guided hybrid prediction framework. Initially, mechanistic features (net specific energy and dynamic heat loss) are formulated based on energy conservation principles to enrich raw EAF data representations. Subsequently, a dual‐layer intelligent filter—integrating an Isolation Forest with a Random Forest residual‐based filtration—systematically eliminates anomalous samples. Finally, linear regression establishes a stable physical baseline, while a customized stacking ensemble (incorporating XGBoost, Random Forest, and Ridge regression) compensates for uncaptured nonlinear residuals. Extensive validation on real‐world industrial datasets demonstrates that Phy‐Stacking achieves a root mean square error (RMSE) of 5.58 °C, a mean absolute error (MAE) of 4.31 °C, and an accuracy of 92.09% within a ±10 °C tolerance window. Comprehensive ablation studies confirm that physical feature reconstruction, dual‐layer data cleansing, and the residual‐driven stacking architecture synergistically enhance predictive accuracy and model robustness.