High‐Precision Prediction Model for Electric Arc Furnace Steel Endpoint Temperature Using Interpretable Machine Learning With In‐Depth Analysis of Driving Factors

This study presents a high‐accuracy, fully interpretable, data‐driven framework for predicting endpoint temperature in Electric Arc Furnace (EAF) steelmaking. Based on a rigorous metallurgical energy balance, 58 process variables were initially assembled. After outlier removal and feature selection using the Distance Correlation Coefficient, 18 variables highly correlated with the prediction target were retained. Twelve state‐of‐the‐art machine learning algorithms were evaluated, with gradient‐boosting ensemble models demonstrating the best performance. Global hyperparameter tuning was conducted using an Improved Grey Wolf Optimizer (IGWO), yielding an optimized LightGBM model (IGWO‐LightGBM). The model achieved an RMSE of 4.26°C and an R 2 of 0.967 on the test set, improving accuracy by 37% over the baseline unoptimized LightGBM model. Industrial validation on 300 unseen heats confirmed robust generalization, with hit rates of 81% within ±5°C and 94.3% within ±10°C. Interpretability was ensured through TreeSHAP, individual conditional expectation, and two‐dimensional partial dependence plots, which verified consistency with metallurgical principles and revealed actionable interactions among electrical energy input, oxygen supply, carbon injection, and hot‐heel management. The proposed framework not only surpasses existing EAF temperature‐prediction methods but also provides transparent, mechanism‐aligned insights to support energy‐efficient, digitally enabled steelmaking.