SARF-enhanced CNN-transformer stacking ensemble for hot-metal ladle temperature drop prediction

Accurate prediction of hot-metal temperature drop during ladle transportation is important for transfer scheduling, energy management, and stable steelmaking operations. This study proposes a hybrid framework that combines a multi-scale CNN–Transformer encoder with a stacking ensemble for hot-metal temperature-drop prediction. The multi-scale CNN captures local variation patterns, the Transformer models long-range temporal dependencies, and a Scale-Adaptive Residual Fusion (SARF) module is introduced to fuse the two types of representations. Based on the extracted deep features, LightGBM, CatBoost, and Random Forest are integrated through a fixed XGBoost meta-learner. Hyperparameters are configured using a stage-wise Whale Optimization Algorithm (WOA) strategy. Under a chronological 80/20 train-test split, the proposed framework achieves R² = 0.924, RMSE = 7.017 °C, and hit rates of 85.754% and 95.531% within ±10 °C and ±15 °C, respectively. In addition, a 5-fold rolling-origin validation conducted within the training set yields mean R² = 0.8811 ± 0.0089 and mean RMSE = 9.0882 ± 0.2719 °C, indicating stable performance across temporally distinct validation segments. To provide process insight, SHAP feature-attribution analysis is performed as a post-hoc explanation to identify dominant factors and support practical process optimization.