This study presents a comprehensive data-driven framework for predicting energy consumption in an industrial aluminum cold rolling mill. Cold rolling is one of the most energy-intensive stages of aluminum production. While traditional physics-based models are insufficient for real-time applications, existing machine learning studies have largely overlooked aluminum-specific process dynamics and the interpretability dimension. An industrial dataset consisting of 1, 345 production records from 2022 was used. Physically meaningful features such as reduction ratio, thickness reduction, and specific energy consumption (SEC, kWh/ton) were derived from raw process variables. Linear regression, Ridge, and Lasso were employed as baseline models, and compared against ensemble learners including Random Forest, Gradient Boosting, XGBoost, LightGBM, and CatBoost. Under the random split, CatBoost achieved the best performance (\ (R² = 0. 78\) ), whereas under the more realistic temporal split, Random Forest and Gradient Boosting emerged as the top performers (\ (R² 0. 81\) ). Linear models degraded dramatically under temporal evaluation (\ (R² 0. 18\) ). SHAP analyses revealed that CVC position and rolling speed are the dominant drivers of model predictions. The ablation study demonstrated that removing the CVC position feature group caused the \ (R²\) score to drop from 0. 80 to 0. 20. The proposed framework addresses a significant gap in the literature by jointly offering: rolling-aware energy indicator derivation, a fair comparison of ensemble model families, SHAP-based interpretability, and a CO \ (₂\) accounting analysis grounded in SEC-based efficiency stratification.
Ağgül et al. (Sat,) studied this question.