Intelligent prediction of energy consumption for the billet reheating of hot-rolled seamless steel pipe

Abstract

Seamless steel pipes are essential engineering materials, yet reducing heating energy consumption in their production remains challenging. Focusing solely on overall furnace energy consumption limits refined energy management. This study proposes an enhanced energy allocation model that incorporates both billet weight and residence time in the furnace, overcoming the difficulty of obtaining reliable energy consumption for single billet reheating (ECSBR). Key factors influencing ECSBR are identified through a hybrid strategy combining thermodynamic mechanism analysis with XGBoost-based feature importance ranking. Four prediction models are compared: two single models (Multilayer Perceptron, Support Vector Regression) and two ensemble models (Random Forest, Extreme Gradient Boosting). Results show that ensemble learning significantly outperforms single models, achieving an R² improvement of over 2.2% on the test dataset. The XGBoost model delivers the best performance (RMSE = 2.065 kgce/t, R² = 0.977), with 95% of absolute prediction errors within 5 kgce/t. Coupled with SHAP-based interpretability, the proposed framework not only provides high-precision ECSBR prediction but also translates black-box outputs into actionable insights for plant operators, providing a scientific reference for refined energy management and abnormal energy consumption diagnosis in hot-rolled seamless steel pipe production.