Scholarlink

<p>This study presents a novel hybrid modeling approach for carbon content prediction in Electric Arc Furnace (EAF) steelmaking by integrating mechanistic models with machine learning (ML)-based error correction. The mechanistic model, based on carbon mass balance principles, achieved limited accuracy (<i>R</i><sup>2</sup> of 0.2498, RMSE of 0.0117) due to simplified assumptions regarding carbon losses and material composition variations. To address these limitations, six decision tree-based ensemble learning algorithms were implemented as error correction modules, forming hybrid prediction models. Feature selection using ReliefF algorithm identified nine optimal features that significantly enhanced model performance. The ReliefF-CatBoost hybrid model demonstrated superior results with 202.31&#xa0;pct improvement in <i>R</i><sup>2</sup> and 52.14&#xa0;pct reduction in RMSE compared to the pure mechanistic model. Hyperparameter optimization using Teaching-Learning Optimization (TLO) and Grey Wolf Optimization (GWO) algorithms further enhanced performance, with the TLO-optimized ReliefF-CatBoost hybrid model achieving an <i>R</i><sup>2</sup> of 0.9018, RMSE of 0.0048, and MAX of 0.0094. SHAP interpretability analysis confirmed that the ML module effectively captured metallurgical mechanisms, with coke powder consumption and oxygen consumption identified as the most influential factors. Actual validation using 300 heats demonstrated exceptional hit rates, with 95.67&#xa0;pct within ±&#xa0;0.03&#xa0;pct accuracy, significantly outperforming traditional mechanistic and pure ML approaches. The proposed hybrid strategy combines metallurgical mechanisms with the non-linear fitting ability of ML, providing a reliable solution for carbon prediction in EAF.</p> Graphical Abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Prediction Model for Endpoint Carbon Content in the Electric Arc Furnace Based on the Fusion of Metallurgical Mechanisms and Decision Tree-Based Ensemble Learning Algorithms

  • 期刊论文

  • 发表日期：

    2026年1月21日

查看全文
  • Hongbin Lu,
  • Hongchun Zhu,
  • Zhouhua Jiang,
  • Huabing Li,
  • Ce Yang

摘要

This study presents a novel hybrid modeling approach for carbon content prediction in Electric Arc Furnace (EAF) steelmaking by integrating mechanistic models with machine learning (ML)-based error correction. The mechanistic model, based on carbon mass balance principles, achieved limited accuracy (R2 of 0.2498, RMSE of 0.0117) due to simplified assumptions regarding carbon losses and material composition variations. To address these limitations, six decision tree-based ensemble learning algorithms were implemented as error correction modules, forming hybrid prediction models. Feature selection using ReliefF algorithm identified nine optimal features that significantly enhanced model performance. The ReliefF-CatBoost hybrid model demonstrated superior results with 202.31 pct improvement in R2 and 52.14 pct reduction in RMSE compared to the pure mechanistic model. Hyperparameter optimization using Teaching-Learning Optimization (TLO) and Grey Wolf Optimization (GWO) algorithms further enhanced performance, with the TLO-optimized ReliefF-CatBoost hybrid model achieving an R2 of 0.9018, RMSE of 0.0048, and MAX of 0.0094. SHAP interpretability analysis confirmed that the ML module effectively captured metallurgical mechanisms, with coke powder consumption and oxygen consumption identified as the most influential factors. Actual validation using 300 heats demonstrated exceptional hit rates, with 95.67 pct within ± 0.03 pct accuracy, significantly outperforming traditional mechanistic and pure ML approaches. The proposed hybrid strategy combines metallurgical mechanisms with the non-linear fitting ability of ML, providing a reliable solution for carbon prediction in EAF.

Graphical Abstract