Scholarlink

<p>Accurate endpoint temperature prediction is essential for efficient electric arc furnace (EAF) steelmaking but remains a persistent challenge. This study innovatively proposes a hybrid modeling framework integrating metallurgical mechanisms with machine learning (ML), employing Tabular Prior-data Fitted Networks (TabPFN) in the ML component to enhance prediction performance. Three hybrid strategies were evaluated: (1) feature augmentation, using mechanistic predictions as additional inputs; (2) error correction, using ML to correct systematic biases in the mechanistic model; and (3) dynamic weighted fusion, implementing cluster-based adaptive weight allocation across varying operational conditions. All hybrid models achieved R<sup>2</sup> values above 0.88 on the test dataset. The error correction strategy performed best, reducing RMSE by 56.12 pct and MAE by 66 pct compared to the mechanistic baseline, with an 94.89 pct hit rate within ±10&#xa0;°C. Benchmarking against optimized gradient boosting models demonstrated TabPFN’s effectiveness in small-sample industrial prediction, delivering RMSE and MAE improvements of 15.12 and 19.93 pct. SHAP analysis confirmed the ML component’s alignment with metallurgical principles and reliability. The robustness and practical value of this framework were validated through 300 industrial heats. This study pioneers a novel paradigm integrating metallurgical principles with the Transformer architecture, overcoming the inherent limitations of purely mechanism-based modeling or ML approaches, advancing intelligent EAF prediction technology.</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

Assessment of Multiple Hybrid Modeling Approaches Combining Mechanistic and Machine Learning Methods for Endpoint Temperature Prediction in Electric Arc Furnace

  • 期刊论文

  • 发表日期：

    2026年1月9日

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

摘要

Accurate endpoint temperature prediction is essential for efficient electric arc furnace (EAF) steelmaking but remains a persistent challenge. This study innovatively proposes a hybrid modeling framework integrating metallurgical mechanisms with machine learning (ML), employing Tabular Prior-data Fitted Networks (TabPFN) in the ML component to enhance prediction performance. Three hybrid strategies were evaluated: (1) feature augmentation, using mechanistic predictions as additional inputs; (2) error correction, using ML to correct systematic biases in the mechanistic model; and (3) dynamic weighted fusion, implementing cluster-based adaptive weight allocation across varying operational conditions. All hybrid models achieved R2 values above 0.88 on the test dataset. The error correction strategy performed best, reducing RMSE by 56.12 pct and MAE by 66 pct compared to the mechanistic baseline, with an 94.89 pct hit rate within ±10 °C. Benchmarking against optimized gradient boosting models demonstrated TabPFN’s effectiveness in small-sample industrial prediction, delivering RMSE and MAE improvements of 15.12 and 19.93 pct. SHAP analysis confirmed the ML component’s alignment with metallurgical principles and reliability. The robustness and practical value of this framework were validated through 300 industrial heats. This study pioneers a novel paradigm integrating metallurgical principles with the Transformer architecture, overcoming the inherent limitations of purely mechanism-based modeling or ML approaches, advancing intelligent EAF prediction technology.

Graphical Abstract