Scholarlink

<p>Accurate prediction of end point molten steel temperature during the Vacuum Degassing (VD) refining process is critical for improving product quality and energy efficiency. This study proposed a hybrid Transformer–Long Short-Term Memory (LSTM) model integrating global feature representation (via self-attention) and local temporal dynamics, with hyperparameter optimization via Optuna. Industrial data from a domestic steel plant encompassing 12 key process parameters were preprocessed by 3<i>σ</i> criteria and Min–Max normalization. The proposed model achieved a remarkable prediction accuracy of 91.5% within a ± 3&#xa0;°C error margin, with <i>R</i><sup>2</sup> = 0.99, RMSE = 1.82 and MAE = 1.25. This performance significantly surpasses that of baseline LSTM (83%, <i>R</i><sup>2</sup> = 0.97, RMSE = 2.52, MAE = 2.04) and Transformer (75%, <i>R</i><sup>2</sup> = 0.97, RMSE = 2.61, MAE = 2.53) models. Furthermore, the model demonstrated robustness in tracking temperature inflection points amidst process disturbances. By enabling more precise end point temperature control, this approach reduces the overheating margin by over 4.0&#xa0;°C compared with conventional methods, which typically operate within a ± 5–10&#xa0;°C tolerance band. This enhanced precision provides an effective AI-driven solution for low-carbon intelligent control in VD refining, with potential applications in other secondary metallurgy processes and electric arc furnaces, thereby contributing directly to industrial decarbonization and energy optimization goals.</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 Transformer–LSTM Hybrid Model for End Point Molten Steel Temperature Prediction in Vacuum Degassing Refining

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Kai Zhao,
  • Yuhan Ge,
  • Jiayong Qiu,
  • Xiaoyan Lv,
  • Jun Bian,
  • Xinzhe Lu,
  • Jingyu Zhao,
  • Yingjiang Wen

摘要

Accurate prediction of end point molten steel temperature during the Vacuum Degassing (VD) refining process is critical for improving product quality and energy efficiency. This study proposed a hybrid Transformer–Long Short-Term Memory (LSTM) model integrating global feature representation (via self-attention) and local temporal dynamics, with hyperparameter optimization via Optuna. Industrial data from a domestic steel plant encompassing 12 key process parameters were preprocessed by 3σ criteria and Min–Max normalization. The proposed model achieved a remarkable prediction accuracy of 91.5% within a ± 3 °C error margin, with R2 = 0.99, RMSE = 1.82 and MAE = 1.25. This performance significantly surpasses that of baseline LSTM (83%, R2 = 0.97, RMSE = 2.52, MAE = 2.04) and Transformer (75%, R2 = 0.97, RMSE = 2.61, MAE = 2.53) models. Furthermore, the model demonstrated robustness in tracking temperature inflection points amidst process disturbances. By enabling more precise end point temperature control, this approach reduces the overheating margin by over 4.0 °C compared with conventional methods, which typically operate within a ± 5–10 °C tolerance band. This enhanced precision provides an effective AI-driven solution for low-carbon intelligent control in VD refining, with potential applications in other secondary metallurgy processes and electric arc furnaces, thereby contributing directly to industrial decarbonization and energy optimization goals.

Graphical Abstract