Scholarlink

<p>Electric arc furnace (EAF) steelmaking process represents a typical black box system characterized by high temperatures, strong nonlinearity, and multi-variable coupling. Challenges such as insufficient accuracy, poor model stability, and limited interpretability persist in the process for predicting the electricity power consumption (EPC). This paper proposed a physics guided cross fusion network (PGCFN), which integrates physics mechanisms with deep learning. By incorporating multi-layer physical injection, feature attention fusion, a dynamic hybrid loss function, and sharpness aware minimization (SAM) optimization, the model preserves physical consistency and achieves high-precision, stable, and robust predictions of EPC in EAF steelmaking process. The model was trained using real production data from 2595 heats. Results indicate that PGCFN achieves MAE, RMSE, and <i>R</i><sup>2</sup> values of 1372.86 kWh, 2046.32 kWh, and 0.774, respectively, demonstrating superior performance compared to other benchmark models in the references. Relative to the best-performing reference model, MAE and RMSE decreased by 16.24 and 5.19 pct, respectively, while <i>R</i><sup>2</sup> increased by 4.74 pct. Stability tests in random initialization and noise robustness experiments confirmed the stability and robustness of the model. The model converges reliably and shows high resilience to noise. Ablation studies reveal that each module in PGCFN contributes positively to overall performance. The proposed PGCFN maintains both predictive accuracy and physical consistency, offering an innovative and efficient solution for EPC prediction in EAF steelmaking process.</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

Physics-Guided Cross Fusion Network for Electricity Power Consumption Forecasting in Electric Arc Furnace Steelmaking

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Yuchi Zou,
  • Hang Hu,
  • Yuan Cai,
  • Lingzhi Yang,
  • Guangsheng Wei,
  • Yufeng Guo

摘要

Electric arc furnace (EAF) steelmaking process represents a typical black box system characterized by high temperatures, strong nonlinearity, and multi-variable coupling. Challenges such as insufficient accuracy, poor model stability, and limited interpretability persist in the process for predicting the electricity power consumption (EPC). This paper proposed a physics guided cross fusion network (PGCFN), which integrates physics mechanisms with deep learning. By incorporating multi-layer physical injection, feature attention fusion, a dynamic hybrid loss function, and sharpness aware minimization (SAM) optimization, the model preserves physical consistency and achieves high-precision, stable, and robust predictions of EPC in EAF steelmaking process. The model was trained using real production data from 2595 heats. Results indicate that PGCFN achieves MAE, RMSE, and R2 values of 1372.86 kWh, 2046.32 kWh, and 0.774, respectively, demonstrating superior performance compared to other benchmark models in the references. Relative to the best-performing reference model, MAE and RMSE decreased by 16.24 and 5.19 pct, respectively, while R2 increased by 4.74 pct. Stability tests in random initialization and noise robustness experiments confirmed the stability and robustness of the model. The model converges reliably and shows high resilience to noise. Ablation studies reveal that each module in PGCFN contributes positively to overall performance. The proposed PGCFN maintains both predictive accuracy and physical consistency, offering an innovative and efficient solution for EPC prediction in EAF steelmaking process.