Scholarlink

<p>Effective management of thermal load was crucial for maintaining stable operating conditions in blast furnaces. This study proposed a thermal load prediction model based on feature analysis and deep learning to assist operators in proactively monitoring and managing blast furnaces. This yielded interesting research findings. First, data governance techniques were applied to establish a standardized governance framework for blast furnace data, thereby improving data quality. Second, applying variational mode decomposition, the thermal load was adaptively decomposed into seven data components, categorized into periodic and trend-type data. Through maximal information coefficient analysis, blast air humidity and oxygen-enriched flow exhibited their strongest effects on thermal load intrinsic mode function1 after 2 hours, theoretical combustion temperature reached its peak influence after 3 hours, which aligns with real-world blast furnace dynamics. Third, for periodic and trend-based data, the deep forest and bidirectional long short-term memory algorithms were selected, respectively. Concurrently, the sparrow search algorithm was applied to enhance and optimize hyperparameters, constructing the SSA-DL model to predict the next hour’s heat load. The SSA-DL model demonstrated high prediction accuracy, achieving a mean absolute error of 182.88, a root mean squared error of 221.57, an R-squared value of 0.91, and a hit rate of 91.25 pct within ±350 MJ·h<sup>−1</sup>. The integration of the optimization algorithm reduced the base model error by over 20 pct and increased the average hit rate by more than 15 pct. Finally, considering the variability of blast furnace industrial data, an online update mechanism for the predictive model was designed. Industrial validation was conducted using the latest 30-day data. This enhancement improved the model’s actual prediction hit rate by 10.25 pct. The validation results demonstrated that the predictive model achieved high accuracy and strong stability, which contributed to operators maintaining thermal load within stable operational ranges.</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

Thermal Load Prediction Model Based on Feature Analysis and Deep Learning

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Jue Tang,
  • Minghui Ma,
  • Quan Shi,
  • Mansheng Chu,
  • Zhen Zhang,
  • Qichen Hu

摘要

Effective management of thermal load was crucial for maintaining stable operating conditions in blast furnaces. This study proposed a thermal load prediction model based on feature analysis and deep learning to assist operators in proactively monitoring and managing blast furnaces. This yielded interesting research findings. First, data governance techniques were applied to establish a standardized governance framework for blast furnace data, thereby improving data quality. Second, applying variational mode decomposition, the thermal load was adaptively decomposed into seven data components, categorized into periodic and trend-type data. Through maximal information coefficient analysis, blast air humidity and oxygen-enriched flow exhibited their strongest effects on thermal load intrinsic mode function1 after 2 hours, theoretical combustion temperature reached its peak influence after 3 hours, which aligns with real-world blast furnace dynamics. Third, for periodic and trend-based data, the deep forest and bidirectional long short-term memory algorithms were selected, respectively. Concurrently, the sparrow search algorithm was applied to enhance and optimize hyperparameters, constructing the SSA-DL model to predict the next hour’s heat load. The SSA-DL model demonstrated high prediction accuracy, achieving a mean absolute error of 182.88, a root mean squared error of 221.57, an R-squared value of 0.91, and a hit rate of 91.25 pct within ±350 MJ·h−1. The integration of the optimization algorithm reduced the base model error by over 20 pct and increased the average hit rate by more than 15 pct. Finally, considering the variability of blast furnace industrial data, an online update mechanism for the predictive model was designed. Industrial validation was conducted using the latest 30-day data. This enhancement improved the model’s actual prediction hit rate by 10.25 pct. The validation results demonstrated that the predictive model achieved high accuracy and strong stability, which contributed to operators maintaining thermal load within stable operational ranges.