Scholarlink

<p>This study addresses the issue of decreasing accuracy in data-driven converter endpoint prediction models over time due to changing conditions. A self-learning model, coupled with a system to detect and correct abnormal molten iron data, is proposed. It has been confirmed that the model for supplementing abnormal data can improve its adaptability to a certain extent, but there is still significant room for improvement in the selection and quantity of supplementary variables. Comparative analysis of self-learning algorithms revealed random forest (RF) outperformed error back propagation (BP) and convolutional neural networks (CNN) in prediction accuracy, achieving 65.7% (±&#xa0;10°C) and 84% (±&#xa0;15°C) for endpoint temperature, 61.2% (±&#xa0;0.015&#xa0;wt.%) and 74.1% (±&#xa0;0.020&#xa0;wt.%) for carbon content. Meanwhile, self-learning models showed significant improvements over non-self-learning models, with temperature prediction accuracy increasing by 10.5% (±&#xa0;10°C) and 9.1% (±&#xa0;15°C), and carbon content accuracy rising by 3.9% (±&#xa0;0.015&#xa0;wt.%) and 1.1% (±&#xa0;0.020&#xa0;wt.%). This study provides a new method for the long-term application of data-driven converter endpoint prediction models in the field. However, data-driven prediction accuracy has an upper limit, likely due to industrial data anomalies. Accurate data collection and preprocessing are crucial. Therefore, for industrial data-driven prediction models, future data refinement collection and processing are crucial steps.</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

Data Driven Self-Learning Model for Supplementing Abnormal Molten Iron Information and Predicting Converter Endpoint Carbon Content and Temperature

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Tongze Xin,
  • Min Wang

摘要

This study addresses the issue of decreasing accuracy in data-driven converter endpoint prediction models over time due to changing conditions. A self-learning model, coupled with a system to detect and correct abnormal molten iron data, is proposed. It has been confirmed that the model for supplementing abnormal data can improve its adaptability to a certain extent, but there is still significant room for improvement in the selection and quantity of supplementary variables. Comparative analysis of self-learning algorithms revealed random forest (RF) outperformed error back propagation (BP) and convolutional neural networks (CNN) in prediction accuracy, achieving 65.7% (± 10°C) and 84% (± 15°C) for endpoint temperature, 61.2% (± 0.015 wt.%) and 74.1% (± 0.020 wt.%) for carbon content. Meanwhile, self-learning models showed significant improvements over non-self-learning models, with temperature prediction accuracy increasing by 10.5% (± 10°C) and 9.1% (± 15°C), and carbon content accuracy rising by 3.9% (± 0.015 wt.%) and 1.1% (± 0.020 wt.%). This study provides a new method for the long-term application of data-driven converter endpoint prediction models in the field. However, data-driven prediction accuracy has an upper limit, likely due to industrial data anomalies. Accurate data collection and preprocessing are crucial. Therefore, for industrial data-driven prediction models, future data refinement collection and processing are crucial steps.