摘要
转炉终点的温度和碳含量是转炉终点控制的重要参数。转炉终点特征因子的传统分析方法依赖人工经验,但随着数据量的急剧增长和工艺复杂性的增加,人工分析的方法难以充分挖掘转炉冶炼数据中的潜在关系,无法对转炉终点控制的特征因子进行深入理解。提出了一种神经网络特征提取与树模型回归相融合的新方法,用于寻找特征因子中复杂关系,进一步精确预测转炉终点的温度和碳含量。通过构建可调节的神经网络作为特征提取器,采用贝叶斯优化方法自动搜索最佳网络结构,设计适用于转炉流程的特征提取网络。其次,利用LightGBM树模型对特征因子进行回归分析。采用自定义的训练过程,将树模型的预测误差反向传播至特征提取网络,更新网络权重。结果表明,该方法显著提高了预测精度,终点温度±15和±20℃误差内的命中率分别达到91.80%和96.12%,终点碳±0.01%和±0.02%误差内的命中率分别达到90.72%和96.36%,且其在现场落地后,转炉一倒出钢率提升了10%。
The endpoint temperature and carbon content are crucial parameters for controlling the converter endpoint.Traditional analysis methods often rely on long-term accumulated human experience to manually analyze the characteristic factors affecting the converter endpoint.However,with the sharp increase in data volume and the complexity of the process,relying solely on traditional manual analysis has become insufficient to fully exploit the potential relationships in converter smelting data,limiting the deep understanding of the key characteristic factors affecting converter endpoint control.A new method was proposed that integrates neural network feature extraction with tree model regression to identify complex relationships within characteristic factors and further accurately predicts the converter′s endpoint temperature and carbon content.A tunable neural network was constructed as a feature extractor,using Bayesian optimization to automatically search for the optimal network structure,and designing a feature extraction network suitable for converter operations.Secondly,the LightGBM tree model is used for regression analysis of the characteristic factors.A custom training process is adopted,where the prediction error of the tree model is backpropagated to the feature extraction network to update the network weights.The results show that this method significantly improves prediction accuracy,with hit rates of 91.80%and 96.12%within errors of±15 and±20℃for endpoint temperature and 90.72%and 96.36%within errors of±0.01%and±0.02%for endpoint carbon,respectively.Moreover,after its implementation on-site,the steel tapping rate of the converter has increased by 10%.
作者
潘佳
孙中强
刘晓航
李光强
王强
PAN Jia;SUN Zhongqiang;LIU Xiaohang;LI Guangqiang;WANG Qiang(Key Laboratory for Ferrous Metallurgy and Resources Utilization of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China;Steelmaking Plant,Yangchun Xin Iron&Steel Co.,Ltd.,Yangjiang 529600,Guangdong,China)
出处
《钢铁研究学报》
北大核心
2025年第7期866-877,共12页
Journal of Iron and Steel Research
基金
国家自然科学基金区域创新发展联合基金资助项目(U22A20173)。
关键词
终点温度
终点碳含量
神经网络
特征提取
LightGBM
回归分析
endpoint temperature
endpoint carbon content
neural network
feature extraction
LightGBM
regression analysis
作者简介
潘佳(1991-),男,硕士生,工程师,E-mail:1710801141@qq.com;通信作者:王强(1989-),男,博士,教授,E-mail:wangqiangwust@wust.edu.cn。
