摘要
转炉炼钢终点控制的关键是实现熔池内碳含量和温度的准确预测。针对单一即时学习(JITL)模型在相似性度量时无法考虑样本的工况信息且易受噪声影响从而影响预测精度的问题,提出一种互信息加权度量相似样本选择的即时-集成学习(JITL-EL)预测模型。首先,利用谱聚类算法将历史数据样本划分为若干子集,使得各子集之间的差异最大化,而子集内部的相似性最大化,从而实现工况的有效区分;其次,根据类内特征与类内目标变量之间的关联度提出一种类内特征加权的后验概率计算方法得以获得待测样本隶属于不同聚类子集的隶属度;然后,根据待测样本与聚类子集的隶属度,从不同的聚类子集中利用类内特征加权的度量方法动态挑选不同数目的相似样本分别构建JITL基模型学习器;最后,通过当前待测样本隶属于不同聚类子集的隶属度将不同聚类子集的JITL基模型预测值加权融合,得到最终预测结果。通过钢厂采集的真实数据仿真结果表明,碳含量的准确率为80.00%,而温度的准确率则达到了83.00%。
The key to endpoint control in BOF steelmaking is accurately predicting carbon content and temperature within the molten bath.To address the limitations of single just-in-time learning(JITL)models,which fail to account for operational condition information during similarity measurement and are susceptible to noise,thereby compromising prediction accuracy,a mutual information-weighted similarity sample selection-based just-in-time ensemble learning(JITL-EL)prediction model is proposed.First,the spectral clustering algorithm is employed to partition historical data samples into several subsets,maximizing inter-subset differences and intra-subset similarities,thereby effectively distinguishing operational conditions.Second,a posterior probability calculation method weighted by intra-class features is introduced based on the correlation between intra-class features and target variables,enabling the determination of the membership degree of the test sample to different clustered subsets.Subsequently,based on the membership degree of the test sample to the clustered subsets,a dynamic selection of varying numbers of similar samples from different subsets is performed using an intra-class feature-weighted metric method,constructing JITL base model learners.Final,the predicted values from the JITL base models of different clustered subsets are weighted and fused according to the membership degree of the current test sample to each subset,yielding the final prediction results.Simulation results using real data collected from steel mills demonstrate that the accuracy of carbon content prediction reaches 80.00%,while that of temperature prediction achieves 83.00%.
作者
王浩东
刘辉
WANG Haodong;LIU Hui(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,Yunnan,China)
出处
《钢铁研究学报》
北大核心
2025年第6期761-773,共13页
Journal of Iron and Steel Research
基金
国家自然科学基金资助项目(62263016)
云南省应用基础研究基金资助项目(202401AT070375)
云南省“兴滇英才支持计划”资助项目
云南省高校服务重点产业科技资助项目(FWCY-QYCT2024003)。
关键词
互信息加权度量
即时-集成学习(JITL-EL)
特征加权
聚类算法
后验概率
mutual information-weighted metric
just-in-time ensemble learning(JITL-EL)
feature weighting
clustering algorithm
posterior probability
作者简介
王浩东(1999-),男,硕士生,E-mail:1135314593@qq.com;通信作者:刘辉(1984-),男,博士,教授,E-mail:liuhui621@126.com。
