基于CNN-LSSVM的转炉炉后动态合金加入量预测模型  

A prediction model for dynamic alloy addition after converter based on CNN-LSSVM

作　　者：董晓雪 韩啸 杨鑫 何志军[1,2] 乔西亚 朱海琳 DONG Xiaoxue;HAN Xiao;YANG Xin;HE Zhijun;QIAO Xiya;ZHU Hailin(School of Materials and Metallurgy,University of Science and Technology Liaoning,Anshan 114051,Liaoning,China;Key Laboratory of Green Low-Carbon and Intelligent Metallurgy,Liaoning Province,Anshan 114051,Liaoning,China)

机构地区：[1]辽宁科技大学材料与冶金学院,辽宁鞍山114051 [2]辽宁省绿色低碳与智能冶金重点实验室,辽宁鞍山114051

出　　处：《钢铁》2025年第1期75-83,共9页Iron and Steel

基　　金：国家自然科学基金资助项目(52104331)。

摘　　要：转炉炉后脱氧合金化是转炉炼钢过程中非常重要的环节,获取精确的转炉炉后合金收得率及合金加入量,可降低生产成本、提高产品质量。以某钢厂120 t转炉冶炼HRB400E钢种的炉后操作为研究对象,通过RF(random forests,随机森林)结合递归特征消除法对影响硅锰合金收得率的因素进行回归分析,确定了9个转炉冶炼工艺参数作为后续模型的输入项。综合分析了BP、CNN、LSSVM算法的优缺点,分别建立了基于BP、CNN、CNNLSSVM的转炉炉后合金动态收得率预测模型,获取并保存三者预测精度最高的网络结构参数,对比发现CNNLSSVM模型的预测效果更为精确且贴合现场工艺特点,其决定系数为0.952,均方根误差为0.0068,平均绝对误差为0.0045。基于动态合金收得率模型,建立了转炉炉后操作合金加料预测模型,从成本最低角度确定转炉炉后合金配加方式,采用转炉炉后脱氧合金化和物料平衡原理,结合线性回归方法对原有合金加料方案进行优化。结果显示,优化后的预测合金加料成本均低于实际加料成本,并且成品钢中碳元素质量分数从原来的0.220%~0.255%收窄到0.230%~0.248%、硅元素质量分数从原来的0.38%~0.65%收窄到0.40%~0.54%、锰元素质量分数从原来的1.31%~1.64%收窄到1.35%~1.60%,符合钢种内控标准且实现了成分收窄的效果。该模型能够指导实际生产操作,提高企业的经济效益。Post converter deoxidation alloying is a very important process in converter steelmaking.Obtaining accurate alloy yield and alloy addition after converter furnace can reduce production costs and improve product quality.The post furnace operation of a 120 t converter smelting HRB400E steel was studied in a certain steel plant.RF(random forests)combined with recursive feature elimination method was used to perform regression analysis on the factors affecting the yield of silicon manganese alloy,and 9 converter smelting process parameters were determined as input items for the subsequent model.A comprehensive analysis was conducted on the advantages and disadvantages of BP,CNN,and LSSVM algorithms.Based on BP,CNN,and CNN-LSSVM,dynamic yield prediction models for converter rear alloys were established.The network structure parameters with the highest prediction accuracy among the three were obtained and saved.Comparison showed that the CNN-LSSVM model had a more accurate prediction effect and was in line with the characteristics of the on-site process,with a determination coefficient of 0.952,root mean square error of 0.0068,and average absolute error of 0.0045.Based on the dynamic alloy yield model,a prediction model for alloy addition after converter operation was established.The alloy addition method after converter operation was determined from the perspective of lowest cost.The deoxidation alloying and material balance principles after converter operation were adopted,and the original alloy addition scheme was optimized by combining linear regression method.The results show that the optimized predicted alloy addition cost is lower than the actual addition cost,and the content of C,Si,and Mn elements in the finished steel narrows from 0.220%-0.255%to 0.230%-0.248%,0.38%-0.65%to 0.40%-0.54%,and 1.31%-1.64%to 1.35%-1.60%,respectively,which meets the internal control standards of the steel grade and achieves the effect of narrowing the composition.This model can guide actual production operations and improve the eco

关 键 词：CNN-LSSVM模型 合金收得率 合金成本 特征提取 出钢合金化 随机森林 转炉 

分 类 号：TF713.5[冶金工程—钢铁冶金] TP18[自动化与计算机技术—控制理论与控制工程]