基于TCN-DenseNet的烧结矿FeO含量预测  

PREDICTION OF FeO CONTENT IN SINTER BASED ON TCN-DENSENET

作　　者：黄鼎堯 黄晓贤[1,2] 向家发 彭梓塘[1] 周茂军 陈许玲[1,2] 冯振湘[1,2] 范晓慧 Huang Dingyao;Huang Xiaoxian;Xiang Jiafa;Peng Zitang;Zhou Maojun;Chen Xuling;Feng Zhenxiang;Fan Xiaohui(School of Mineral Processing and Bioengineering,Central South University,Changsha 410083,Hunan;Engineering Research Center of Ministry of Education for Carbon Emission Reduction in Metal Resource Exploitation and Utilization,Central South University,Changsha 410083,Hunan;Ironmaking Plant of Baoshan Iron&Steel Co.,Ltd.,Shanghai 200941)

机构地区：[1]中南大学资源加工与生物工程学院,湖南长沙410083 [2]中南大学金属资源开发利用碳减排教育部工程研究中心,湖南长沙410083 [3]宝山钢铁股份有限公司炼铁厂,上海200941

出　　处：《河北冶金》2024年第10期14-19,49,共7页Hebei Metallurgy

基　　金：国家自然科学基金基础科学中心项目(72088101);中南大学研究生自主探索创新项目(2024ZZTS0381)。

摘　　要：烧结矿FeO含量是烧结工序的一项重要质量和能耗指标,也对高炉冶炼有直接影响。针对目前化学检测法检测烧结矿FeO含量时存在较长时间滞后的现状,本文提出了一种时域卷积网络(Temporal Convolutional Network,TCN)与密集连接卷积神经网络(Densely Connected Convolutional Network,DenseNet)混合的烧结矿FeO含量预测方法。首先采用TCN建立烧结矿FeO含量的时间序列预测模型,同时采集烧结机尾断面红外图像,采用DenseNet建立烧结矿FeO预测模型,通过自适应加权平均方法将两者的输出结果进行整合,获得最终的烧结矿FeO含量预测值。针对烧结矿层断面红外图像的特征,对DenseNet进行了添加注意力层、修改卷积块结构,并修改了浅层卷积层大小和步长等改进措施。在国内某钢铁公司的大型烧结机的实际生产数据上对模型进行了验证,经过数据处理、模型参数优化等操作后,本文所提的TCN-DenseNet混合模型的烧结矿FeO含量预测在测试集绝对误差±0.4%以内命中率可达94.34%,均方根误差为0.21,优于单独使用TCN或者DenseNet进行建模时的预测效果。该方法对提高烧结矿FeO含量预测的准确性和稳定性效果显著,可以为烧结现场的生产操作提供数据支撑。The FeO content in sinter is an important quality and energy consumption indicator of the sintering process,and also has a direct impact on blast furnace smelting.In response to the current situation of long time lag in chemical detection methods for detecting FeO content in sinter,this paper proposes a method for predicting FeO content in sinter by mixing Temporal Convolutional Network(TCN)and Dense Connected Convolutional Network(DenseNet).Firstly,TCN is used to establish a time series prediction model for FeO content in sinter,and infrared images of the tail section of the sintering machine are collected.DenseNet is used to establish a FeO prediction model for sinter,and the output results of the two are integrated through adaptive weighted averaging method to obtain the final FeO content prediction value.Based on the characteristics of infrared images of sinter layers,DenseNet was improved by adding attention layers,modifying the convolutional block structure,and adjusting the size and step size of shallow convolutional layers.The model was validated on the actual production data of a large sintering machine in a domestic steel company.After data processing,model parameter optimization,and other operations,the TCN DenseNet hybrid model proposed in this paper achieved a hit rate of 94.34%and a root mean square error of 0.21 for predicting the FeO content of sinter within an absolute error of±0.4%in the test set,which is better than the prediction effect of using TCN or DenseNet alone for modeling.This method has a significant effect on improving the accuracy and stability of FeO content prediction in sinter,and can provide data support for production operations in sintering sites.

关 键 词：烧结 FEO含量 复合预测模型 TCN DenseNet 注意力机制 

分 类 号：TF046.4[冶金工程—冶金物理化学]