摘要
热连轧生产为多钢种、多规格混杂的带钢连续轧制过程,现有的机器学习方法不能考虑各带钢层的影响,将各带钢的轧制力预测过程视为独立的而不是关联的,这种做法不符合实际情况。提出一种预测带钢轧制力的梯度提升树-图卷积神经网络(gradient boosting decision tree-graph convolutional networks, GBDT-GCN)模型。首先,构建用于轧制力预测的带钢关系图结构,将数据集中的每块带钢作为图结构中的节点,根据带钢的轧制时序、层别关系生成各带钢节点之间的连接边,将连续轧制、相同层别的带钢关联起来;接着,将图结构输入结构调整后的GCN模型,采用平均绝对误差作为损失函数进行模型训练,采用GBDT对轧制力的影响因素进行重要性排序,并根据GCN模型的预测精度变化筛选出重要的因子作为最终的节点特征向量。最后,利用国内某热连轧机组的实际生产数据进行试验验证,结果表明,GBDT-GCN模型在测试集上的平均绝对误差为405.6 kN,相对误差在±10%以内的数据所占比例为91.5%,相较于传统SIMS模型、RF随机森林算法、MLP多层感知机模型,利用带钢关系图结构预测轧制力的GBDT-GCN模型具有更高的预报精度。
Hot strip rolling is a continuous rolling process of strip steel with multiple steel grades and specifications,however the current machine learning methods could not consider the influence of strip layers,and treat the rolling force prediction process as an independent process of each strip rather than a related process,which is not in line with the actual situation.A gradient boosting decision tree-graph convolutional networks(GBDT-GCN)model is proposed to predict the rolling force of strip steel.Firstly,a strip relational graph structure for rolling force predic-tion was constructed.Each strip in the data set was regarded as a node in the graph structure,and the connecting edges were generated according to the rolling sequence and layer relationship of the strip.The same layer strips or the strips which were continuously rolled are associated by graph edges.Then,the graph structure was input into the adjusted GCN,and the mean absolute error was used as the loss function for model training,and the GBDT was utilized to rank the influencing factors of rolling force,the important factors were selected as the final node feature vector according to the change of prediction accuracy of GCN model.Finally,the actual production data of a domes-tic hot strip mill are used for experimental verification.The results show that the average absolute error of the GB-DT-GCN model on the test set is 405.6kN,and the proportion of data with relative error within±10%is 91.5%.Compared with the traditional SIMS model,RF random forest algorithm and MLP multi-layer perceptron model,the GBDT-GCN model of predicting rolling force by strip graph structure has higher prediction accuracy.
作者
李维刚
刘玮汲
谢璐
赵云涛
LI Wei-gang;LIU Wei-ji;XIE Lu;ZHAO Yun-tao(Engineering Research Center for Metallurgical Automation and Measurement Technology of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China;School of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China)
出处
《钢铁》
CAS
CSCD
北大核心
2023年第3期89-96,127,共9页
Iron and Steel
基金
国家自然科学基金资助项目(51774219)。
关键词
热连轧带钢
轧制力预测
特征选择
梯度提升树
图卷积神经网络
hot strip rolling
rolling force prediction
feature selection
gradient boosting decision tree
graph conv-olutional networks
