基于WOA-ELM-LSTM的非稳态热轧过程轧制力预测  

Prediction of hot rolling force based on WOA-ELM-LSTM in unsteady process

作　　者：丁敬国[1] 刘方路 于琨 李旭[1] 张殿华[1] DING Jingguo;LIU Fanglu;YU Kun;LI Xu;ZHANG Dianhua(State Key Laboratory of Rolling and Automation,Northeastern University,Shenyang 110819,Liaoning,China;School of Materials Science and Engineering,Liaoning University of Technology,Jinzhou 121001,Liaoning,China)

机构地区：[1]东北大学轧制技术及连轧自动化国家重点实验室,辽宁沈阳110819 [2]辽宁工业大学材料科学与工程学院,辽宁锦州121001

出　　处：《钢铁研究学报》2024年第1期85-94,共10页Journal of Iron and Steel Research

基　　金：国家重点研发计划资助项目(2022YFB3304800);国家自然科学基金资助项目(U21A20475)。

摘　　要：热连轧生产过程中,因换辊、换钢种、换规格等非稳态轧制条件下,轧制力的预测精度降低,导致产品厚度命中率降低、秒流量控制失衡、宽度拉窄等质量问题,究其原因发现,机制模型在非稳态条件下的模型误差存在较大差异,仅通过层别表模型参数切换无法实现精准设定。为解决该问题,首先,构建具有计算速度快和预测精度良好的极限学习机作为浅层神经网络,同时构建具有挖掘工业数据特征能力强的短时记忆网络作为深度神经网络。其次,采用鲸鱼算法对极限学习机参数寻优,构建了基于鲸鱼算法优化极限学习机协同长短时记忆网络(whale algorithm to optimize extreme learning machine cooperative long-term and short-term memory network,WOA-ELM-LSTM)的热轧轧制力预测模型,然后增加误差值评判机制,利用长短时记忆网络对轧制力偏差值进行训练并结合极限学习机模型轧制力预测值进行二次修正,将该混合模型与支持向量机、经鲸鱼算法优化后的支持向量机(WOA-SVR)、极限学习机、经鲸鱼算法优化后的极限学习机(WOA-ELM)进行模型预测精度对比。对比结果表明,基于WOA-ELM-LSTM的热轧轧制力模型预测精度明显高于其他方法,该模型的R2值为99.34,轧制力预测偏差在±5%以内,在板带材热连轧生产中有着良好的应用前景。In the hot strip rolling process,the prediction accuracy of rolling force decreases due to the unsteady rolling conditions such as roll change,steel type change and specification change,which leads to the quality problems such as the reduction of thickness hit ratio,the imbalance of metal flow control and the narrowing of width.It is found that the model error of the mechanism model is quite different under the unsteady rolling condition,and the precise setting cannot be achieved only through the model parameters switching of the layer table.To solve this problem,firstly,extreme learning machine model with fast computing speed and good prediction accuracy was constructed as the shallow neural network,and the short-term memory network with strong mining ability of industrial data features was constructed as the deep neural network.Secondly,whale algorithm was used to optimize the parameters of extreme learning machine model,and a hot rolling force prediction model based on whale algorithm was constructed to optimize the extreme learning machine collaborative long and short time memory network(WOA-ELM-LSTM).Then,an error value evaluation mechanism was added.The deviation value of rolling force was trained by using long-short memory network,and combined with the predicted value of rolling force of extreme learning machine model,the hybrid model was compared with support vector machine,support vector machine optimized by whale algorithm(WOA-SVR),extreme learning machine and extreme learning machine optimized by whale algorithm(WOA-ELM).The comparison results show that the prediction accuracy of hot rolling force model based on WOA-ELM-LSTM is obviously higher than other methods.The R2 value of this model is 99.34,and the prediction deviation of rolling force is within±5%,which has a good application prospect in hot strip rolling production.

关 键 词：轧制力预测 热连轧 极限学习机 鲸鱼优化算法 非稳态过程 

分 类 号：TG335.11[金属学及工艺—金属压力加工]