基于改进CNN-LSTM和RF的铁水KR脱硫预测模型  

Model for predicting KR desulfurization of hot metal based on improved CNN-LSTM and random forest

作　　者：胡佳辉 熊凌[1] 但斌斌 吴经纬[3] HU Jiahui;XIONG Ling;DAN Binbin;WU Jingwei(College of Information Science and Engineering,Wuhan University of Science and Technology,Wuhan 430081,China;Engineering Research Center for Metallurgical Automation and Detecting Technology of Ministry of Education,Wuhan University of Science and Technology,Wuhan 430081,China;Intelligent Manufacturing Division,WISDRI Engineering and Research Incorporation Limited,Wuhan 430223,China)

机构地区：[1]武汉科技大学信息科学与工程学院,湖北武汉430081 [2]武汉科技大学冶金自动化与检测技术教育部工程研究中心,湖北武汉430081 [3]中冶南方工程技术有限公司智能制造事业部,湖北武汉430223

出　　处：《武汉科技大学学报》2024年第4期254-263,共10页Journal of Wuhan University of Science and Technology

基　　金：国家自然科学基金项目(62173261);湖北省重点研发计划项目(2020BAB021).

摘　　要：为实现较高精度的脱硫剂加入量预测,有效提高生产效益,本文提出一种基于改进卷积神经网络(CNN)-长短期记忆(LSTM)网络和随机森林(RF)结合的铁水脱硫两步预测模型。考虑到模型输入数据的相关性,利用皮尔逊相关系数确定各输入参数的相关性并筛选特征。模型以CNN-LSTM为基础,增加卷积层和残差连接,在提高挖掘数据的高维特征信息的同时避免网络退化。为增加网络对特征的区分和关注能力,引入多头注意力机制,让网络更加关注特征中的重要信息。使用贝叶斯优化RF超参数构建误差预测模型从而实现残差推理,对改进的CNN-LSTM模型预测结果进行修正。以现场采集的数据进行实验,结果表明,与CNN-LSTM模型相比,本文模型的拟合精度R2提升了17.11%,平均绝对值误差MAE降低了24.85%,均方根误差RMSE降低了30.18%,平均绝对百分比误差MAPE降低了28.33%。In order to achieve higher precision prediction of desulfurization agent addition and effectively improve production efficiency,a two-step model for predicting hot metal desulfurization was proposed based on the combination of improved CNN-LSTM and random forest.With the correlation of the input data of the model considered,the Pearson correlation coefficient was used to determine the correlation of each input parameter and filter the features.The model was based on CNN-LSTM,adding convolutional layers and residual connections to improve the high-dimensionality of the mining data feature information while avoiding network degradation.In order to increase the network’s ability to distinguish and focus on features,a multi-head attention mechanism was introduced to enable the network to pay more attention to important information in features.The Bayesian optimization of random forest hyperparameters was used to construct an error prediction model to achieve residual reasoning and correct the prediction results of the improved CNN-LSTM model.The experimental results show that compared with those of the CNN-LSTM model,the R-square of the proposed model increases by 17.11%,the average absolute value error(MAE)reduces by 24.85%,the root mean square error(RMSE)reduces by 30.18%,and the average absolute percentage error(MAPE)reduces by 28.33%.

关 键 词：KR脱硫模型 卷积神经网络 长短期记忆网络 注意力机制 随机森林 

分 类 号：TF704.3[冶金工程—钢铁冶金] TP183[自动化与计算机技术—控制理论与控制工程]