基于代价敏感和贝叶斯优化XGBoost的KR工序脱硫符合预测  

Prediction of desulfurization conformity in KR process based on cost-sensitive and Bayesian optimized XGBoost

作　　者：戴兆汉 于艳[2] 张宇军[2] 何飞 DAI Zhaohan;YU Yan;ZHANG Yujun;HE Fei(School of Metallurgical Engineering,Anhui University of Technology,Ma'anshan 243002,China;Central Research Institute,Baoshan Iron&Steel Co.,Ltd.,Shanghai 201900,China)

机构地区：[1]安徽工业大学冶金工程学院,安徽马鞍山243002 [2]宝山钢铁股份有限公司中央研究院,上海201900

出　　处：《冶金自动化》2024年第5期44-52,共9页Metallurgical Industry Automation

摘　　要：KR脱硫是铁水预处理的典型方法之一。随着低硫钢种产品需求的增加,实现铁水终点硫含量的稳定控制,进而降低脱硫综合处理成本非常重要,而其关键环节是如何预测脱硫终点是否符合目标问题。因此,针对KR工序铁水终点硫预测问题,本文提出一种基于代价敏感策略的样本类别平衡处理方法和贝叶斯优化的极限梯度提升(extreme gradient boosting,XGBoost)算法以及脱硫终点硫含量二分类分析方法相结合的建模方法。首先,通过基于二分类分析方法对脱硫数据的终点硫含量进行符合或不符合二分类处理,并基于代价敏感策略调整类别样本权重缓解不平衡问题,构建特征数据集;然后,利用某钢铁企业实际生产数据,基于代价敏感策略和贝叶斯优化XGBoost交叉验证训练模型,同时通过Macro-F1指标优选最优参数形成最终的KR工序脱硫符合预测模型,实现了对脱硫目标符合和不符合的数据预测。与支持向量机(support vector machine,SVM)、反向传播神经网络(back propagation neural network,BPNN)预测模型的对比实验结果表明,本文方法能够有效处理脱硫数据不平衡问题,具有较好的脱硫符合预测实践效果。KR desulfurization is a typical approach for hot metal pretreatment.With the increasing demand of low sulfur steel products,achieving stable control of end sulfur content in molten iron and subsequently reducing the overall cost of desulfurization processes are of paramount importance.A critical aspect of this process is the ability to predict whether the desulfurization endpoint complies with the required standards.Therefore,a modeling method combining a sample class balance processing method based on cost-sensitive strategy and Bayesian-optimized extreme gradient boosting(XG-Boost)algorithm with a binary classification analysis method for solving end sulfur prediction problem in KR desulfurization process is proposed.Firstly,the end sulfur content of the desulfurization data is processed into conforming or non-conforming two categories based on the binary classification analysis method.The category sample weights are adjusted based on the cost-sensitive strategy to alleviate the imbalance issue for constructing the feature dataset.Then,using actual production data from a steel company,the model is trained via cross-validation with cost-sensitive strategy and Bayesian-optimized XGBoost with the optimal parameters are selected based on Macro-F1 metric to form the final desulfurization conformity prediction model for the KR process,achieving the data prediction for desulfurization conformity and non-conformity targets.Experimental results comparing with support vector machine(SVM)and back propagation neural network(BPNN)prediction models show that the proposed method can effectively deal with the imbalance issue in desulfurization data,showing good practical effects in desulfurization conformity prediction.

关 键 词：KR 脱硫符合 预测 二分类 代价敏感 贝叶斯优化 XGBoost 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TF704.3[自动化与计算机技术—控制科学与工程]