基于类别特征梯度提升的冷轧带钢板形预测模型  

Prediction model of cold rolled strip flatness based on CatBoost

作　　者：丁肇印 丁成砚 孙杰[1] 张殿华[1] DING Zhaoyin;DING Chengyan;SUN Jie;ZHANG Dianhua(State Key Laboratory of Rolling and Automation,Northeastern University,Shenyang 110819,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

出　　处：《轧钢》2022年第6期99-105,共7页Steel Rolling

基　　金：国家自然科学基金重点项目(U21A20117,51634002);国家重点研发计划项目(2018YFB1308700);辽宁省兴辽英才计划项目(XLYC1907065)。

摘　　要：板形控制是冷轧带钢生产过程的核心技术。为了提升板形预设定和闭环反馈控制效果,建立高精度的板形预测模型非常必要。提出了一种基于类别特征梯度提升的冷轧带钢板形预测模型,通过某1 450 mm冷连轧生产线采集的生产数据建立模型,采用贪婪搜索和交叉验证的方式进行超参数设置,以自适应提升模型、梯度提升决策树模型和深度学习神经网络模型作为对比。结果表明:类别特征梯度提升模型的RMSE为0.666 IU,并且有90.397%的预测数据绝对误差小于1 IU,较其他3种模型有更好的表现,对冷轧带钢板形预测有更好的鲁棒性和预测精度。Flatness control is the core technology of cold rolled strip production process. To improve the effect of flatness preset and closed-loop feedback control, constructing a high-precision flatness prediction model is necessary. A CatBoost-based predictive model for cold rolled strip flatness was proposed. Production data collected from a 1 450 mm tandem cold rolling production line were used to establish the model, and greedy search and cross validation were utilized to tune hyper-parameters. Moreover, Adaboost, GBDT and DNN were selected as the comparison models. The experimental results show that RMSE of CatBoost model is 0.666 IU and 90.397% of prediction data has an absolute error less than 1 IU, which prove the proposed CatBoost model outperform than three comparison models and has better robustness and accuracy for cold rolled strip flatness prediction.

关 键 词：冷轧带钢 板形预测 类别特征梯度提升模型 集成学习 

分 类 号：TG335.12[金属学及工艺—金属压力加工]