基于BO-LightGBM混合模型的热轧带钢板凸度预测

Crown Prediction of Hot Rolled Strip Based on BO-LightGBM Hybrid ModelZHAO Zhiting

为了满足日益增长的热轧带钢板凸度预测要求,建立了基于贝叶斯优化(Bayesian Optimization,BO)结合轻量梯度提升机(Light Gradient Boosting Machine,Light GBM)的板凸度预测模型BO-Light GBM。首先通过贝塞尔公式去除异常值,并经过五点三次平滑降噪。其次,建立Light GBM模型进行板凸度预测,最后通过贝叶斯优化算法选择最优参数,得到最优的模型,并与梯度提升决策树(Gradient Boosting Decision Tree,GBDT)、极端梯度增强(Extreme Gradient Boosting,XGBoost)和随机森林(Random Forest,RF)模型XGBoost、GBDT、RF算法进行比较。实验表明,基于BO-light GBM的板凸度预测模型优于XGBoost、GBDT、RF模型,对测试集预测的决定系数(Coefficient of Determination,R2)、平均绝对误差(Mean Absolute Error,MAE)和均方根误差(Root Mean Squared Error,RMSE)分别为0.97、1.49、2.28μm,均优于其他预测模型,且98%以上样本点的预测误差在-4~4μm,满足板凸度预测的精度要求。同时为了进一步验证该模型的预测稳定性,将模型运行100次后,R2、MAE和RMSE的分布仍处于最优位置,所以认为BO-LIght GBM是值得推荐的板凸度预测模型。

In order to meet the increasing requirements of crown prediction of hot rolled strip,this paper establishes a crown prediction model BO-LightGBM based on Bayesian optimization(BO)combined with light gradient boosting machine(LightGBM).Firstly,the outliers are removed by Bessel's formula and undergo five-spot triple smoothing.secondly,a LightGBM model is built for plate convexity prediction,and finally the optimal parameters are selected by Bayesian optimization algorithm to obtain the optimal model and compared with the gradient boosting decision tree(GBDT),extreme gradient boosting(XGBoost)and random forest(RF)model XGBoost,GBDT,RF algorithms are compared.The experiments show that the BO-lightGBM-based plate convexity prediction model outperforms the XGBoost,GBDT,RF model.The coefficient of determination(R2),mean absolute error(MAE)and root mean squared error(RMSE)of the test set prediction are 0.97,1.49,2.28μm,respectively.The prediction errors of more than 98%of the sample points are in the range of-4~4μm,which meets the accuracy requirement of plate convexity prediction.In order to further verify the prediction stability of the model,the distributions of R2,MAE and RMSE are still in the optimal position after running the model 100 times,so BO-LIghtGBM is considered as a recommended prediction model for plate convexity.