School of Mechanical and Power Engineering,Shenyang University of Chemical Technology
为了满足日益增长的带钢板凸度预测精度和速度要求,本文建立了一种主成分分析(principal component analysis,PCA)结合随机森林(random forest,RF)算法的板凸度预测模型。首先,应用pauta准则去除异常值,用五点三次平滑公式进行降噪处理；其次采用主成分分析法对数据进行降维,利用载荷矩阵选取关键控制变量；最后利用关键控制变量建立基于随机森林的板凸度预测模型,并与?持向量机回归(support vector regression,SVR) 、最近邻(K Nearest Neighbor,KNN) 、梯度提升决策树(Gradient Boosting Decision Tree,GBDT)、极端梯度增强(extreme gradient boosting,XGBoost)、轻量梯度提升机(light Gradient Boosting Machine,LightGBM)模型进?比较。结果表明,PCA-RF模型将参数由93维降低到15维,极大的减少了建模时间,且PCA-RF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0.982 0、1.485 2 μm和2.260 3 μm , 均优于其他预测模型,且98%以上样本点的预测误差在-3~3 μm,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,从而证明该模型能够通过降维减少建模时间的同时实现了带钢板凸度的高精度预测,为热轧带钢板凸度的研究提供了一定的参考。
To meet the increasing accuracy and speed requirements for strip crown prediction, a strip crown prediction model based on principal component analysis (PCA) and the random forest (RF) algorithm was established in this paper. Firstly, the pauta criterion was used to remove outliers, and the five-point cubic smoothing formula was employed for noise reduction. Secondly, the dimension of the data was reduced using principal component analysis, and the load matrix was utilized to select the key control variables. Finally, a strip crown prediction model based on random forest was established by utilizing the key control variables. The model was compared with the support vector regression (SVR), K-Nearest Neighbor (KNN), Gradient Boosting Decision Tree (GBDT), extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM) models. The results showed that the PCA-RF model reduced the parameters from LightGBM dimensions to 15 dimensions, significantly reducing the modeling time. The coefficient of determination (R2), mean absolute error (MAE), and root mean squared error (RMSE) of PCA-RF based on the test set were 0.982 0, 1.485 2 μm and 2.260 3 μm, respectively. The prediction error of over 98% of sample points was within the range of -3~3 μm, which meets the accuracy requirements for strip crown prediction. Additionally, to further verify the prediction stability of the model, the distribution of R2, MAE, and RMSE remained in the optimal position after 100 model operations. It is concluded that the PCA-RF model can reduce modeling time through dimensionality reduction while achieving high-precision strip crown prediction, providing a reference for further studies on strip crown.
赵志挺,朱亮宇,高珣洋,等. 基于PCA-RF的热轧带钢板凸度预测[J]. 科学技术与工程, , ():复制