基于机理与数据驱动的热连轧板凸度组合预测

Combined prediction of hot strip crowns of hot tandem rolling based on mechanism and data driving

针对传统热连轧出口板凸度预测方法存在的模型精度低、解释性差等缺陷,提出了一种将机理与数据驱动相结合的热连轧板凸度组合预测模型。通过热连轧板凸度机理预测模型得到热连轧板凸度基准值,将该基准值与实际值之间的偏差量作为机器学习模型的预测变量,再将偏差量预测值与基准值进行求和得出组合预测模型的板凸度预测值,并将该组合预测策略应用至多个神经网络进行方法验证。研究结果表明,提出的热连轧板凸度组合预测模型相较于传统预测模型具有更好的预测性能,其中有97%以上预测数据的绝对误差小于0.02 mm,82%以上预测数据的绝对误差小于0.01 mm,同时该组合预测方法具有较好的可行性与普适性,所提出的模型能够实现机理模型与数据驱动模型的优势互补,使得模型更加符合实际物理意义,该组合模型既缓解了神经网络预测结果由于过程黑箱导致解释性差、可信度低的问题,又弥补了机理模型预测结果偏离生产工况、无法实时修正的缺陷,对热连轧板带钢的板形控制以及热连轧产品质量的改善具有重要意义。

To address defects of the traditional method for predicting the strip outlet crown of hot tandem rolling,which suffers from low accuracy and poor interpretability,a model for combined prediction of hot strip crowns based on mechanism and data driving is proposed.The strip crown reference value is obtained using the strip crown mechanism prediction model.The deviation between the reference value and the actual value is used as the prediction variable of machine learning models,and then the deviation prediction value and the reference value are summed to obtain the strip crown prediction value of the combined prediction model.This combined prediction strategy is verified using multiple neural networks.It is found that the proposed strip crown combined prediction model has better prediction performance than the traditional model,with over 97%of the predicted data having an absolute error of less than 0.02 mm and more than 82%of the predicted data showing an absolute error of less than 0.01 mm.Additionally,the model is both satisfactorily feasible and widely applicable.The proposed model integrates the relative strengths of the mechanism model and the data-driven model,resulting in a representation that is more closely aligned with the actual physical phenomena.The combined model not only alleviates the problems of poor interpretation and low reliability with the results from the black-box neural network prediction,but also compensates for the defects of the mechanism model,which often produces results that deviate from the production conditions and cannot be adjusted in real time.This proposed model makes a significant contribution to the shape control of hot strip and the improvement of hot strip product quality.