Multistep networks for roll force prediction in hot strip rolling mill
Shen Shu-hong, Denzel Guye, Xiaoping Ma, Stephen Yue, Narges Armanfard
Abstract
Hot rolling processes consist of multiple single rolling stand operating at high temperature and speed to achieve desired steel shapes and superior properties, via exerting roll forces that need to be accurately predicted by a model. The currently used model of the mill of this study shows prediction instability and is unable to accurately accommodate changes in steel grade. In this paper, we propose a machine learning based framework to establish a model that accurately predicts roll forces at each mill stands of the hot strip rolling mill. In contrast to the traditional models, the proposed expert system considers an individual model for each rolling stand and employs rolling history when predicting roll forces. The proposed model includes both steel chemistry and physical process parameters for its predictions. Our experimental results demonstrate that the proposed framework improves both prediction accuracy and stability by 40%–50% over the currently used mill model. The enhanced prediction accuracy will greatly improve dimensional and microstructural control, as well as ensuring the avoidance of mill overloads.