Deviation Prediction for Online Calculation Model of Hot-Rolled Strip Crown

Precise online crown calculations are essential for accurate control of hot-rolled strip shape, whose accuracy is limited by the assumptions made in mechanism model development and the variability of rolling production conditions. The current model self-learning strategy combines exponential smoothing-based short-term compensation and proportional archiving-based long-term compensation. However, increasingly stringent requirements for the shape quality of hot-rolled strip reveal deficiencies in the strategy, particularly in learning coefficient determination and cooperative performance. To address this, this article proposes a machine learning-based model to predict deviations between calculated and measured crown values, with the objective of replacing the current self-learning strategy to provide compensation for the model. Industrial data from hot-rolling production line are collected to construct a modeling dataset, where key process parameters impacting online crown calculations are selected as input features. Four machine learning methods—multilayer perceptron, regression tree, support vector regression, and locally weighted linear regression (LWLR)—are utilized to develop a predictive model. Results show that the LWLR model achieves the best performance, with a root mean square error of 5.77 and a coefficient of determination of 0.931, demonstrating its potential of model compensation and accuracy enhancement.