Abstract: In order to improve the predicted precision of rolling force in unsteady process， an optimization method for rolling force self-learning model was proposed. The self-learning coefficient in the model was decomposed into the layer learning coefficient and rolling state learning coefficient， which characterized the genetic characteristics of rolling force prediction deviation between racks， and the effect of actual roll state on the model prediction. In the process of coefficient updating， the learning coefficients were updated according to the layer distance so that the prediction error of rolling force could be reduced， especially when the rolling specifications were switched. The proposed self-learning method has been successfully applied into a hot rolling process. Compared with the original model， the predicted deviation of the optimized self-learning method is reduced from 2.8% to 1.4%， and the mean square error is reduced from 3.3% to 1.7%， which effectively improve the accuracy and robustness of rolling force in unsteady process.

Key words: hot rolling, rolling force, unsteady process, layer distance, prediction deviation