Abstract: With the improvement of computer operation level，data-driven technology is widely used in

metallurgical industry. Accurate prediction of rolling force is helpful to shorten the length of strip head and tail and

improve the yield of strip. In order to solve the problem of low prediction accuracy of data-driven model due to

blind feature extraction during pre-training，a deep hierarchical supervised preprocessing framework based on semisupervised stacked autoencoder（SS-SAE）was proposed for modeling rolling force prediction. In SS-SAE，several semi-supervised autoencoders（SS-AEs）were trained in turn to extract target related features. Each SS-AE took features from the previous hidden layer as new input to generate higher-order features. By stacking multiple SSAEs，the features related to deep targets could be gradually learned，while the deep network structure will gradually reduce the irrelevant information. The simulation results show that the prediction accuracy of this model can be controlled within 2%，and the high precision prediction of rolling force is realized.

      Key words: deep learning；feature extraction；autoencoder （AE）；rolling force prediction