The performance of rolling mill vibration prediction model depends on the features extracted from input variables. Therefore, aiming at the characteristics of large sample size and strong nonlinearity for the vibration data of tandem cold rolling mill, and the forward and backward dependencies in the time, a prediction method of rolling mill torsional vibration based on SSAE-LSTM was proposed. Firstly, for the rolling process parameters with small numerical differences and indistinct relationship representation of the same parameters, a stacked sparse autoencoder（SSAE）network was used for unsupervised adaptive feature extraction to mine the deep-level features of production data. Then, taking the advantage of long short-term memory （LSTM） network in dealing with time series, the deep-level features extracted by SSAE network were used as the input of the prediction model，and the rotational angular acceleration was used as the output to establish the rolling mill torsional vibration prediction model based on LSTM. The simulation results show that the prediction accuracy of SSAE-LSTM model is 98.5%. Compared with RNN model and LSTM model, the prediction accuracy of SSAE-LSTM model is improved by 24.8% and 12.2% respectively, and the validity of the method is verified, which provides the reference for the real-time prediction of the rolling mill torsional vibration state.