Considering the horizontal vibration, vertical vibration and dynamic rolling process parameters of work roll, a CNN-LSTM dynamic rolling force model based on a combination of one-dimensional convolutional neural network (1D CNN) and long short-term memory network (LSTM) was established. Then, aiming at the problems of unsteady state and variable working conditions such as multiple product specifications and frequent roll changes in the hot rolling process, the dynamic rolling force prediction of small sample target data was enhanced by the transfer learning method. The results show that the CNN-LSTM model has better generalization ability. With 10354 sets of data, the prediction accuracy of the CNN-LSTM dynamic rolling force model is 96.7%, and the relative error is within 0.3%. For variable working conditions, the transfer learning model needs 600 sets of data to achieve the prediction accuracy of more than 90%.      The time of data training and parameter adjustment is saved by the transfer learning method, the prediction accuracy of small sample target data is improved to better adapt to the actual production, which provides a new idea for the rapid prediction of dynamic rolling force and vibration suppression.