Aiming at the problem of reduced prediction accuracy of rolling force prediction model under small number of samples when predicting the rolling force of different frames,which is caused by the complex rolling condition of each frame for cold rolling mill to make the distribution of actual collected process data different, a rolling force prediction model based on transfer learning and multi-aspect feature extraction was proposed. This method could transfer the parameters of established rolling force prediction benchmark model to the target frame by transfer learning to realize the rapid prediction of rolling force in different frames with small number of samples. When predicting, Inception-LSTM-Attention rolling force prediction benchmark model was established by using the source domain data, which combined the spatial feature extraction ability of Inception network module and the temporal feature prediction ability of LSTM. At the same time, the attention mechanism Attention was introduced to adjust the weight of information vector learned by the neural network, and the prediction accuracy of the benchmark model was as high as 98.8%. Then, some network modules of the benchmark model were frozen and transfered to the target domain model, and the model parameters were fine-tuned to establish the final rolling force prediction model. The experimental results show that the prediction accuracy of the model established based on the transfer learning method under small number of samples is higher than that of the traditional neural network prediction model.