It is very important for the prediction of temper mills rolling force to control the rolling process optimally. For the problem of low precision in rolling force prediction, a neural network model with ReLU (Rectified Linear Units) activation function was proposed to predict the rolling force of temper mill. After the principal components of data were analyzed, the main factors affecting the rolling force were obtained, which were selected as the input layer of neural network, and the rolling force of temper mill was taken as the output layer. The experiment was carried out by the python programming language, and related parameters and algorithms in the hidden layer of neural network model were chosen with the principle of single variable. Then a neural network forecasting model was established to ensure the highest accuracy of rolling force prediction. The results show that by adjusting the number of hidden layers, the number of neurons, the propagation algorithms and the regularization methods, the model is able to ensure the prediction error less than 10%. And the experimental method can accurately predict the rolling force of the temper mill with different input parameters.