In order to overcome the problems of slow speed and low accuracy of convergence and the shortcomings of poor stability of the traditional logical prediction of breakout system. This paper designs a breakout predicting model based on BP neural network which is capable of self-organize and self-learn, and improving the stability and accuracy in breakout prediction. In this paper, we modified the BP algorithm to improve its learning speed such as changing study rate, adding momentum item and avoiding vibration item, so the network can escape from the local minimum while it is training. The drawing speed and temperature of molten steel in tundish are regarded as the influencing factors of breakout in model to extend the range of breakout factors. The experimental results show that the system predicted to get exact results based on practical data from field in a steel plant, so it has good anticipant practical application on line in predicting breakout.