Abstract: The conventional back-propagation for neural network is improved by introducing the variable-step learning rate with a momentum term added in so as to prevent the network from error surge and accelerate its convergence rate. Then, the causes and influencing factors on the longitudinal cracks on slab surface in the continuous casting process are analyzed, and a prediction system of longitudinal surface cracks of slab is set up with the HP295 steel supplied by Meishan Steelworks as example, based on the improved BP neural network. It is found that the root cause of the longitudinal surface cracks on HP295 is the nonuniform distribution of the secondary cooling water. So, adjusting the proportion of the secondary cooling water is the efficient way to reduce the formation of the longitudinal surface cracks in practical production.