In this paper, we propose a new neural network-GA neural network with high-dimensional multi-input layers and then a mechanical properties prediction model of hot-rolling production is established based on its structure flgure and training algorithm. As to the multi-stage characters of process in the hotrolling strip, input nodes are added to some hidden layers of neural network according to the conducting order of technological process. Furthermore, a two-stage training algorithm which combines genetic algorithm with BP algorithm is adopted to achieve higher prediction accuracy. Finally, testing results given by the actual manufacture data of hot-rolling products from an iron and steel enterprise show that the predicted result satisfles the requirement of actual prediction of mechanical property, and has higher accuracy and stability than classical BP and RBF neural network.