In this paper, the forecasting method for the furnace bottom temperature and carbon content in smelting process is proposed to aim at the shortcomings of great energy and low productivity from control strategy, through the analysis of submerged arc furnace smelting process. The method provides a theoretical basis for dynamical control of submerged arc furnace smelting process. The forecasting model is established base on improved BP neural network, and 10 major factors affect the furnace bottom temperature and carbon content are selected to be as input variables. Variable step-size learning algorithm is used to achieve the purpose of global optimization and fast convergence. Data information from the regular and consecutive smelting process of the 801 furnace in 8th branch of Sinosteel Jilin Ferroalloy SCNB are selected to be as training samples and test samples, MATLAB simulation software is used to verify the accuracy and usefulness of the model.