Abstract: To predict the molten steel temperature in the smelting process of ladle furnace (LF), a hybrid model is proposed, where a BP neural network trained by an improved differential evolutionary algorithm (DEA) is introduced to determine the model parameters which are hard to exactly obtain by mechanism models. In the model, the mutation factor is randomly selected and the individuals in population are re-initialized in due time, thus rising above the difficulty efficiently in training BP neural network and avoiding the premature of DEA. Simulation results showed that the hybrid model has good precision of prediction for molten steel temperature and meets practical production requirements.