In this paper, a dynamic model of a walking beam billet reheating furnace is constructed. The model is based on a multilayer perception neural network, which is trained using a sequential window batch learning algorithm. To avoid the lack of BP algorithm such as initial condition sensitivity and solving complex partial differential equations, a hybrid pattern search (PS) and particle swarm optimization (PSO) algorithm is introduced. Considering the different relations between data, a modified performance function is employed to improve the model training. Verification results show that the model has a favorable adaptation to dynamics of furnace, and capability of predicting furnace temperatures precisely.