To overcome the local convergence shortcoming of the gradient method used by the BP approach in the Elmannetwork training process, a combined training method of the exhaustion particle swarm optimization(EPSO) and gradient method is built. An enhanced PSO algorithm(EPSO), which has better global property, is proposed by simulating feeding behavior of birds. Then, the EPSO algorithm is applied to optimizing Elman network weights across the whole solution space, and the gradient method is used in the exhaustion process of EPSO to improve the local convergence property. Finally, the proposed training method is applied to flight trajectory prediction experiment, and simulation results show the effectiveness of this method.