A prediction model for time series was introduced, which uses the hybrid algorithm to optimize the neural networks. The main idea was to build the new hybrid algorithm by combining particle swarm optimization with Simulated Annealing in sudden jump, and then to optimize the neural networks with this hybrid algorithm. Therefore, many shortcomings, like the slow convergence of common neural networks, partial optimization and the prediction precocity of simplex particle swarm optimization, were overcome. In addition, in order to prove the validity and the value of the model, the Mackey-Glass chaotic time series and the Henon map were simulated. The results have shown the fast convergence, good stability and the high precision of this model.