Abstract: A new multi-branch time delay neural network is adopted to conduct prediction research on chaotic time series.In the cases that the initial network states are not strictly equal to those of the practical systems,the approximation ability of this network to nonlinear system is discussed.The structure of the network is defined by integrating the theory of reconstructing phase space,which makes the efficient prediction information be contained in the network.The chaotic time series generated by Rossler chaotic equation and the practically observed yearly sunspot time series are respectively taken as examples.Simulations show that the presented network can be used to model and predict the chaos system successfully,and the method can get a higher precision.