In this paper,a new nonlinear prediction model of chaotic time series which has both the global and local properties has been given and explained in detail. Firstly, the Kohonen Self Organization Network is used to produce K centers and to cluster the points of the reconstructed attractor using time delay coordinate into K classes; Secondly. a nonlinear model based on radial basis functions is fitted to every class. So the prediction model is composed of K functions which are all defined on the whole whole attractor and also have the good properties near their correspondent centers，such as the predicting ability can be improved by increasing size of the learning set．Compared with the local prediction model，this new model has an explicit form (only K functions)and avoids the long time searching for the local points as in the previous one．