An evolutionary structure optimisation method for the Gaussian radial basis function network is presented for modelling and predicting nonlinear time series. The generalisation performance is significantly improved with a much smaller network, compared with that of the previous clustering and least square learning method.