Starting from the support vector domain model, the paper establishes SVD predictive models of chaos time series as well as chaos phase trace of non-linear map, based on Takens phase space delay reconstructing theory. We adopted the methodof data set as support object elements. Machine self-learning reduces error upper limit of the generalized model. The three chaos time series, Henon/Lorenz/Rossler are predicted by least square. The prediction result indicates that the predictive model makes the set to be mapped into an eigen space of higher dimensions, and the series is predicted by embed dimensions. The predictive error changes with the increase of embed dimension to a constant. Compared with SVM, the SVD requires smaller support vector, and has faster convergence rate. It has robustness characteristics with adaptive flexible kernel function choice. The predicted net points are ten to twenty times more than SVM. Under the conditions of small sample, non-linear, and unknown probability density, the predicted series is in concordance with the series′ true value.