The hyper-parameters of support vector regression influence the performance of its model. In the normal gradient descent method, kernel functions or estimation functions must approximately differential. And this method sensitively depends on initial value. Therefore, a novel two-stage optimization selection method for hyper-parameters is proposed. In first stage, the search extent of each hyper-parameter is determined according to the reqirement of issues. In second stage, optimal hyper-parameters are obtained by adaptive chaotic culture algorithm during above search space. Adaptive chaotic cultural algorithm uses implicit knownledge extracted from evolution process to control mutation scale of adaptive chaotic mutaion operator. This strategy can ensure the diversity of population and exploitation in the latter evolution. Taken prediction of Mackey-Glass time series as examples, simulation results indicate that the selection method can effectively avoid the premature convergence and has better computation stability and precision. And it is not related on the stucture of functions. SVR model corresponding to optimal hyper-parameters by this method has better generaliztion.