Aiming at the problem of traditional iterative way to achieve method of multi-step prediction, new method of multi-step prediction based on parallel Support Vector Regression (SVR) was proposed. To begin with the time delay of time series will be calculated in this method, resample the time series according to the interval of time delay. What’s more the time series will be classified into several sets of data, and it sets up SVR model for the sets of each. Finally, the parallel prediction of each set is composed to get multi-step prediction result. This method not only eliminates the accumulated error, improves the accuracy of prediction, but also saves the computation time.