Considering the impact of temporal-spatial features of traffic flow can improve the prediction accuracy. Therefore, this paper introduces a radial kernel function to convert the complex predictive problem into a regression algorithm in high-dimensional space. Then, based on support vector regression, an online short-term traffic flow prediction model considering temporal-spatial features is built. Grid search method is used to optimize the parameters. Finally, state vector is built to analyze the influence of temporal-spatial features. Based on the dataset of detectors in highway, different models are compared and the validity and feasibility of the prediction model are verified. The results indicate that online model is superior to traditional support vector. If considering the influence of temporal-spatial features the prediction model is more accuracy and steady.