The present prediction methods for SO₂ concentration suffer from the disadvantages that there is no uniform understanding of pollutant sources and influencing factors, small sample data is sensitive, and prediction methods are easy to fall into local optimum etc. In order to solve these problems, a method for the prediction of SO₂ concentrations on expressway is proposed which is based on fuzzy time series and support vector machine (SVM), and provides a reliable theoretical support for building the highway environmental health monitoring system. Based on the seasonal variation of SO₂ concentrations, the method takes the season as time series, 24h for graining window width. Through the Gaussian kernel function to extract the eigenvalues of the original sample data, which are input support vector machine (SVM) model for training, and k-fold cross validation method combined with the grid division is used to optimize model parameters. Finally, a SO₂ concentrations prediction model is established with the method in this paper. By using 1h average SO₂ concentrations as sample data which are obtained by Shanxi taijiu expressway monitoring station from April 2014 to March 2014, the LIBSVM tool is used to realize the calculation process on the MATLAB platform. The results show that based on fuzzy time series and support vector machine (SVM), the forecasting methods of SO₂ concentration is not restricted by the research of machine rational theory, and supports small-sample learning, otherwise, the nonlinear fitting effect is perfect, and the ability of generalization is well.