In order to establish an accurate and efficient air quality forecasting system, three forecasting models based on pollutants, meteorological factors and pollutant mixed meteorological factors were established and used as input variables for support vector machine regression (SVR) for the daily forecast to look for the best predictor mode. The support vector machine regression model was optimized by using grey wolf optimization (GWO) to form the GWO-SVR model. The experimental results show that the meteorological factors of pollutants mixing acted as input variables is the optimal predictor model, and the determination coefficients of the test set of SVR and GWO-SVR model are R2=0.79 and R2=0.81, respectively, which indicates both of the models have high prediction accuracy. By comparision, GWO-SVR model prediction performance is better. After that, the data is classified according to the wind direction conditions and the better GWO-SVR is used to predict the PM10 concentration. The prediction results show that when the prevailing southwest wind prevails, the evaluation index of prediction set is R=0.91 and MSE=47.15, which is better than the status with prevailing northeasterly wind where R=0.87 and MSE=125.80 and the whole data with R=0.90 and MSE=107.94.