Abstract: Based on the data in 2013-2014 at automatic weather station, radiosonde and the concentrations of air pollutants, the identi fi cation models of haze and the prediction models of visibility at 1400 BT were respectively carried out by using the Support Vector Machine (SVM) at Jinhua Meteorological Bureau. The results show that: 1) The identi fi cation models of haze may be used in the actual business forecast, because the satis fi ed TS scores were all over 0.65, except for that only three days were judged completely wrong by eight optimal models; 2) The prediction of the visibility forecast models for haze-day at 1400 BT in Jinhua were concentrated in 6-16km, which is much small than the actural range, however. That means the models have little cability to forecast the extreme values, so have a little guidance to distinguish the moderate and severe haze.