In recent years

severe haze pollution accidents occurred frequently

which caused heavy losses in national economy and health of residents in China. Accurate early warning of severe haze pollution episodes can not only remind people of refraining from the hazards

but also set aside enough time for the government of emergency management before substantial improvement of air quality. According to the non-Gaussian distribution characteristics of PM2.5 precursors and meteorological factors and limitation of the number of known hidden states in traditional hidden Markov models (HMMs)

generalized hidden Markov models (GHMMs) were employed to make predictions of PM2.5 concentrations at 11 nationally controlled monitoring sites (except Dingling site) in Beijing from January 2013 to January 2017. Data from Jan. 2013 to Dec. 2015 was used to train the GHMM models and data from Jan. 2016 to Jan. 2017 was used to validate these models. The same data was also used to train traditional HMM which contained 2 hidden states and 6 Gaussian distributions to make comparison with GHMM. Results illustrate that true prediction rate of GHMMs is significantly higher than that of traditional HMMs when applied on the prediction of samples over 250 μg·m−3; while both GHMMs and HMMs have similar performances when applied on the prediction of samples blow 150 μg·m−3.