In order to solve the problem of simultaneous multi-site PM2.5 prediction, a hierarchical autoregressive spatio-temporal model is proposed under the Bayesian framework. The true daily average concentration of PM2.5 is regarded as a potential spatio-temporal process. The first-order autoregressive process is used to describe the temporal correlation, and the spatial correlation is captured based on the Matérn process, which greatly improves the efficiency of dimension reduction and synchronous prediction. In addition, meteorological factors such as daily maximum temperature, relative humidity and wind speed are used as explanatory variables to improve the accuracy of prediction. Thanks to the hierarchical structure of the model, the parameter estimation and prediction process can be realized by Bayesian method and MCMC. The empirical analysis of daily PM2.5 concentration in Beijing shows that the proposed model has good interpolation or prediction effect in spatial and temporal dimensions.