PM2.5　elements　have　a　great　impact　on　air　quality,　so　it　is　of　great　significance　to　predict　PM2.5　concentration　for　People＇s　Daily　life　and　health.　Aiming　at　the　problem　of　low　prediction　accuracy　of　existing　models,　we　propose　a　spatial-temporal　attention　neural　network　(STAN).　Firstly,　we　introduce　a　spatial　attention　module　to　adaptively　extract　spatial　features　between　monitoring　stations.　Then,　we　use　a　temporal　attention　to　extract　features　from　encoder　hidden　states　across　time　series.　We　evaluate　the　STAN　on　PM2.5　prediction　with　data　from　Beijing　observation　stations,　and　the　results　show　that　it　is　superior　to　ARIMA　LSTM　and　Seq2seq　models　in　predicting　PM2.5　concentration.