Environment　protection　department　need　to　grasp　the　concentration　of　PM2.5　in　a　future　moment　when　monitoring.　However,　the　existing　PM2.5　prediction　studies　only　forecast　short-term　time　points,　and　cannot　accurately　give　the　trend　of　the　next　period　of　time.　In　this　paper,　a　PM2.5　prediction　model　based　on　Att-ConvLSTM　model　integrated　training　method　is　established　by　the　advantage　of　ConvLSTM　to　obtain　spatiotemporal　information.　Then,　Experiments　were　performed　using　DNN,　ARIMA　and　LSTM　as　control　model　with　Att-ConvLSTM　model　and　used　it　for　application　test.　The　result　demonstrated　that　prediction　model　can　extract　spatiotemporal　features　with　attention　mechanism　and　ConvLSTM.　The　model　can　reduce　the　generalization　error　of　the　model　when　predicting　other　observation　points.　©　Springer　Nature　Switzerland　AG,　2020.