As　the　availability　of　air　quality　data　collected　at　ground-based　monitoring　stations　increases,　the　researchers　use　the　data　in　sophisticated　models　to　predict　the　concentration　of　different　pollutants.　This　study　analyzed　the　concentration　of　PM2.5　in　Beijing　to　mine　the　long-term　trend　of　air　quality.　The　results　showed　that　PM2.5　is　in　the　trend　of　decreasing　year　by　year　but　still　above　the　annual　maximum　limit　(35　mu　g/m(3))　of　WHO　with　strong　seasonality.　Besides,　this　study　proposed　an　attention　mechanism　(AM)-based　prediction　method,　named　MSAQP.　Firstly,　attention　mechanism　was　introduced　into　the　decoding　phase　of　MSAQP　to　calculate　the　context　vector.　The　attention　mechanism　learned　the　weight　distribution　strategy　of　the　original　data　and　integrates　all　the　coding　states　into　the　context　vector　to　enhance　the　representation　ability　of　time　characteristics.　Secondly,　due　to　the　problems　of　gradient　explosion　and　gradient　disappearance　in　Recurrent　Neural　Network　(RNN),　this　study　adopted　long　short-term　memory　network　(LSTM).　In　addition,　three　different　loss　functions　were　applied　to　the　training　experiment　of　the　model,　respectively.　The　experimental　results　showed　that　the　prediction　accuracy　was　improved,　among　which　the　MAE　was　reduced　by　3.42,　the　NMSE　was　reduced　by　0.01,　and　the　R-2　was　improved　by　0.24.