PM2.5　concentration　is　an　important　evaluation　index　of　urban　air　quality.　Improving　the　prediction　and　early　warning　ability　of　PM2.5　concentration　is　of　great　significance　for　guiding　residents＇　real-time　protection　and　efficient　control　of　air　pollution.　A　large　number　of　studies　have　shown　that　the　temporal　and　spatial　correlation　between　PM2.5　concentration　of　surrounding　stations　and　target　stations　directly　affects　the　prediction　accuracy　of　the　model.　There　is　still　a　lack　on　how　to　select　strongly　correlated　and　representative　temporal　and　spatial　information,　and　the　selection　methods　are　somewhat　subjective.　In　this　paper,　air　pollutants　and　meteorological　factors　are　taken　as　the　influencing　factors,　and　a　PM2.5　prediction　model　(ST-CCN-PM2.5)　based　on　causal　convolution　network　is　proposed　to　improve　the　performance　of　PM2.5　concentration　prediction　at　a　fine-grained　spatialoral　scale.　We　use　spatial　attention　mechanism　to　fuse　spatial　correlation　information,　and　optimize　threshold　parameters.　Based　on　time　attention　mechanism,　the　sliding　window　size　is　adjusted　to　dilate　the　time-series　input.　Finally,　the　hourly　records　of　air　pollutants　and　meteorological　factors　monitored　by　95　monitoring　stations　in　Haikou　city　are　employed,　and　the　performance　is　compared　with　baseline.　Compared　with　AR　and　ARMA　models,　the　MSE　values　decrease　by　38.9%　and　40.9%,　and　R2　values　increase　by　2.3%　and　2.6%,　respectively.　Compared　with　GRU,　SVR,　LSTM　and　ANN　models,　the　MSE　value　of　ST-CCN-PM2.5　decreases　by　44.8%,　45.4%,　46.5%　and　49.0%,　respectively,　and　its　R2　value　increases　by　3.6%.　The　final　results　show　that　the　prediction　performance　of　ST-CCN-PM2.5　model　is　significantly　improved　compared　with　the　baseline,　which　proves　that　the　model　has　stronger　reflection　ability　and　prediction　generalization　ability　for　dynamic　nonlinear　system,　and　proves　the　potential　of　the　model　in　the　prediction　of　fine　particle　PM2.5　concentration.　©　2021　ACM.