To　achieve　accurate　and　fine-grained　PM2.　5　concentration　prediction,　this　paper　proposed　a　PM2.　5　concentration　prediction　model　based　on　spatial-temporal　cognitive　dilated　convolution　network　(STC-DCN).　The　model　applied　spatial-temporal　factors　and　meteorological　factors　to　PM2.　5　concentration　prediction,　extracted　spatial-temporal　features　based　on　causal　convolution　network,　and　optimized　the　extraction　of　spatial-temporal　features　by　using　spatial-temporal　attention　mechanism.　Experimental　results　based　on　Haikou　air　pollution　data　show　that　for　a　single　monitoring　station,　compared　with　the　baseline　model,　the　RMSE　of　ST-C-DCN　decreases　by　24.　7%　on　average,　the　MAE　decreases　by　9.　93%　on　average,　and　the　R2　increases　by　3.　35%　on　average.　ST-C-DCN　outperforms　other　models　in　terms　of　prediction　accuracy　for　all　monitoring　stations,　achieving　the　highest　scores　in　win-tie-loss　experiments　(including　MSE,　RMSE,　MAE,　and　R2),　with　values　of　68,　63,　and　64,　respectively.　The　Friedman　test　conducted　under　different　data　sampling　conditions　confirms　that　ST-C-DCN　exhibits　significant　performance　improvement　compared　to　the　baseline　model.　In　conclusion,　ST-C-DCN　provides　a　potential　direction　for　fine-grained　PM2.　5　prediction.　©　2024　Beijing　University　of　Technology.　All　rights　reserved.