Accurate　and　fine-grained　prediction　of　PM2.5　concentration　is　of　great　significance　for　air　quality　control　and　human　physical　and　mental　health.　Traditional　approaches,　such　as　time　series,　recurrent　neural　networks　(RNNs)　or　graph　convolutional　networks　(GCNs),　cannot　effectively　integrate　spatial-temporal　and　meteorological　factors　and　manage　dynamic　edge　relationships　among　scattered　monitoring　stations.　In　this　paper,　a　spatial-temporal　causal　convolution　network　framework,　ST-CCN-PM2.5,　is　proposed.　Both　the　spatial　effects　of　multi-source　air　pollutants　and　meteorological　factors　are　considered　via　spatial　attention　mechanism.　Time-dependent　features　in　causal　convolution　networks　are　extracted　by　stacked　dilated　convolution　and　time　attention.　All　the　hyper-parameters　in　ST-CCN-PM2.5　are　tuned　by　Bayesian　optimization.　Haikou　air　monitoring　station　data　are　employed　with　a　series　of　baselines　(AR,　MA,　ARMA,　ANN,　SVR,　GRU,　LSTM　and　ST-GCN).　Final　results　include　the　following　points:　(1)　For　a　single　station,　the　RMSE,　MAE　and　R-2　values　of　ST-CCN-PM2.5　decreased　by　27.05%,　10.38%　and　3.56%　on　average,　respectively.　(2)　For　all　stations,　ST-CCN-PM2.5　achieve　the　best　performance　in　win-tie-loss　experiments.　The　numbers　of　winning　stations　are　68,　63,　and　64　out　of　95　stations　in　RMSE　(MSE),　MAE,　and　R-2,　respectively.　In　addition,　the　mean　MSE,　RMSE　and　MAE　of　ST-CCN-PM2.5　are　4.94,　2.17　and　1.31,　respectively,　and　the　R-2　value　is　0.92.　(3)　Shapley　analysis　shows　wind　speed　is　the　most　influencing　factor　in　fine-grained　PM2.5　concentration　prediction.　The　effects　of　CO　and　temperature　on　PM2.5　prediction　are　moderately　significant.　Friedman　test　under　different　resampling　further　confirms　the　advantage　of　ST-CCN-PM2.5.　The　ST-CCN-PM2.5　provides　a　promising　direction　for　fine-grained　PM2.5　prediction.