In　this　paper,　a　temporal-spatial　fusion　model　is　proposed　for　PM2.5　concentration　prediction.　The　model　uses　historical　PM2.5　concentration　and　meteorological　data　as　input　of　the　model　to　make　hourly　predictions　of　PM2.5　concentration.　This　model　consists　of　three　parts:　1)　Long　short-term　memory　neural　network　predictor　based　on　time　dimension,　2)　Artificial　neural　network　predictor　based　on　spatial　dimension,　3)　Model　tree　predictor　based　on　temporal-spatial　fusion.　This　method　combines　the　forecast　results　of　two　dimensions　in　space　and　time　dynamically,　as　the　spatial　and　temporal　correlation　of　data　is　considered.　Experimental　results　show　this　model　performs　better　than　predicting　from　a　single　dimension,　confirming　the　effectiveness　of　the　model.