With　rapid　modernization,　air　quality　is　becoming　gradually　deteriorate.　To　avoid　the　adverse　effects　of　severe　air　pollution　on　human　health,　we　need　accurate　real-time　air　quality　prediction.　The　factors　relevant　to　air　pollutant　concentration　forecasting　contain　simultaneously　numeric　type　(temperature　etc.)　and　non-numeric　type　(wind　direction　etc.).　Random　Forest　has　many　advantages,　which　includes　it　can　deal　with　numeric　features　and　non-numeric　features.　In　this　study,　an　online　forecasting　method　based　on　Random　Forest　is　proposed　to　predict　the　concentrations　of　three　kinds　of　air　pollutants　(PM2.5,　NO2,　SO2),　24　hours　in　advance.　The　sliding　window　is　used　to　take　the　recent　data　to　retrain　Random　Forest　prediction　model　and　the　well-trained　models　is　used　to　predict　the　dependent　variable　at　target　moment.　Before　prediction　model　is　trained,　a　variable　selection　method　based　on　Random　Forests　(VSURF)　is　used　to　select　the　factors　that　are　relevant　to　the　forecast　of　air　pollutant　concentrations.　We　evaluate　our　method　with　dataset　from　Microsoft　Research.　Comparison　with　baseline　methods　shows　that　our　method　achieve　state-of-art　performance　on　air　pollutant　concentration　forecasting.　Experimental　results　also　indicate　that　the　features　we　selected　using　VSURF　method　are　most　important　predictors　for　the　prediction　of　three　kinds　of　air　pollutant　concentrations.