In　order　to　solve　the　problem　of　low　prediction　accuracy　of　single　model,　this　paper　proposes　a　PM2.5　prediction　method　based　on　multi-model　fusion.　We　have　constructed　multiple　Least　Angle　Regression　(LARS)　models　based　on　atmospheric　data　onto　different　weather　conditions,　and　the　prediction　results　are　sent　to　BP　neural　network　for　decision-level　fusion.　Using　the　monitoring　data　of　Beijing　University　of　Technology　in　Chaoyang　District,　the　atmospheric　data　of　2018-01-01　∼　2018-10-31　was　selected　as　the　experimental　research　object.　The　simulation　result　shows　that　the　single　model　is　not　ideal　for　predicting　PM2.5　concentration　on　complex　weather　conditions.　The　multi-model　fusion　method　can　effectively　solve　the　error　preference　problem　of　single　model　and　improve　the　prediction　accuracy.　©　2019　IEEE.