At　present,　there　are　widespread　air　pollution　problems　in　most　parts　of　China,　the　accurate　prediction　of　atmospheric　pollutant　concentration　has　become　a　hot　issue　for　people　to　study.　This　paper　proposes　the　NDFA-LSSVM　model　to　predict　the　concentration　of　PM2.5.　The　hyper-parameter　of　Least　Square　Support　Vector　Machine　(LSSVM)　were　optimized　by　using　the　New　Dynamic　Firefly　Algorithm　(NDFA)　to　establish　a　PM2.5　concentration　prediction　model　NDFA-LSSVM.　The　air　quality　data　of　monitoring　stations　at　Chaoyang　Agricultural　Exhibition　Hall　District　was　used　as　source　data　to　compare　the　performance　of　the　optimized　model　with　LSSVM　model　and　General　Regression　Neural　Network　(GRNN)　model.　The　experimental　results　show　that　the　NDFA-LSSVM　model　proposed　in　this　paper　effectively　improves　the　prediction　accuracy　of　PM2.5　concentration.