PM2.5　is　difficult　to　accurately　forecast　due　to　the　influence　of　multiple　meteorological　and　pollutant　variables　in　the　complex　nonlinear　dynamic　atmosphere　system.　In　this　paper,　an　Elman　neural　network　prediction　method　based　on　chaos　theory　is　put　forward　for　the　problem.　Firstly,　the　chaotic　characteristics　of　the　concentration　of　the　PM2.5　are　analyzed　and　verified　from　the　correlation　dimension,　the　maximum　Lyapunov　exponent　and　the　Kolmogorov　entropy.　Then,　phase　space　reconstruction　technique　of　chaotic　theory　is　adopted　to　reconstruct　the　phase　space　of　PM2.5　time　series.　The　reconstructed　phase　space　and　the　future　concentration　of　PM2.5　are　taken　as　the　input　and　output　of　the　Elman　neural　network　with　chaos　theory　(Elman-chaos)　respectively.　The　numerical　and　experimental　analyses　show　that　this　method　is　proportionally　superior　to　that　without　considering　the　chaos　characteristics　and　other　approaches.　The　Elman-chaos　prediction　model　has　better　prediction　performance　and　application　value.　©　2016　TCCT.