Air　pollution　forecasting　provides　early　warning　before　air　pollution　issue　occurs,　thus　protects　human　health　and　living　environment.　A　neural　network　model　optimized　by　genetic　algorithm　was　developed　in　order　to　predict　PM10　concentrations　in　Beijing.　The　genetic　algorithm　was　used　to　optimize　the　initial　weights　and　threshold　of　the　BP　neural　network　in　simulation.　Astringency　of　network　and　accuracy　of　prediction　were　effectively　improved.　The　improved　network　and　Models-3　Community　Multi-scale　Air　Quality　(CMAQ)　modeling　system　were　both　applied　in　the　prediction　of　short-term　PM10　concentration　in　autumn　2002　in　Beijing.　Results　showed　good　prediction　capability　of　both　models,　and　the　mean　relative　errors　were　separately　0.21　and　0.26.　When　applied　in　short-term　air　pollution　forecasting,　neural　network　is　of　similar　prediction　accuracy　compared　with　CMAQ.　It　is　an　effective　alternate　method　for　air　pollution　forecasting　in　areas　where　mathematical　model　on　air　pollution　can＇t　be　widely　applied.