With　the　development　of　auto-monitoring　technology,　mass　data　were　acquired.　How　to　process　the　mass　data　to　extract　useful　information　becomes　an　urgent　problem.　Data　mining　is　one　of　the　methods　to　solve　the　problem.　In　this　paper　BP　neural　network　which　was　trained　using　Bayesian　Regularization　method　based　on　mass　monitoring　data　was　used　to　forecast　the　hourly　average　concentration　of　PM2.5.　Three　strategies　in　order　to　improve　the　predictions　are　analyzed:　sample　data　classification　according　to　seasons,　using　a　larger　training　data　set,　and　consideration　of　the　values　of　meteorological　variables.　The　results　show　that　data　mining　based　on　Neural　Networks　has　good　prediction　precision.