Atmospheric　pollution　prediction　is　helpful　to　find　effective　ways　to　control　air　pollution　and　improve　air　quality.　BP　neural　network　which　was　trained　using　Bayesian　Regularization　method　and　early　stopping　method　was　used　to　forecast　the　hourly　concentration　of　PM2.5.　The　data　of　PM2.5　hourly　concentration　were　obtained　from　monitoring　site　of　Marylebone　Road　in　London,　UK.　The　prediction　relative　error　goes　from　20%　to　49%.　The　results　show　that　compared　with　networks　which　were　trained　using　other　methods,　Bayesian　Regularization　method　and　early　stopping　method　can　improve　the　generalization　ability　of　BP　neural　network.