This　paper　uses　the　data　of　hourly　concentration　of　PM2.5　in　London　and　the　traditional　BP　neural　network　to　build　forecast　model　and　to　quantitively　forecast　the　hourly　concentration　of　PM2.5　in　London,　discusses　the　impacts　of　enlarging　sample　data,　reducing　noise　of　sample　data　and　adding　weather　factors　in　inputing　vector　on　setting　up　the　model　of　the　air　pollution　forecast　network.　Finally,　it　comes　to　the　conclusion　that　properly　selecting　sample　data　and　adding　weather　factors　is　beneficial　to　improving　forecasting　precision　of　the　network　model.