Accurate　PM2.5　concentration　prediction　can　provide　reliable　air　pollution　warning　information　to　the　public.　However,　previous　studies　have　often　focused　on　the　data　of　the　target　city　itself,　ignoring　the　interaction　among　cities　in　the　same　region.　In　this　paper,　we　develop　a　multi-scale　ensemble　learning　approach　to　forecast　daily　PM2.5　concentrations　of　the　target　city　by　modeling　its　air　and　climate　indicators,　and　PM2.5　value　of　its　neighboring　cities.　First,　the　proposed　approach　smooths　the　multivariate　data　by　singular　spectrum　analysis　and　performs　multi-feature　selection　based　on　distance　factor　and　predictive　power　of　data.　Second,　the　inherent　association　among　the　obtained　multiple　features　is　captured　by　multivariate　empirical　modal　decomposition.　Third,　the　Hurst　exponent　is　applied　to　match　each　time　scale　with　the　corresponding　predictor　for　multi-step　prediction.　Finally,　the　forecasting　values　of　all　time　scales　are　summed　to　obtain　the　PM2.5　concentration　forecasting　results　of　the　target　city.　Four　experiments　involving　Beijing,　Wuhan,　and　Shenzhen　are　carried　out　to　verify　the　accuracy　and　robustness　of　the　proposed　approach.　The　experimental　results　show　that　our　approach　outperforms　all　benchmark　models,　and　introducing　city　synergy　strategy　can　improve　the　forecasting　performance　significantly.