The　time-varying　property　of　traffic　networks　has　brought　a　problem　of　modeling　large-scale　dynamic　networks.　Based　on　the　real-time　traffic　sensing　data　of　the　road　network,　community　division　and　prediction　can　effectively　reduce　the　complexity　of　local　management　in　urban　regions.　However,　traffic-based　communities　have　complex　topology　and　real-time　dynamic　features,　and　traditional　community　division　and　topology　prediction　cannot　effectively　be　applied　to　this　structure.　Therefore,　we　propose　a　dynamic　traffic　community　prediction　model　based　on　hierarchical　graph　attention　network.　It　uses　the　hierarchical　features　fusion　with　spatiotemporal　convolution　and　the　ADGCN　proposed　in　this　paper　to　compose　a　hierarchical　graph　attention　architecture.　In　which,　each　layer　component　coordinates　to　perform　different　features　extraction　for　capturing　traffic　community　of　road　network　in　different　time　periods　respectively.　Finally,　the　output　features　of　each　layer　are　combined　to　represent　the　dynamically　divided　regions　in　the　traffic　network.　The　effectiveness　of　the　model　was　verified　in　experiments　on　the　Xi＇an　urban　traffic　dataset.　©　2021,　Springer　Nature　Switzerland　AG.