Accurate　traffic　forecasting　on　citywide　networks　is　one　of　the　crucial　urban　data　mining　applications　that　accurately　provide　congestion　warning　and　transportation　scheduling.　While　previous　work　has　made　significant　efforts　to　learn　traffic　temporal　dynamics　and　spatial　dependencies,　two　key　limitations　exist　in　current　models:　(1)　most　existing　approaches　solely　capture　spatial　correlations　among　neighbors　on　predefined　graph　structure,　and　genuine　relation　may　be　missing　due　to　the　incomplete　graph　connections;　and　(2)　existing　methods　are　defective　to　capture　the　temporal　trends　because　the　recurrent　and　stacking　structure　employed　in　these　methods　suffers　from　the　long-range　temporal　dependency　vanquish　problem.　To　overcome　the　difficulty　in　multistep　prediction　and　further　capture　the　dynamic　spatial-temporal　dependencies　of　traffic　flows,　we　propose　a　new　traffic　speed　prediction　framework　for　multiscale　graph　attention　networks　(MS-GATNs).　In　particular,　MS-GATNs　is　a　hierarchically　structured　graph　neural　architecture　that　learns　not　only　the　local　region-wise　geographical　dependencies　but　also　the　spatial　semantics　from　a　global　perspective.　Furthermore,　a　multiheads　attention　mechanism　is　introduced　to　empower　our　model　with　the　capability　of　capturing　complex　nonstationary　temporal　dynamics.　Experiments　on　real-world　traffic　data　sets　demonstrate　that　MS-GATNs　outperforms　the　state-of-the-art　baselines　in　long-term　forecasting.