Flow　and　speed　are　crucial　for　assessing　the　evolution　of　traffic　conditions.　Current　traffic　prediction　research　faces　two　critical　limitations.　First,　existing　methods　fail　to　effectively　capture　multi-scale　temporal　features　in　traffic　sequences,　where　traffic　variations　exhibit　different　patterns　across　time　scales,　particularly　daily　and　weekly　periodicities.　Second,　the　shared　parameter　paradigm　in　Graph　Convolutional　Networks　limits　nodes’　ability　to　express　unique　characteristics,　making　it　challenging　to　capture　personalized　traffic　patterns　when　nodes　have　significantly　different　contexts.　This　work　proposes　a　novel　Multi-scale　Temporal　Enhanced　Graph　Convolutional　Recurrent　Network　(MTEGCRN)　to　address　these　limitations　simultaneously.　In　the　temporal　domain,　a　temporal　feature　enhancement　module　introduces　daily　and　weekly　periodic　features　through　trainable　temporal　embeddings,　enabling　multi-scale　feature　learning　with　standard　input　lengths.　A　continuous　temporal　learning　module　integrates　spatial　and　temporal　learning　within　a　unified　framework,　while　a　global　temporal　fusion　module　employing　Transformers　captures　global　dependencies　for　long-term　prediction.　In　the　spatial　domain,　a　node-oriented　graph　convolutional　network　breaks　the　shared　parameter　paradigm　by　allocating　personalized　parameter　spaces　and　high-dimensional　temporal　periodic　features　to　each　node,　enabling　the　capture　of　node-specific　traffic　patterns.　Experiments　on　five　public　datasets　demonstrate　that　MTEGCRN　significantly　outperforms　all　baseline　methods.　Compared　to　the　Dynamic　Graph　Convolutional　Recurrent　Network　model,　MTEGCRN　reduces　the　mean　absolute　error　by　9.51　%,　3.70　%,　12.02　%,　7.29　%　and　9.94　%　on　the　five　datasets,　respectively.　The　code　for　MTEGCRN　is　available　at　https://github.com/OvOYu/MTEGCRN.　©　2025