Traffic　prediction　is　indispensable　for　constructing　transportation　networks　in　smart　cities.　Due　to　the　complex　spatio-temporal　correlations　of　traffic　data,　this　task　presents　challenges.　Recent　studies　mainly　use　graph　neural　networks　to　simulate　complex　spatio-temporal　relationships　through　fixed　or　adaptive　graphs.　While　fixed　graphs　may　not　adapt　to　data　drift　caused　by　changes　in　road　network　structures,　adaptive　graphs　overlook　critical　information　of　the　original　roads.　To　address　this　challenge,　we　propose　a　principal　spatio-temporal　causal　graph　convolutional　network　(PSTCGCN)　to　accurately　predict　traffic　flow.　In　response　to　the　data　drift　problem,　we　introduce　a　data-driven　semi-principal　generated　graph　embedding　(SPGGE)　that　first　extracts　the　principal　features　of　the　original　roads　to　model　the　spatio-temporal　sequence　distribution　and　then　remodels　the　data　after　drift　through　adaptive　transformation.　Traffic　flow　data,　while　showcasing　fundamental　spatial　relationships,　also　exhibit　temporal　dynamics.　We　propose　an　effective　temporal　causal　convolution　component　comprising　SPGGE,　graph　convolution,　both　local　and　global　causal　learning　models　to　jointly　learn　short-term　and　long-term　spatio-temporal　correlations.　PSTCGCN　was　evaluated　using　two　actual　highway　datasets,　PEMS03　and　PEMS07,　achieving　a　notable　improvement　of　6.12%　in　RMSE　on　PEMS03　compared　to　STGATRGN.　Our　code　is　available　at　https://github.com/OvOYu/PSTCGCN.　©　The　Author(s),　under　exclusive　licence　to　Springer-Verlag　London　Ltd.,　part　of　Springer　Nature　2024.