Urban　traffic　congestion　significantly　affects　economic　productivity,　environmental　sustainability,　and　quality　of　life.　Traditional　traffic　congestion　prediction　models,　which　are　primarily　based　on　single-mode　data,　often　fail　to　capture　the　intricate　interactions　and　dependencies　present　in　multimodal　urban　transportation　systems.　In　this　paper,　we　propose　a　novel　Mixture　of　Spatial–Temporal　Graph　Transformer　Networks　(MoSTGTN)　for　urban　congestion　prediction　using　multimodal　transportation　data.　Our　approach　integrates　a　pattern-aware　dynamic　graph　neural　network　for　analyzing　grid-based　traffic　big　data,　a　dynamic　spatial–temporal　graph　transformer　to　learn　the　distinct　features　of　each　traffic　modality,　and　a　mixture　of　experts　framework　to　capture　both　intra-modality　characteristics　and　inter-modality　interactions.　Real-world　multimodal　transportation　data　from　Tianjin,　China,　including　taxi　and　shared　bicycle　data,　is　used　to　validate　the　proposed　model.　The　MoSTGTN　model　consistently　outperforms　state-of-the-art　baseline　models　across　multiple　metrics,　as　well　as　in　various　traffic　regions,　time　periods,　and　directions.　The　robustness　of　the　model　is　further　confirmed　through　extensive　variance　analysis.　Our　key　findings　highlight　the　critical　role　of　inter-modal　interactions　and　underscore　the　significant　contribution　of　shared　bicycle　data　in　improving　urban　traffic　congestion　prediction　accuracy,　especially　in　high-density　traffic　areas.　©　2025　Elsevier　Ltd