MDDF

Short-term multimodal travel demand forecasting is essential for intelligent transportation systems, yet remains challenging due to spatiotemporal heterogeneity and complex inter-modal dependencies.A critical limitation of existing approaches lies in the inadequate separation of shared patterns from modality-specific dynamics, leading to feature interference and spatial misalignment across transportation modes with heterogeneous coverage.To address these issues, this paper proposes a Multimodal Disentangled Dynamic Fusion (MDDF) framework comprising three key components: (i) a spatiotemporal disentangled representation learning module with dual-path encoding and orthogonality constraints to decouple shared and modality-specific features; (ii) a distribution alignment module employing Jensen-Shannon divergence (JSD) regularization and geographically masked attention to handle spatial inconsistencies; and (iii) a cross-modal attention fusion module with ranking loss to dynamically evaluate modality importance.Experiments on Shenzhen multimodal transportation data demonstrate that MDDF significantly outperforms state-of-the-art baselines, achieving MAE reductions of up to 19.0% for metro, 10.1% for bus, and 10.0% for taxi at 60-minute horizons compared with the strongest baseline.The framework effectively addresses feature entanglement and spatial misalignment, providing a principled approach for spatiotemporal multimodal data integration in urban mobility prediction.