Zero-shot urban flow generation via KG-guided spatiotemporal diffusion

Urban flow data is crucial for smart city decisions but challenging to acquire in areas lacking historical observations. Existing methods rely heavily on region-specific data, limiting generalization, and focus primarily on prediction rather than generation. To address this, we propose a generative AI approach (UKG-STLDM) using an Urban Knowledge Graph (UKG) to model the urban environment and relationships between regions. UKG embeddings condition a Latent Diffusion Model (LDM) enhanced with a transformer-based temporal block capturing time dynamics and a GCN-based spatial block modeling spatial correlations. This spatio-temporally augmented model generates high-fidelity urban flow data with complex patterns. Experiments demonstrate significant improvements over benchmarks: a 10.07% reduction in MAE and a 13.0% improvement in MMD compared to the best baseline, confirming superior accuracy and distribution fidelity. This provides a practical solution for data-scarce scenarios in urban planning and traffic simulation.