Spatio-temporal feature fusion and few-shot learning framework with cross-city knowledge transfer for traffic flow prediction

Accurate traffic prediction across urban networks faces four fundamental challenges: (1) nonlinear spatio-temporal correlations varying across cities and time, (2) data scarcity in newly deployed regions due to sparse sensor coverage, (3) computational inefficiency when scaling to large metropolitan networks, and (4) heterogeneous road topologies hindering cross-city generalisation. We propose ST-FSFusion, a Spatio-Temporal Feature Fusion and Few-Shot Learning Framework with Cross-City Knowledge Transfer that addresses these challenges through an integrated modular architecture comprising five components. First, the Spatio-Temporal Meta-Knowledge Learner (ST-MKL) captures transferable traffic patterns using dual spatial–temporal processing pathways and adaptive graph transformers while preserving city-specific characteristics. Second, the Temporal Encoder Wrapper (TEWrapper) provides dual-pathway adaptive encoding (MetaT-RobustT) that switches between meta-enhanced and standard encoders based on data availability, ensuring robust performance under both data-rich and data-scarce conditions. Third, the Cross-Domain Feature Fusion (CDFusion) hierarchically combines spatial and temporal representations while promoting topological consistency via graph reconstruction regularisation. Fourth, the Cross-City Adapter (CAdapter) and Few-Shot Learner (FSL) jointly enable efficient deployment in new cities, with CAdapter enabling selective knowledge transfer and FSL supporting rapid adaptation via prototype-based learning constrained by road-network structure. Experiments on four real-world urban datasets demonstrate that ST-FSFusion achieves consistent improvement over state-of-the-art methods in terms of both MAE and RMSE (Mean Absolute Error and Root Mean Square Error) under few-shot settings (3 days, 5%–10% of training data) and exhibits strong noise resilience and robustness under missing data conditions, with ablation studies confirming ST-MKL as the most critical component. The framework offers a scalable, practical solution for intelligent transportation systems in cities with limited historical data or computational resources.The implementation code is available at https://github.com/afofanah/ST_FSFusion.