Federated Spatial–Temporal Synchronous Graph Convolutional Network for Cross-City Traffic Prediction

Accurate traffic forecasting plays a vital role in modern intelligent transportation systems and enables proactive traffic management and well-informed decision-making. Although existing deep learning-based models have achieved remarkable progress, they often rely heavily on complete data and struggle to generalize across cities with diverse spatial-temporal patterns and limited observations. Moreover, the direct sharing of raw data between cities is frequently infeasible because of increasing privacy concerns. To address these limitations, we propose a novel framework named federated spatial-temporal synchronous graph convolutional network (Fed-STSGCN), which enables cross-city traffic prediction without exposing private data. Fed-STSGCN adopts a decentralized learning paradigm that is aligned with federated and privacy-preserving learning, which enables autonomous local model training and collaborative knowledge sharing across distributed traffic agents. The model uses a spatial-temporal synchronous graph construction strategy that jointly captures spatial correlations and temporal causality. On top of this structure, we design stacked spatial-temporal synchronous graph convolution modules to extract high-level representations and employ a cosine similarity-aware aggregation mechanism to improve global knowledge transfer under the federated setting. By supporting privacy-aware cross-region adaptation under data-scarce conditions, the proposed framework is well suited for consumer-facing intelligent transportation applications, in which traffic data are generated and processed in a distributed manner across heterogeneous sensing and computing environments. Extensive experiments are conducted on twelve cross-city forecasting tasks derived from four real-world datasets. The results show that Fed-STSGCN achieves notable error reductions across all the considered evaluation metrics, which range from 0.32%-13.22% in traffic flow prediction and from 2.04%-12.78% in traffic speed prediction, compared with the state-of-the-art baseline, thus demonstrating its robustness and effectiveness in cross-city and data-scarce settings.