Efficient Traffic State Prediction With Dynamic Joint Spatio-Temporal Relation Inference

Researchers introduce STEI-PCN, a convolutional neural network designed to improve traffic flow prediction by efficiently modeling spatial interactions, temporal patterns, and their dynamic relationships across road networks. The method achieves competitive accuracy on standard benchmarks while maintaining lower computational costs than existing complex spatio-temporal models.

Traffic prediction remains a fundamental challenge in smart city infrastructure and autonomous systems, requiring models that capture how conditions at neighboring locations and past timestamps influence future traffic states. STEI-PCN addresses this by combining spatio-temporal encoding with dynamic relation inference, using position-aware attention mechanisms to weight interactions between road sensors. The approach separates local joint dependencies from long-range temporal patterns through specialized convolutional components, enabling more interpretable and efficient predictions.

The research emerges from ongoing efforts to balance model complexity with practical deployment requirements. Previous architectures either treated spatial and temporal dimensions independently, missing crucial interactions, or employed unified structures that became computationally prohibitive at scale. STEI-PCN's pure convolutional design avoids recurrent layers and complex graph operations while maintaining expressiveness through clever encoding strategies and multi-view fusion.

The results across PeMS datasets demonstrate practical utility for transportation systems, ride-sharing platforms, and city planning applications that depend on accurate traffic forecasting. By reducing training and inference costs while preserving accuracy, the method makes deployment on edge devices and real-time systems more feasible. This efficiency matters for cities managing dense sensor networks and continuous prediction tasks where computational budgets are constrained.

Future developments likely focus on handling extreme traffic disruptions, integrating external events, and scaling to heterogeneous network topologies. The public code release enables broader adoption and benchmarking against specialized baselines.

• →STEI-PCN achieves competitive traffic prediction accuracy with lower computational overhead than existing spatio-temporal models
• →Dynamic relation inference through position and distance encodings enables efficient capture of complex traffic interactions
• →Pure convolutional architecture avoids recurrent layers while maintaining performance, improving real-time deployment feasibility
• →Multi-view prediction module fuses multiple temporal scales for improved multi-step forecasting accuracy
• →Empirical validation on five public traffic datasets demonstrates consistent performance under varying prediction horizons