Dynamic Knowledge Driven Multi-Scale Spatiotemporal Fusion Framework for Reliable Multi-Horizon Traffic Flow Forecasting

Abstract

Accurate traffic flow forecasting remains a core challenge in intelligent transportation systems. Urban traffic patterns are characterised by complex spatiotemporal dependencies and inherent uncertainties, which complicate traffic forecasting. This study proposes a Dynamic Knowledge-Driven Spatiotemporal Fusion (DKDMSF) architecture to extract features of Multi-Scale traffic dynamics across various temporal traffic horizons. DKDMSF integrates Dynamic Spatial Attention (DSA) and Adaptive Graph Learning (AGL), a context-aware parallel encoder, and cross-modal fusion of variational inference for uncertainty quantification. DSA constructs a context-aware spatial inference graph by adaptively modifying node weights in response to changing traffic patterns. The AGL uncovers the optimal traffic network topology by learning node connectivity using dynamic traffic flow data. Context-aware parallel encoders extract temporal patterns via periodic temporal encoding and are fused with probabilistic uncertainty estimation for traffic forecasting. The proposed framework was evaluated on the METR-LA, PEMS-BAY, NYC-Taxi, SZ-Taxi, and TomTom datasets to demonstrate the model's robustness. The results demonstrate better performance and narrower confidence intervals than existing models across different prediction horizons, making it suitable for various traffic applications.