Spatio-Temporal Fusion Network for Traffic Flow Prediction

Abstract

Intelligent transportation systems (ITSs) rely heavily on traffic flow forecasting to facilitate efficient traffic management and alleviate congestion. This paper proposes an innovative hybrid network-spatio-temporal fusion network (STFN), designed to forecast traffic flow with accuracy and efficacy. To capture spatial dependencies and temporal dynamics in traffic data effectively, the proposed model incorporates two key components: residual graph attention networks (RGAT) to model complex spatial correlations between adjacent traffic sensors and bidirectional LSTM (Bi-LSTM) to extract latent temporal features and sequential patterns. An attentional feature fusion (AFF) mechanism is introduced to adaptively integrate complementary features from heterogeneous sources. Experimental results indicate the superior performance of STFN compared to state-of-the-art methods regarding accuracy and efficiency.