STFCT: A Multi-Timescale Spatial-Temporal Transformer for Traffic Flow Prediction

Accurate traffic flow prediction plays a crucial role in modern transportation management systems, enabling extensive applications ranging from congestion warning to optimized route planning. While current approaches have achieved progress in specific areas, they continue to face challenges such as multi-scale dynamics and constrained spatiotemporal modeling capacity. Addressing these limitations, we introduce a innovative model termed the Spatial-Temporal Fusion Convolution Transformer (STFCT). This framework integrates periodic patterns and traffic characteristics via adaptive spatiotemporal embeddings to produce a unified representation capturing both spatial and temporal relationships. Our architecture incorporates a gating mechanism for dynamic spatiotemporal integration, along with a temporal convolution component to simultaneously capture both short- and medium-term patterns. Experimental results from three different traffic datasets reveal STFCT’s advantages over competing methods in terms of all assessment indicators.