STGCAM: frequency-domain enhanced asymmetric spatiotemporal graph attention network for traffic flow forecasting

Traffic flow forecasting is a key task in intelligent transportation systems, supporting congestion warning, signal control, and route guidance. However, real traffic flow is affected by periodic travel patterns, abrupt disturbances, and asymmetric upstream-downstream propagation, making long-horizon forecasting challenging. To model these complex dependencies, this paper proposes STGCAM, a spatio-temporal graph attention network with frequency-domain enhancement and asymmetric spatial fusion. Gated dilated temporal convolution captures multi-scale temporal evolution, adaptive graph convolution learns directed spatial dependencies without relying on fixed road-network priors, and frequency-domain channel attention extracts periodic components while suppressing high-frequency noise through discrete cosine transform. Experiments on four public PEMS datasets show that STGCAM achieves competitive forecasting accuracy, robustness, and computational efficiency. Ablation studies further verify the effectiveness of the asymmetric graph structure, frequency-domain enhancement, and multi-scale temporal modelling.