Frequency Enhanced Dynamic Graph Convolutional Networks for Traffic Flow Forecasting

With rapid urbanization and technological advancements, increasing population density has posed significant challenges for urban traffic management. Traffic flow forecasting is essential for optimizing traffic signals, regulating traffic flow, and improving overall transportation efficiency. Recent advancements in deep learning, particularly Graph Convolutional Networks (GCNs), offer more effective solutions for traffic forecasting. However, existing approaches face challenges in modeling traffic trends, periodicity, and randomness. To address these issues, this paper proposes a Frequency-enhanced dynamic Graph Convolutional Network (FGCN) that incorporates sequence decomposition and reconstruction to extract periodic and trend components. Additionally, the proposed method integrates the Fourier transform into the loss function to mitigate label autocorrelation and improve long-term forecasting accuracy. Furthermore, this paper employs a dynamic graph convolution method to separately model spatial and temporal dependencies, capturing traffic flow variations more effectively. Experimental results on four open datasets demonstrate that FGCN not only outperforms state-of-the-art methods but also provides interpretable insights into the dynamic spatial relationships between road segments, while significantly improving forecasting accuracy.