Dynamic Graph Adaptive Spatio-Temporal Network for Smart Traffic Systems

Currently, the increasing complexity of urban traffic systems demands intelligent solutions capable of accurate and real-time traffic flow predictions. The Graph Neural Networks (GNNs) have shown strong potential by modeling spatial dependencies across road networks. However, the existing GNN-based models lack adaptability to dynamic traffic conditions, which limits their effectiveness in real-time scenarios. Hence, this study proposes a Dynamic Graph Adaptive Spatiotemporal Network (DGA-STNet) for real-time adaptability. Initially, the input data is collected from PeMSo4 dataset which contains large-scale urban traffic data. Then, the data were preprocessed with linear interpolation, z-score standardization, sliding window, and chronological order to handle missing data, normalize features, transform continuous time-series data, and avoid data leakage. Subsequently, traffic data are embedded into a high-dimensional feature space and enriched through adaptive graph construction by combining the static topology and dynamic feature similarity. Next, multi-scale gated temporal convolution and adaptive spatial attention were applied to extract the complex ST patterns. Finally, a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1 \times 1$</tex> convolutional layer is employed to produce multi-step traffic flow predictions. The proposed DGA-STNet achieved better results in terms of the mean square error (9.87) than the existing ST Gated Graph Attention network (STGGAN) model.