Asymptotic mixed graph convolutional with trend-aware self-attention networks for traffic flow forecasting

Accurate traffic flow forecasting plays a critical role in enhancing the efficiency and safety of transportation systems. Nevertheless, this task remains challenging due to the complex spatio-temporal correlations embedded in traffic data. Existing adaptive-graph-based methods usually learn graph structures during training, making them insufficiently responsive to testing-phase traffic states. Consequently, their forecasting performance may deteriorate under unexpected fluctuations and irregular traffic variations. To overcome this limitation, this paper proposes a novel Asymptotic Mixed Graph Convolutional with Trend-Aware Self-Attention Network (AMG-TSA). By learning trend similarities among traffic nodes, AMG-TSA constructs an asymptotic adjacency matrix to support both asymptotic structure-aware graph convolution and asymptotic dynamic graph convolution, enabling synchronized modeling of dynamic spatial dependencies. Meanwhile, a trend-aware self-attention module is introduced to capture local contextual patterns and nonlinear temporal dependencies, while dilated causal convolution enlarges the temporal receptive field and supports long-range temporal dependency modeling. Experiments on two real-world traffic datasets show that AMG-TSA consistently outperforms representative baselines, demonstrating its effectiveness in traffic flow forecasting.