Dual-perception spatiotemporal graph attention for traffic flow prediction fusing recent and periodic features

ObjectiveTraffic flow data exhibit high nonlinearity, dynamic spatiotemporal correlations, and significant periodic patterns. Existing methods often fail to sufficiently mine the implicit long-term dependencies and capture the dynamic spatial associations, which limits the prediction accuracy in complex road conditions. To address these challenges, a dual-perception spatiotemporal graph attention (DPST-GAT) network is proposed, which integrates both recent and periodic perception features.MethodThe model features a dual-channel framework. The recent perception module incorporates the temporal convolutional network with self-attention mechanism (TCN-SAM), utilizing positional encoding to capture temporal dependencies. This module further integrates an improved distance-based graph attention network, where spatial distance thresholds are applied to filter neighbor nodes, thereby enhancing the local spatial feature extraction. The periodic perception module introduces a spatiotemporal similarity aggregation algorithm, employing the deep embedding clustering to identify periodic patterns. This algorithm recognizes nodes with similar variation trends that are often physically non-adjacent, effectively capturing global long-term dependencies.ResultComparative experiments were conducted on the datasets, i.e., PeMS03, PeMS04, PeMS07, and PeMS08. DPST-GAT demonstrates superior performance in both short-term and long-term forecasting. Compared with eight baseline models, e.g., SVR, LSTM and STFGNN, the proposed model outperformed the best-performing STFGNN with the RMSE decreasing by 6.30%, 3.16%, 5.66%, and 7.01% respectively.ConclusionThe integration of self-attention mechanism, spatial distance constraints and clustering strategy effectively strengthens the modeling of dynamic spatiotemporal heterogeneity. The model significantly improves the prediction accuracy and robustness in complex intelligent transport environments.