Dynamic Spatiotemporal Feature Extraction Integrating Temporal Information and Fluctuation Features For Traffic Prediction

Short-term traffic flow prediction reflects changes in the traffic system, aiding in signal control, route planning, and congestion reduction. However, existing methods often overlook fluctuation features and lack sufficient integration of temporal position information, limiting spatiotemporal modeling. To address this, we propose the dynamic spatiotemporal feature extraction model (DSTFE-TIFF). The DSTFE-TIFF model encodes temporal location using trigonometric functions and extracts fluctuation features via an exponentially decaying weighted method with a sliding window. A dynamic relational module enhances adaptability by integrating these features into traffic data. Additionally, a spatiotemporal attention fusion module uses temporal additive attention to capture local fluctuations and spatial attention to link geographically distant yet similar nodes. Finally, an adaptive fusion method dynamically adjusts the focus on temporal and spatial features. Experimental results on PeMSD4, PeMSD8, and METR-LA datasets show that DSTFE-TIFF outperforms baseline models in RMSE, MAE, and MAPE, confirming its effectiveness.