Abstract
Short-term traffic flow prediction is essential for intelligent transportation systems, enabling signal control, route planning, and congestion mitigation. However, existing methods often overlook flow fluctuations and underutilize temporal positional information, limiting spatiotemporal modeling performance. Therefore, this study proposed a dynamic spatiotemporal feature extraction (DSTFE) model that integrated temporal information and fluctuation features (TIFF) for enhanced traffic prediction—DSTFE-TIFF. Specifically, DSTFE-TIFF employed trigonometric time encoding to capture multiscale patterns and an exponentially weighted sliding window to emphasize recent, abrupt changes. These features were then integrated through a time-weighted attention framework, which modeled both local and long-range dependencies to form comprehensive spatiotemporal representations. Experiments on three data sets—the traffic-speed data set in the Los Angeles County road network, traffic-flow data set in the San Francisco Bay area, and traffic-flow data set in California—showed that DSTFE-TIFF achieved state-of-the-art performance, reducing key error metrics (mean absolute error, root mean square error, mean absolute percentage error) by over 22% on average versus baselines, demonstrating its effectiveness and robustness.

This publication has 30 references indexed in Scilit: