Work

DSTF: A Diversified Spatio-Temporal Feature Extraction Model for traffic flow prediction

Year: 2024

Type: article

Abstract: The layout of urban transportation facilities struggles to accommodate the rapidly increasing traffic demand, thereby giving way to a series of traffic issues. Traffic flow prediction serves as an essential foundation for the management and planning of urban transportation systems, holding significant importance in enhancing road transport efficiency and ameliorating urban traffic conditions. However, the complex and multifaceted spatio-temporal dependencies among segments within the transportation network cause significant difficulties in enhancing the accuracy of traffic flow forecasting. To fully explore and integrate diverse spatio-temporal features to improve the prediction accuracy, this paper proposes a Diversified Spatio-Temporal Feature Extraction (DSTF) Model for traffic flow forecasting. Firstly, in the spatial dimension, we construct two views: geographic neighborhood and semantic similarity, then we introduce Geom-GCN to capture local spatial dependencies from these perspectives, effectively addressing the intricate and diverse local spatial correlations within urban road networks. Secondly, we utilize a spatial attention mechanism to capture dynamic global spatial dependencies, enabling the model to learn the overarching spatial structure patterns of urban traffic flows. In the temporal dimension, we employ a gated convolutionally improved Res2Net to capture multiple scales of temporal dependencies, allowing us to analyze traffic flow patterns across different time periods. Additionally, in the spatio-temporal dimension we develop a dynamic spatio-temporal synchronous aggregation module that utilizes spatial and temporal attention to simultaneously extract spatio-temporal features. Finally, we fuse the extracted diverse spatio-temporal features. To evaluate the effectiveness of DSTF, we conduct experiments using four real-world public traffic datasets. The results demonstrate the excellent performance of DSTF.

Source: Neurocomputing

DOI: 10.1016/j.neucom.2024.129280

Authors: Xing Wang, Xiaojun Wang, Faliang Huang, Fumin Zou, Lyuchao Liao, Ruihao Zeng

Institutions: Fujian Normal University, Nanning Normal University, Fujian University of Technology, The University of Sydney

Language: English

Topic: Traffic Prediction and Management Techniques

Subfield: Building and Construction

Field: Engineering

Domain: Physical Sciences

SDGs: Sustainable cities and communities, Industry, innovation and infrastructure

Open Access status: hybrid

Estimated APC paid: $2,470