Traffic flow prediction with multi-feature spatio-temporal coupling based on peak time embedding

• Published in: The Journal of supercomputing 2024-11, Vol.80 (16), p.23442-23470
• Main Authors: Wei, Siwei, Hu, Dingbo, Wei, Feifei, Liu, Donghua, Wang, Chunzhi
• Format: Article
• Language: English
• Subjects: Compilers; Computer Science; Deep learning; Embedding; Graphical representations; Intelligent transportation systems; Interpreters; Nonlinear dynamics; Peak hour traffic; Peak periods; Prediction models; Processor Architectures; Programming Languages; State-of-the-art reviews; Synchronism; Time synchronization; Traffic flow; Traffic speed; Transportation networks
• ISSN: 0920-8542; 1573-0484
• DOI: 10.1007/s11227-024-06378-1

Traffic flow prediction plays a crucial role in intelligent transportation systems (ITS), offering applications across diverse domains. However, current deep learning models face significant challenges. Real-world traffic conditions, especially during peak hours, exhibit complex spatio-temporal dynamics and intricate nonlinear relationships. Existing studies often overlook variations in traffic flow across different time periods, locations, and scenarios, resulting in prediction models lacking robustness and accuracy across diverse contexts. Furthermore, simplistic models struggle to accurately forecast traffic flow during peak periods, as they typically focus on isolated features such as traffic speed, flow rate, or occupancy rate, neglecting crucial interdependencies with other relevant factors. This paper introduces a novel approach, the peak hour embedding-based multi-feature spatio-temporal coupled traffic flow prediction model (PE-MFSTC), to address these challenges. The PE-MFSTC model incorporates peak time embedding within a multirelational synchronization graph attention network structure. The peak time-based embedding involves mapping daily, weekly, and morning/evening peak periods into low-dimensional time representations, facilitating the extraction of nonlinear spatio-temporal features. The network framework employs a multirelational synchronized graph attention network, integrating multiple traffic features and spatio-temporal sequences for learning. Additionally, a spatio-temporal dynamic fusion module (STDFM) is introduced to model correlations and dynamically adjust node weights, enhancing the model’s sensitivity. Experimental evaluations on four real-world public datasets consistently demonstrate the superior performance of the PE-MFSTC model over seven state-of-the-art deep learning models. These results highlight the efficacy of the proposed model in addressing the complexities of traffic flow prediction across various scenarios.