Details

Autor(en) / Beteiligte

• Zhu, Yuntian
• Ni, Qingjian

Titel

A period-extracted multi-featured dynamic graph convolution network for traffic demand prediction

Ist Teil von

• Applied intelligence (Dordrecht, Netherlands), 2024, Vol.54 (1), p.722-737

Ort / Verlag

New York: Springer US

Erscheinungsjahr

2024

Quelle

Alma/SFX Local Collection

Beschreibungen/Notizen

• Urban online car-hailing demand prediction poses a significant challenge in developing intelligent transportation systems due to its intricate and dynamic spatio-temporal correlation. Prior research has demonstrated promising outcomes in demand forecasting by employing graph neural networks. However, these studies either solely rely on static prior information or allow the model to independently capture spatial associations. In terms of temporal considerations, effectively modeling both long-term and short-term dependencies remains a crucial factor that significantly impacts overall performance. To tackle these challenges, we propose a novel Period-Extracted Multi-featured Dynamic Graph Convolution Network (PE-MDGCN) for traffic demand prediction. Specifically, our proposed model introduces the Period Dynamic Arrival Learning module and the Static Feature Dynamic Adaptation module, to effectively capture shorter-term relations based on time intervals and arrival connections, as well as the dynamic changes based on static multi-featured graphs. Furthermore, our proposed spatio-temporal multi-graph learning framework leverages a temporal contextual gated mechanism and multi-visual field convolution to efficiently capture global, long-term, and short-term information. By conducting comprehensive experiments on two real-world traffic demand datasets, our model consistently surpasses all baseline models in terms of various evaluation metrics.