Song, Yaofeng (Nanyang Technological University), Ma, Yifan (Nanyang Technological University), Chu, Zhiyao (Bohai University), Luo, Ruikang (School of Electrical and Electronic Engineering, Nanyang Technol), Su, Rong (Nanyang Technological University)

GMDNet: A Graph-Dynamics-Aware Multi-Scale Decoupling Network for Spatio-Temporal Traffic Flow Prediction

Scheduled for presentation during the Regular Session "S08c-Intelligent Modeling and Prediction of Traffic Dynamics" (WE-LA-T8), Wednesday, November 19, 2025, 16:20−16:40, Coolangata 2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 19, 2025

Abstract

Traffic flow prediction continues to face challenges in complex scenarios where dynamic spatial correlations coexist with multiscale temporal fluctuations. Existing methods struggle to effectively model transient spatial dependencies triggered by emergencies such as traffic accidents: traditional static graph convolutional networks can only capture steady-state correlations between road nodes based on physical connectivity, but fail to characterize instantaneous spatial correlation reconfigurations caused by dynamic events like real-time navigation rerouting or temporary traffic control. Simultaneously, traditional signal decomposition techniques (e.g., Fourier transform, wavelet decomposition) relying on predefined basis functions cannot adaptively decouple cross-scale coupling features between abrupt fluctuations and periodic patterns in nonstationary traffic signals. To address these challenges, this paper proposes the GMDNet model: In the dynamic graph convolutional module, event-driven adaptive adjacency matrices are generated based on real-time traffic states, breaking the path dependence of traditional static adjacency matrices on physical road network structures and effectively modeling instantaneous spatial correlation reconfigurations induced by real-time navigation path planning. The Empirical Mode Decomposition (EMD) technique adaptively decomposes traffic pattern signals into multiscale Intrinsic Mode Functions (IMFs), where high-frequency components precisely capture minute-level abrupt fluctuations, while low-frequency components resolve macroscopic trends such as daily periodic patterns. For the decomposed multiscale features, the model constructs parallel processing branches for better temporal pattern extraction. Experiments are conducted on 2 real-wor