General Decoupled Graph Convolutional Recurrent Network for Traffic Prediction

Abstract

Accurate traffic prediction plays a pivotal role in the development of intelligent transportation systems. However, traffic data often exhibits multi-scale temporal dependencies, spatiotemporal heterogeneity, and highly nonlinear characteristics, which pose significant challenges to existing prediction models. To address these challenges, we propose a General Decoupled Graph Convolutional Recurrent Network (GDGCRN). At the core of GDGCRN is its ability to handle multi-scale temporal dynamics. By learning temporal embeddings for daily and weekly cycles, the model adaptively captures short-term fluctuations as well as long-term patterns. On the spatial side, GDGCRN incorporates a sensor-specific mechanism that adapts to the unique behavior of individual traffic sensors, allowing the model to capture complex spatial dependencies across diverse and heterogeneous traffic environments. Moreover, GDGCRN introduces a signal decoupling mechanism that separates steady-state and non-steady-state signals, enabling the model to robustly handle both predictable traffic trends and sudden disruptions. Through comprehensive experiments on multiple benchmark datasets, including traffic flow prediction on PeMS04 and PeMS08, demand prediction on the NYCBike dataset, and speed prediction on PeMSD7(M), we demonstrate that GDGCRN achieves highly competitive performance across these datasets. Notably, on the PeMS08 dataset, GDGCRN reduces the MAE metric by 9.92% compared to the MVSTT model.

Topics & Concepts