.png){kind=icon}
.png){kind=icon}
Return
Decoupled multi-spatio-temporal fusion graph convolutional recurrent network for traffic prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2025.112956.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Precise traffic prediction is essential for building smart city transportation systems. Although significant progress has been made, there are still limitations in capturing complex spatio-temporal relationships. First, current traffic prediction methods generate static graphs that fail to adapt to time-varying traffic conditions, unable to capture how spatial dependencies evolve throughout the day or across the week. Moreover, they often process heterogeneous traffic signals uniformly without distinguishing between steady-state and non-steady-state components. In this work, we propose a Decoupled Multi-spatio-temporal Fusion Graph Convolutional Recurrent Network (DMFGCRN) to address these limitations simultaneously. Firstly, we introduce a dynamic embedding graph learner that integrates real-time traffic signals with calendar-aware temporal patterns, generating unique time-varying adjacency matrices for each time step. Further, we propose a multi-layer architecture that progressively separates steady-state from non-steady-state signals through cascaded convolutional filtering, enabling each layer to refine different signal components. Additionally, our multi-spatio-temporal fusion module combines dynamic graph convolution with bidirectional recurrent processing across multiple refinement layers, where historical context enriches future predictions at multiple abstraction levels. Experimental results on seven real traffic datasets show that DMFGCRN outperforms 24 state-of-the-art methods. Specifically, on the Performance Measurement System District 8 dataset, our model showed improvements of 7.58% in Mean Absolute Error compared to the Multi-spatio-temporal Fusion Graph Recurrent Network model. The source code for DMFGCRN is publicly available at: https://github.com/OvOYu/DMFGCRN .
Journal
IF:
8
Papers:
5.3K
Citations:
3.5W