.png){kind=icon}
.png){kind=icon}
Return
Adaptive spatial-temporal dependence graph convolution neural network for traffic flow prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127564.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The real-time and accurate prediction of road network traffic flow provides information on future road condition for the intelligent transportation system (ITS). Due to the complex network topology and real traffic conditions, extracting the temporal and spatial correlations are challenging, which is essential for achieving high accuracy prediction. In general, the traffic flow correlation will be affected by the distance between nodes, the traffic conditions, the flow propagation process, etc. This paper proposes a Spatial-Temporal Graph Model (STGM) for traffic flow prediction. In view of the spatial characteristics, an improved Graph Convolution Network (GCN) model is developed, which incorporates a reconstruction of the adjacency matrix and a spatial attention mechanism. In the GCN module, we design and incorporate four different graphs to the adjacency matrices to explicitly capture the effects of node distance, flow dispersion, similarity of travel speed and time occupancy. Besides, the spatial attention mechanism is applied to extract the network traffic dynamic spatial features. Furthermore, a Long-Short Term Memory (LSTM) model that integrates the temporal attention mechanism is introduced to analyze the time characteristics. By integrating the GCN and LSTM modules, the proposed STGM is able to capture the spatial-temporal characteristics of road network traffic flow more effectively. Experimental tests are conducted on two different datasets, i.e., the US highway dataset PeMSD8 and China city network dataset Xuancheng. Numerical results show that compared with a series baseline models, our proposed STGM achieve higher flow prediction accuracy.
Keywords:
Traffic flow prediction
Enhanced adjacency matrix
Improved graph convolution network
Long short-term memory
Attention mechanism
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W