A traffic accident prediction model based on a Spatial-Attention Temporal Graph Convolutional Network

Traffic accidents often result in significant casualties and economic losses. Accurate and timely prediction of traffic accidents is crucial for public safety and reducing economic impact. However, predicting traffic accidents is challenging due to their complex causes, involving spatial, temporal dynamics, and external influences within heterogeneous traffic-related data. To address these issues, this paper proposes a Spatial-Attention Temporal Graph Convolutional Network (SATGCN) for predicting traffic accidents. In the process of predicting traffic accidents, at the first stage, the spatial convolution layer leverages GCN and attention mechanisms to extract spatial features. Then, the temporal convolution layer extracts spatiotemporal features through stacked spatiotemporal modules, followed by dimensionality reduction via average pooling. Finally, the processed spatial, temporal, and external features are fused and passed through a fully connected layer to generate the prediction result. The simulation demonstrated that the model performed better than the baseline on the Beijing Traffic Accident Dataset, with its precision improved by 5% compared to DSTGCN.