Multi-factor embedding GNN-based traffic flow prediction considering intersection similarity

Existing studies on traffic flow prediction primarily rely on on-board devices to collect vehicle trajectory data, which can potentially infringe upon the privacy of users and limit the applicability of the method. Additionally, traffic flow prediction remains challenging due to the complex spatial and temporal dependencies within real-world traffic networks. To address these limitations, this paper introduces a framework for analyzing discrete vehicle trajectory data at urban intersections. By incorporating various external physical factors into traffic flow prediction, this framework derives embedding vectors from vehicle trajectory sequences and road network topology, modeling their spatio-temporal dependencies using Skip-Gram and GraphSAGE, respectively. Additionally, the intersection similarity is introduced to capture and integrate traffic flow patterns between the target intersection and similar intersections. A Spatio-Temporal Graph Convolutional Neural Network (ST-GCN) algorithm, which combines Graph Convolutional Networks (GCN) with Long Short-Term Memory (LSTM), is developed to achieve precise traffic flow prediction. Extensive experiments on a real-world traffic flow dataset from Qingdao, China, validate that the proposed method outperforms state-of-the-art baseline methods.