A Spatiotemporal Graph Convolutional Network for Vehicle Speed Prediction

Accurate and real-time vehicle speed prediction is of significant importance for vehicle trajectory forecasting, motion planning, and energy management. However, in real-world traffic scenarios, due to the fact that a vehicle’s future movement is a result of its interaction with surrounding vehicles, and the interactions between vehicles dynamically change over time, traditional speed prediction models struggle to handle these complex spatiotemporal data. Addressing the intricate spatiotemporal correlations in vehicle motion, this paper proposes a spatiotemporal graph convolutional network. This network combines graph convolutional network (GCN) and long short-term memory network (LSTM) for vehicle speed prediction. The GCN extracts spatial interaction features among vehicles, while the LSTM captures temporal correlation features among historical speed data. The proposed model is trained and tested on the publicly available NGSIM US-101 dataset. Simulation experiments demonstrate that, compared to other baseline models, the proposed model can extract spatiotemporal features from traffic data with minimal prediction errors. When the prediction horizons are 1 second, 3 seconds, and 5 seconds, the root mean square errors are only 2.0219,2.1574, and 3.2149, respectively. Visualizations indicate that the predicted results reflect the variation characteristics of vehicle speed.