Traffic Flow Prediction Model Based on Attention Spatiotemporal Graph Convolutional Network

Abstract

In order to fully capture the hidden dynamic spatiotemporal correlations in traffic flow data and improve traffic flow prediction accuracy, a multi-head attention spatiotemporal graph convolutional network model (ASGCN) is proposed. Firstly, the multi-head attention mechanism is used to automatically assign attention weights to the traffic sensing nodes in the road network, achieving adaptive matching of weight values for different neighboring nodes and obtaining spatial correlations. Secondly, a spatiotemporal convolutional network structure with gate and attention mechanisms is used to extract time series correlations, and residual block connections are used to improve the model's generalization ability. Finally, weekly, daily, and adjacent time series data are respectively extracted, and input into three parallel spatiotemporal components to explore time periodicity correlations between different time windows. The final traffic flow prediction results are obtained by fully connected layers. The proposed method is evaluated using the high-speed highway traffic datasets KM-Data and CQ-Data for 15-min, 30-min, 45-min, and 60-min traffic flow prediction experiments. The experimental results show that the ASGCN model has better modeling ability than existing baseline models for both short-term and long-term traffic flow prediction.