A Self-Attention Enhanced Hypergraph Convolution Network for Traffic Speed Forecasting

Accurate traffic speed prediction is important in modern society for its effectiveness in route navigation, estimated time of arrival calculations and other practical applications. As the road network is complicated, traffic speed exhibits high-order correlations among regions, namely many-to-many spatial correlations, while also displaying long-term temporal dependencies. However, existing studies have not effectively modelled these characteristics. In this context, this study proposes a self-attention enhanced hypergraph convolution network (SE-HCN) for accurate speed prediction. The proposed SE-HCN consists of four modules. Specifically, we design a relation extraction module, which can obtain the similarity of road sections from geographical information and clustering. Subsequently, the model contains a spatial correlation hypergraph convolutional module and a long-term temporal dependencies transformer module to capture spatio-temporal features comprehensively. Two public real-world datasets - PeMSBAY and PeMSD7-M - were tested to validate the model's performance, and the result demonstrates that our approach achieved state-of-the-art performance.