Orthogonal Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting

Abstract

It is vital to forecast traffic flow in circumstances of large population size since accurate prediction not only enhances road traffic efficiency but also boosts reasonable traffic plans. However, the majority of prediction methods suffer from the problems of structured 2D or 3D grid data and complex spatial-temporal relationships in traffic data. To deal with the above issues, this paper proposes a novel orthogonal spatial-temporal graph convolutional network (OSTGCN) to forecast traffic flow, which allows direct inputs of graph-based traffic networks. The proposed OSTGCN model consists of two major parts, i.e., i) an orthogonal spatial-temporal block to capture long-range spatial-temporal dependency from traffic data, which can be further strengthened by multi-orthogonality fusion; and ii) a graph convolution block based on a non-binary adjacency matrix to capture spatial patterns among neighbor nodes. Extensive experiments on real-world benchmark datasets demonstrate that, OSTGCN can improve the prediction performance of baselines by at most 6.42% and guarantee an acceptable accuracy with a long prediction interval.