A Graph Convolution Network with Temporal Convolution for Long-term Traffic Flow Forecasting

Abstract

High-efficiency and high-precision forecasting of traffic flow are conducive to the improvement of intelligent transportation systems. The traditional traffic flow forecasting models do not take into account the actual topological relationship of the road network. These methods primarily consider the road network to be a regular Eulerian structure or a regular time series. Therefore, for the large and complex traffic network, the forecasting of traffic flow is usually inefficient. In addition, the long-term characteristics of traffic flow are often overlooked. In this paper, we propose a graph convolution network with temporal convolution for long-term traffic flow forecasting, which is distinct from the traditional methods. The proposed model considers the real road topology relationship as a non-Eulerian graph and can also learn long-term traffic characteristics. Our experiments have been verified on two real data sets, and several test indicators have been significantly improved.