Asymmetric long-term traffic prediction with graph multi-attention network

Abstract

Accurate and reliable traffic prediction is one of the challenges in the fields of traffic operation and management. In order to deal with non-linearity and spatio-temporal characteristics of the traffic, much research has been conducted on traffic prediction by using deep learning-based models recently. However, most models have been focused on short-term prediction, which is forecasting within 1-hour. Long-term traffic prediction can be applied in making earlier and efficient decisions for traffic management as well as individual travel. A graph neural network, as one of the deep learning methods, is suitable to predict traffic by considering the connectivity and globality of real road networks. In this study, we propose Asymmetric Long-Term Graph Multi-Attention Network (ALT-GMAN) based on the GMAN model that uses graph neural networks and a multi-attention algorithm. As an extension of the existing model, the proposed model can predict traffic in both short-term and long-term perspectives. In addition, new spatial and temporal embedding techniques are applied in order to consider the effect of real traffic situations and traffic flow theory such as backpropagation of congestion. ALT-GMAN is tested in six months amount of PeMS Bay highway area and in two months amount of Gangnam Yeoksam urban area. MAPE for 3-hour and 6-hour prediction in PeMS Bay is 5.53% and 6.05% respectively. In addition, the prediction of urban areas, where commuting time forecast is important, shows more than six times the accuracy of the existing prediction method. Our proposed model outperforms the existing model in both short-term and long-term traffic speed prediction.