Adaptive Spatial-Temporal Convolution Network for Traffic Forecasting

Zhao Li, Yong Zhang, Zhao Zhang, Xing Wang, Lin Zhu

Abstract: As a vital part of intelligent transportation system (ITS), traffic forecasting is crucial for traffic management and travel planning. However, accurate traffic forecasting is challenging due to complex spatial-temporal correlations among traffic series and heterogeneities among prediction hirizons. Most existing works focus on extracting spatial features by graph convolution network and ignore the heterogeneities of timestamps in long-term forecasting. Moreover, the fixed spatial-temporal correlations extraction pattern is not sensitive to the changes of traffic environments. In this paper, we propose a general and effective framework called adaptive spatial-temporal convolution network (ASTCN) for traffic flow forecasting. ASTCN can capture inherent spatial-temporal correlations adaptively according to specific prediction task without any prior knowledge, and has strong adaptability to different traffic environments. Moreover, each timestamp over prediction horizon can learn its own unique features aggregation pattern, so as to improve the accuracy of long-term and short-term forecasting concurrently. The experimental results on four public highway datasets show that ASTCN outperforms state-of-the-art baselines and achieves optimal accuracy in all prediction horizons for the first time.