Prior Knowledge and Data-Driven Based Long- and Short-Term Fusion Network for Traffic Forecasting

Abstract

Urban traffic flow prediction has remained a chal-lenging topic in the intelligent transportation system, due to the complicated spatial-temporal dependency and essential uncer-tainty brought about by the complex road network and dynamic traffic conditions. However, existing methods either rely too much on prior knowledge or the data itself when modeling spatial-temporal dependency and few researchers consider them in combination. In this paper, a long- and short-term fusion net-work for traffic forecasting, which can comprehensively capture the complex spatial and temporal dependency based on prior knowledge and data-driven, is proposed. In particular, in the perspective cross-region spatial dependency considering local and global influences in road networks, a dynamic weighted graph is constructed by finding the spatial and semantic neighborhoods of traffic nodes based on traffic networks and the similarities between traffic flow data on different roads. Besides, to integrate cross-time information, a fusion network that considers long term and short term is applied, and in the temporal layer, the temporal trend module and implicit temporal dependency module are combined to capture the temporal transitivity of traffic flow and implicit dependencies between time point pairs. The experiment results of our proposed model outperform the state-of-the-art baselines.