Spatial-temporal Structures of Deep Learning Models for Traffic Flow Forecasting: A Survey

Abstract

Traffic forecasting is important for the success of intelligent transportation systems. In recent years, deep learning methods, such as convolution neural networks, recurrent neural networks and graph neural networks are introduced to model the spatial and temporal dependencies of the traffic data and have achieved state-of-the-art performance. In this survey, the traffic flow forecasting models based on deep learning are summarized from the perspective of spatial and temporal structure design. Specifically, the existing models can be divided into combinatorial and integrative structures, where the spatial and temporal submodules are considered stepwise with the combinatorial mode but considered comprehensively as a whole with the integrative mode. The functions and structures of each submodule are described and summarized in detail, and the combined mode of the submodules is analyzed as well. On the other hand, the integrative pattern is discussed with two model design paradigms. Furthermore, this paper summarizes the open data sets and source code of the surveyed papers to help upcoming researchers. Finally, the challenges and prospect research directions are discussed thoroughly so as to inspire for more accurate and efficient models development.