Dynamic spatial–temporal graph convolutional recurrent networks for traffic flow forecasting

Traffic flow forecasting is crucial for making appropriate route guidance and vehicle scheduling schemes in intelligent transportation systems. However, recent graph-based methods leverage pre-defined static graphs to describe the spatial–temporal characteristic of road networks. The interactions of two road segments are changeable under the influence of natural environmental and socioeconomic factors, while these methods are not sufficient to capture the complicated dynamic correlations of different nodes. To address these problems, we propose a novel dynamic graph-based deep learning framework with dynamic graph recurrent network for traffic flow forecasting, called Dynamic Spatial–temporal Graph Recurrent Neural Networks. In this framework, a novel dynamic graph generator is designed to obtain the dynamic representation of nodes, which employs multi-head attention network and dynamic node embedding to capture hidden spatial dependency more effectively. To infer the edge status of dynamic graph at different times, the generated dynamic graph is trained as special time series data via dynamic graph recurrent neural network for downstream time-series prediction. In contrast to methods straightforwardly concatenating static graphs and dynamic graphs, a novel fusion framework integrates two-channel convolutional networks with penalty terms and a gate fusion layer to extract dynamic spatial dependency from multiple graphs for improving forecasting accuracy and reducing computational consumption. Experiments on three real-world datasets are carried out to evaluate the superior performance of our model. Compared with previous state-of-the-art baselines, the proposed method performs much better with 10%–26% improvements on three datasets. The results also indicate that our model is robust against emergent traffic situations.