Graph Convolutional Recurrent Neural Networks: A Deep Learning Framework for Traffic Prediction

Road traffic state prediction is a challenging task for urban traffic control and guidance due to the complicated spatial dependencies on the roadway network and the time-varying traffic flow data. In this work, a novel traffic flow prediction method named the graph convolutional recurrent neural network (GCRNN) is proposed to tackle this challenge. First, we address the problem on a graph and build the model with graph embedding techniques. Second, the proposed model employs the GCN model to learn the interactions of the roadways to capture the spatial dependence and uses the long short-term memory (LSTM) neural network (NN) to learn dynamic changes of traffic data to capture temporal dependence. An experiment is conducted on a Hangzhou transportation network with several typical intersections under the Sydney coordinated adaptive traffic system (SCATS), the results of which indicate that our model yields excellent performance in terms of different prediction error measures.