Short-Term Traffic Flow Prediction Based on Graph Convolutional Network Embedded LSTM

Short-term traffic flow prediction, which is useful to improve traffic congestion and road efficiency, has been a hot issue in the field of transportation. However, only considering Euclidean space, conventional methods are always unable to make good use of the spatial-temporal correlation of traffic flow data which is usually a topological structure. In this paper, a deep learning model, GCN-LSTM (graph convolutional network-LSTM), was proposed with encoder and decoder structure. GCN-LSTM will simultaneously capture the spatial and temporal characteristic of traffic flow by embedding GCN into the structure of LSTM. Training with the traffic flow data of previous T moments and adjacent section, GCN-LSTM effectively perform short-term traffic flow prediction. Experiments on real data demonstrate that our method, considering both of spatial and temporal features, has a more powerful representation ability and higher prediction accuracy compared with LSTM.