A Multi-step Traffic Speed Forecasting Model Based on Graph Convolutional LSTM

Abstract

Traffic speed is an indicator reflecting the traffic state and it is important for traffic control and optimization. Traffic speed forecasting is a typical application of spatial-temporal forecasting problems. However, it is very difficult to accurately predict the traffic speed especially when we need a long-term forecasting result. And how to capture the spatial features is also a huge challenge. In this paper, a sequence to sequence (Seq2Seq) model based on Graph Convolutional LSTM (GC-LSTM) was proposed, which is an end-to-end deep learning model that uses the network topology and historical speed data. The Graph Convolutional Networks (GCN) is used to deal with spatial relationship features while LSTM for temporal relationship features. The Seq2Seq model is a structure that can predict the multi-step traffic speed. Then an experiment on a real-world dataset shows that our model can get a good result for the long-term traffic forecasting problem.