Graph Attention LSTM Network: A New Model for Traffic Flow Forecasting

Abstract

For the road networks containing multiple intersections and links, the traffic flow forecasting is essentially a time series forecasting problem on graphs. The task is challenging due to (1) complex spatiotemporal dependence among traffic flows of the whole road network and (2) sharp non-linearity and dynamic nature under different conditions. In this paper, by extending the LSTM to have graph attention structure in both the input-to-state and state-to-state transitions, we propose the Graph Attention LSTM Network (GAT-LSTM) and use it to build an end-to-end trainable encoder-forecaster model to solve the multi-link traffic flow forecasting problem. Experiment results show that our GAT-LSTM network could capture spatiotemporal correlations better and has achieved improvement of 15% - 16% over state-of-the-art baseline.