Multivariate Long-Term Traffic Forecasting with Graph Convolutional Network and Historical Attention Mechanism

Abstract

Due to the complexity of the traffic system and the constantly changing characteristics of many influencing factors, long-term traffic forecasting is extremely challenging. Many existing methods based on deep learning perform well in short-term prediction, but do not perform well in Long-Term Time Series Forecasting (LTSF) tasks. These existing methods are difficult to capture the dependencies of long-term temporal sequences. To overcome these limitations, this paper introduces a new graph neural network architecture for spatial-temporal graph modeling. By using simple graph convolutional networks and developing novel spatial-temporal adaptive dependency matrices, our model can capture the hidden spatial-temporal internal dependency in the data. At the same time, we add external dependency to the model. We utilize the periodicity between long-term time series and historical data and introduce a Historical Attention Mechanism to capture historical dependencies in combination with historical data, which can expand the receptive field of the model from local relationships to historical relationships to help improve the prediction accuracy and avoid the problem of too long sequence and too much useless information caused by taking the entire historical sequence as input. Experimental results on two public traffic datasets, NYC-TLC and England-Highways, demonstrate the superior performance of our method.