A Frequency-Aware Spatio-Temporal Network for Traffic Flow Prediction

Abstract

Predicting traffic flow is crucial for transportation management and resource allocation, which has attracted more and more attention from researchers. The traffic flow in a city generally changes over time periods but always exhibits certain periodicity. Previous works focused on modeling spatial and temporal correlations using convolutional and recurrent neural networks respectively. Typically, a method that can effectively absorb more time-interval inputs and integrate more periodic information will achieve better performance. In this paper, we propose a Frequency-aware Spatio-temporal Network (FASTNet) for traffic flow prediction. In addition to modeling the spatio-temporal correlations, we dynamically filter the inputs to explicitly incorporate frequency information for traffic prediction. By applying Discrete Fourier Transform (DFT) on traffic flow, we obtain the spectrum of traffic flow sequence which reflects certain travel patterns of passengers. We then adopt a frequency-based filtering mechanism to filter the traffic flow series based on the explored spectrum information. To utilize the filtered tensor, a 3D convolutional network is designed to extract the spatio-temporal features automatically. Inspired by the frequency spectrum of traffic flows, this spatio-temporal convolutional network has various kernels with different sizes on temporal dimension, which models the temporal correlations with multi-scale frequencies. The final prediction layer summarizes the spatio-temporal features extracted by the spatio-temporal convolutional network. Our model outperforms the state-of-the-art methods through extensive experiments on three real datasets for citywide traffic flow prediction.