Abstract
Intelligent transportation systems heavily rely on forecasting urban traffic flow, and a variety of approaches have been developed for this purpose. However, most current methods focus on exploring spatial and temporal dependencies in historical traffic data, while often overlooking the inherent spectral characteristics hidden in traffic time series. In this paper, we introduce an approach to analyzing traffic flow in the frequency domain. By integrating attention mechanisms, we comprehensively capture the hidden correlations among space, time, and frequency dimensions. By leveraging deep learning to capture spatial correlations in traffic flow and applying spectral analysis to fuse time series data with underlying periodic correlations in both the time and frequency domains, we develop an innovative traffic prediction model called the Space-Time-Frequency Attention Network (STFAN). The core of this network lies in the application of attention mechanisms, which project the hidden states of current traffic features across the space, time, and frequency domains onto future hidden states. This approach enables a comprehensive learning of the relationships between each dimension and the future states, ultimately allowing for accurate predictions of future traffic flow. We carry out experiments on two publicly available datasets from the California Department of Transportation, PeMS04 and PeMS08, to assess the performance of the proposed model. The results demonstrate that the proposed model outperforms existing baseline models in terms of predictive accuracy, particularly for mid- and long-term traffic flow forecasting. Finally, the ablation study confirmed that the frequency domain characteristics of traffic flow significantly influence future traffic conditions, demonstrating the practical effectiveness of the model.