Unifying spatiotemporal and frequential attention for traffic prediction

Abstract Urban traffic flow prediction is an essential task within intelligent transportation systems, and numerous methodologies have been proposed to address it. However, most existing approaches focus on historical traffic information in terms of spatial and temporal aggregation, neglecting the implied spectral analysis of traffic time series. In this paper, we introduce the traffic flow in the frequency domain and, in conjunction with attention mechanisms, comprehensively learn the hidden correlations between spatial, temporal, and frequential. By deeply learning the spatial graph topological correlations of traffic flow, and using spectral analysis, fusing time series and implied periodic correlations in the temporal and frequential, we have constructed an innovative traffic prediction network model known as the Spatial Temproal-Frequential Attention Network (STFAN). The core of this network is the application of attention mechanisms to project the hidden states of traffic features in the current spatial, temporal, and frequential onto future hidden states, thereby comprehensively learning the hidden relationships of each dimension to future states and achieving the prediction of future traffic flow. To validate the performance of the proposed model, experiments were conducted on two public datasets from the California Department of Transportation (PeMS04 and PeMS08). The results demonstrate that the proposed model outperforms existing baseline models in terms of predictive accuracy, especially in medium and long-term traffic flow forecasting. Additionally, the ablation study confirmed the influence of frequency domain characteristics of traffic flow on future traffic states, thus proving the theoretical and practical effectiveness of the model.