Global spatio‐temporal dynamic capturing network‐based traffic flow prediction

Abstract

Capturing the complex spatio‐temporal relationships of traffic roads is essential to accurately predict traffic flow data. Traditional models typically collect spatial and temporal relationships and increase the complexity of the model by considering connected and unconnected roads. However, global road networks are dynamic and hidden connectivity relationships generally undergo variations over time. A deterministic single‐connection correlation inevitably limits the learning capability of the model. In this paper, the authors propose a global spatio‐temporal dynamic capturing network (GSTDCN) for traffic flow prediction. First, the global encoding module based on the attention mechanism is set up to describe the dynamic spatio‐temporal relationships. It is shown that GSTDCN can learn the hidden node information by spatial correlation at different times. Meanwhile, an effective temporal prediction module is constructed, which facilitates the data augmentation and improves the prediction results of GSTDCN. The model is experimented on four public transportation datasets, and the results show that the GSTDCN outperforms the state‐of‐the‐art baseline.

HD‐Net: A hybrid dynamic spatio‐temporal network for traffic flow prediction

“…Thus, the attention mechanism faces limitations in integrating the correlations between the local and global features. To achieve a more accurate and comprehensive prediction model, it may be necessary to combine other methods, or employ different modelling techniques [36,45]. To address these challenges, a novel hybrid dynamic traffic flow prediction model, namely HD-Net, has been proposed in this paper.…”

“…}at the T consecutive historical time steps from the global respective. Herein, the input of the SA (45).…”