Traffic prediction is a crucial task in intelligent
transportation systems, which can help achieve effective management
and optimization of traffic congestion. However, due
to the complexity and uncertainty of traffic systems, accurate
traffic prediction has always been a challenging problem. The
specific challenge of this task is how to model traffic dynamics
along the dimensions of temporal and spatial in a reasonable
manner while respecting and utilizing the spatial and temporal
heterogeneity of traffic data. To address the aforementioned
challenges, this paper proposes a new Transformer-based approach
for traffic prediction. Specifically, to accurately model
complex spatial correlations, we design a spatial Transformer
layer combined with clustering, which reduces computational
complexity and mitigates the risk of over-fitting. To model
dynamic nonlinear temporal correlations, we introduce dilated
attention, which benefits from a global receptive field conducive
to long-term predictions. To validate the effectiveness of our
proposed model, we conduct experiments on four real-world
traffic datasets. The experimental results demonstrate that our
model outperforms state-of-the-art baselines. Furthermore, we
conduct comparative experiments to demonstrate that both the
spatial clustering and dilated attention modules contribute to
the overall improvement of the model’s performance.