Pyramid Temporal-Spatial Attention Network for Traffic Forecasting

Traffic flow prediction is a crucial component in optimizing highway traffic within intelligent transportation systems. Achieving high-accuracy traffic flow predictions can lead to better management of intelligent transport systems, alleviating traffic congestion and reducing travel costs. However, traffic flows exhibit high spatiotemporal dynamics and nonlinear characteristics. Existing methods primarily focus on the spatiotem-poral features of traffic flows and often overlook the complexity of prediction. To fully explore the implicit spatiotemporal features of traffic flow data and reduce prediction loss, we propose a traffic flow prediction model that integrates a pyramid spatiotemporal attention mechanism. The model employs a dual-channel approach, embedding date and time encoding to capture temporal and periodic information, and combines it with traffic flow graphs to generate enhanced traffic flow data features. Based on this, a two-stage attention mechanism layer containing low-rank temporal attention and low-rank spatial attention is constructed, with a pyramid attention mechanism embedded externally beyond the two-stage attention mechanism layer that includes residual units. This forms a pyramid spatiotemporal attention module. Experiments conducted on public datasets indicate that the predictive performance of our model surpasses certain baseline models in terms of prediction accuracy.