STAE-TA: Spatio-Temporal Frequency Adaptive Embedding with Multi-Scal Trend Attention
Jing Tian, Sheng Su, Mingyang Wang, Chenggang Wang · 2024
Time series forecasting plays a crucial role in facilitating decision-making processes, particularly in areas such as traffic flow analysis, situational forecasting, and disease prediction. Traffic prediction has grown in importance with the expansion of road networks and the escalating demand for effective trafficmanagement strategies that leverage historical data. Recently, spatio-temporal graphical neural networks (STGNNs) and transformer-based models have garnered significant attention owing to their enhanced performance. However, the complex network structure did not lead to significant performance improvements, and researchers began to focus on the accurate representation of data. Although standard attention mechanisms is good in extracting crucial information, they often overlook the modeling of local dynamic changes and spatial relationships. To address these limitations, we introduce a novel model, STAE-TA, which integrates a spatio-temporal frequency adaptive embedding layer with a multi-scale trend-aware attention mechanism. Extensive experimentation with four real datasets demonstrates that STAE-TA surpasses existing models across most metrics, thereby enhancing the accuracy and adaptability of traffic prediction.