Spatial-Temporal Adaptive Graph Convolution Network with Channel Attention

The paper introduces a novel STACGCN model for enhancing traffic speed prediction in urban traffic management and planning, addressing the limitations of traditional models in handling complex spatial-temporal dependencies. By integrating time-series and graph-structured data features, STACGCN efficiently predicts dynamic traffic speeds. It features a parallel time-convolution module combining gated recurrent units (GRUs) and a channel attention mechanism for capturing both long and short-term dependencies and highlighting crucial temporal information. Spatially, an adaptive dynamic graph learning method is designed to adjust the traffic network graph in real time, reflecting actual traffic conditions and extracting accurate spatial features. A spatial-temporal feature aggregation module further consolidates these features for improved prediction accuracy. Tests on urban traffic datasets show STACGCN superior accuracy and generalization capabilities over baseline models.