关键词: 动态图卷积网络, 时空注意力机制, 交通流预测, 残差图卷积, 解耦模块, 时空依赖性

Abstract: Abstract: Aiming at the problem that existing traffic flow prediction models rely on predefined static graph structures and are difficult to capture dynamic spatio-temporal dependencies, a traffic flow prediction model based on spatio-temporal attention and decoupled dynamic residual graph convolution is proposed. The model adaptively generates node association relations through Dynamic Graph Convolutional Recurrent Network, and captures temporal features in combination with gated recurrent unit. The Spatio-Temporal Attention Block is introduced to model dynamic spatial and nonlinear temporal correlations. The decoupling module is utilized to decompose the traffic flow into independent mode representations, and the multi-mode features are processed by the residual graph convolution module. Experiments on the PEMS03, PEMS04, and PEMS08 datasets show that the prediction performance of STAR-DDRGC significantly outperforms that of the existing 18 benchmark models, with an average reduction of MAE, RMSE, and MAPE metrics of 12.7%, 10.3%, and 14.2%, respectively, and a maximum reduction of MAE of 21.33% especially on the PEMS08 dataset. The ablation experiments verify the effectiveness of each module, and the feature enhancement layer and spatio-temporal attention module reduce the MAE by 0.53 and 0.61, respectively.The computational cost analysis reveals that the number of STAR-DDRGC parameters is 527,534, and the training time is reduced by 63.23% compared with that of STGODE, which ensures the prediction accuracy and improves the efficiency at the same time. Dynamic graph generation and multi-module co-design can effectively model the complex spatio-temporal dependencies of traffic flow and provide high-precision and low-latency prediction solutions for intelligent transportation systems.

Key words: dynamic graph convolutional networks, spatio-temporal attention mechanism, traffic flow prediction, residual graph convolution, decoupling module, spatio-temporal dependence