Abstract: Due to the complexity of traffic flow and the insufficient extraction of spatio-temporal features, it is difficult for D2STGNN to capture the dynamic changes of traffic networks, which limits the improvement of prediction accuracy. In this paper, a Bi-EMHGRU model combining an efficient multi-head self-attention mechanism (EMHSA) and a bidirectional gated recurrent unit (BiGRU) is proposed. This model captures the sequential dependencies of both forward and backward timings through BiGRU and dynamically allocates weights to each time step by using the multi-head self-attention mechanism to focus on key sequential features. Meanwhile, a multi-scale time feature extraction module is introduced, which enhances the modeling ability for short-term fluctuations and long-term trends and improves the modeling effect of complex spatio-temporal dynamics. The experimental results show that Bi-EMHGRU performs excellently on the PEMS04 and PEMS08 datasets. The root mean square error value has decreased by approximately 0.55~1.55, the mean absolute error has decreased by approximately 0.89~1.40, and the mean absolute percentage error has decreased by approximately 0.86~1.77 percentage points. It can still maintain stable prediction performance when the prediction step length increases and has strong generalization ability. Compared with the existing benchmark models, Bi-EMHGRU can capture the dynamic spatio-temporal features of traffic flow more effectively and significantly improves the prediction accuracy and robustness.

Key words: traffic flow prediction, multi-head self-attention, Bi-EMHGRU, dynamic spatio-temporal features, multi-scale temporal features