Abstract: In the field of modern intelligent transportation systems, accurate prediction of traffic speed is of great significance to alleviate traffic congestion, improve road safety, and optimize traffic management. To improve the performance of existing traffic speed prediction models in medium and long-term prediction tasks, this paper proposes an adaptive residual dynamic fusion graph attention network for traffic speed prediction, in which the bimodal graph architecture can capture static topology and dynamic spatio-temporal correlation features of the road network through parallel processing and dynamic fusion of adaptive and dynamic adjacency matrices. Applying gated temporal convolution to realize the feature screening, and using multi-head attention mechanism to enhance the spatio-temporal feature expression ability, designing dynamic feature fusion unit, retaining static topological information through residual connection, and combining cross-layer multi-scale feature fusion to avoid feature degradation. The experimental results show that the root mean square error of this model is reduced by 22.5% and 22.6% compared with Graph WaveNet in the 60 min prediction task for the METR-LA and PEMS-BAY datasets, respectively. The model can adapt to the changes of the traffic state in real time, provide accurate speed prediction for the traffic management department, and assist in the congestion diversion, dynamic path planning, and emergency response. The model has high practical application value.

Key words: city traffic, traffic speed prediction, bimodal map fusion, dynamic graph convolutional network, residual connection, attention mechanism