基于自适应门控图卷积网络的交通流量预测OA

Traffic flow data in road networks generally have noise interference,information loss,and complex spatio-temporal dynamic evolution characteristics,and the existing spatio-temporal graph convolutional models based on a fixed graph structure are prone to problems such as over-smoothing when characterizing node features,resulting in limited prediction accuracy and low computational efficiency.To address the above challenges,an Adaptive Gated Spatio-Temporal Graph Convolutional Network(AG-STGCN)model was proposed,which adopted a multi-module collaborative design strategy.The feature fusion module used convolution operations for multi-scale smoothing,and combined them with the original feature reconstruction and splicing to effectively suppress noise and missing data in the input.The causal convolution module mined the nonlinear evolution patterns of traffic flow within each node over time through temporal convolution operations.The unidirectional graph DropEdge mechanism was introduced when constructing the dynamic sparse road network to sparsify the topological connections and alleviate the over-smoothing problem caused by multi-layer graph convolution.The adaptive gated graph convolution layer combined the auto-regressive moving average(ARMA)filter with an attention mechanism to dynamically adjust the node feature update weights to enhance spatial feature extraction.The output module achieved multi-scale feature fusion through hierarchical attention aggregation to generate the final traffic flow prediction results.Extensive experimental validation on four benchmark datasets,PEMS03,PEMS04,PEMS07 and PEMS08,shows that the proposed AG-STGCN model outperforms seven baseline models,such as STFGNN and HSTGCNT,in the three key metrics,MAE,RMSE and MAPE.Especially on the PEMS07 dataset with larger node numbers,the MAE,MAPE and RMSE were reduced by 15.6％,42.9％and 4.7％,respectively,compared with the baseline model HSTGCNT.The average training time of the model on the four datasets was reduced by about 50％compared with that of HSTGCNT.These results reflect the advantages of the proposed model in terms of generalization ability and computational efficiency in complex road networks,thereby providing effective technical support for traffic flow prediction in intelligent transportation systems.Its modular design also provides a useful reference for related spatio-temporal prediction research.