基于动态图的时空解耦交通流预测模型OA

Traffic flow prediction is beneficial for traffic management and scheduling,and it is an important component of Intelligent Transportation Systems(ITS).This paper proposed a spatio-temporally decoupled traffic flow prediction model based on dynamic graphs to address the problems of existing methods,which often fail to effectively address abnormal signal disturbances and inadequately capture spatial features.The proposed model adopted a spatio-temporal decoupling framework,which separately models the periodic patterns of time series and the spatio-temporal dependency features.Considering that traffic flow data often contains missing values and anomalies,a filtering gate mechanism was designed to reduce the impact of small-scale abnormal signals on temporal prediction.In order to capture the temporal dependency of traffic flow,a global attention mechanism was introduced to extract periodic global signals.To model the complex spatio-temporal dependencies in traffic flow,a Spatio-temporal Local Graph Convolutional Network(STLGCN)was constructed based on dynamic graphs.The dynamic graph constructed in this study not only incorporates daily and weekly temporal embeddings to provide contextual time information but also introduces target and source node embeddings to distinguish the directionality of information propagation and aggregation between nodes.This design enables a more fine-grained characterization of the time-varying dynamic dependencies within the traffic network.Experiments were conducted on four real-world traffic flow datasets.The results show that the model outperformed baseline models,across all three metrics:MAE,RMSE,and MAPE.On the larger-node PEMS-BAY dataset for long-term prediction,the model achieved reductions of 6.78%in MAE,8.57%in MAPE,and 7.43%in RMSE compared with the previous state-of-the-art model.Furthermore,ablation experiments confirmed the contributions of the filtering gate,the global attention module,and the dynamic graph modeling strategy to the model's performance.This study provides an effective new approach for traffic flow prediction while offering significant technical support for the development and optimization of intelligent transportation systems.