| Traffic flow prediction is widely used in the research of intelligent transportation systems,and can be used to solve traffic problems such as traffic congestion.High accuracy traffic flow prediction can ensure the efficient and safe operation of transportation networks.In the current era of transportation big data,using rich and multi-source transportation data and mining its spatial-temporal correlation to accurately predict traffic flow can provide data support for urban transportation network planning and travel plan formulation.This thesis proposes a hierarchical graph convolution network based on spatial-temporal dynamics for highway traffic flow prediction,combining the natural hierarchical structure of the traffic network to model.Firstly,aiming at the natural hierarchical structure contained in the transportation network,a graph clustering algorithm is used to calculate the observed traffic network map data to generate regional map data.Three time interval traffic data,namely,near cycle,daily cycle,and weekly cycle,are selected for data splicing as input data for the hierarchical graph convolution network.Adding a spatial-temporal embedding module allows for simultaneous consideration of spatial and temporal information,enhancing the ability of the model to obtain spatial-temporal correlations.An adaptive graph convolutional network is further added to the model to maintain a better balance between local and global information dissemination,obtain more distinctive node characteristics,and explore the interaction between global spatial adaptability and local dynamics.Then,time gated convolution layer is used to effectively capture long-term dependencies when processing long sequence data,and time graph convolution layer is used to extract spatial-temporal features.After that,a dynamic conversion module based on dynamic graph learning is added to the interaction layer between the region map and the road network map to fuse the region features and the road network features.Subsequently,a Laplace matrix hiding network is introduced to learn node representations from traffic hierarchy data and process the Laplace matrix of the highway hierarchy.Then,feature sampling operations are used to reduce computational complexity and improve the training speed of the model when training on large-scale data sets.Propose a multi headed self attention mechanism to explore the relationships between more nodes,comprehensively utilize causal trend attention mechanism and scaling dot product attention mechanism to reveal the interdependencies between input traffic flow information or features,selectively focus on important data information,and effectively explore the spatial-temporal correlation between any element in the sequence.Later,in order to further explore the internal relationship between adjacent matrix sequences,long-term and short-term memory units were used to learn temporal correlation.Experiments are conducted on two traffic flow datasets,Highway KF and Highway CQ.The STDHGCN model performs traffic flow prediction based on feature fusion,and outputs the final traffic flow prediction results.After comparing and analyzing the prediction results with other baseline models,the results show that the STDHGCN model has a high accuracy in traffic flow prediction at different time intervals.The traffic flow prediction accuracy of the STDHGCN model is improved by about 13% compared to other models,and has a high degree of fit compared to the real traffic flow data collected by the coil sensors.At the same time,ablation experiments are conducted,and the results show that the hierarchical structure of the STDHGCN model can obtain natural hierarchical information in the traffic network,and the use of Laplace matrix to hide the network can more fully obtain the spatial-temporal correlation between nodes in the traffic data.Adding an adaptive adjacency matrix to the STDHGCN model can analyze the impact of certain nodes on other nodes,and introduce more useful new information to the model.The dynamic conversion module fuses regional features and road network features to integrate feature information,improving the traffic flow prediction accuracy of the STDHGCN model.The experimental results verify the effectiveness of setting up hierarchical diagrams,Laplace matrix hidden networks,dynamic conversion modules,and adaptive adjacency matrices to predict traffic flow. |
