Research On Traffic Flow Prediction Method Based On Attention Mechanism And Convolutional Neural Network

Since 2022,China’s urbanization rate of permanent residents has exceeded 60%,and with the continuous advancement of urbanization,the number of motor vehicles in stock is also increasing,leading to increasingly severe traffic congestion.Traffic flow prediction is a crucial aspect of intelligent transportation systems,and accurate predictions will provide technical support for traffic management,traffic signal optimization,travel route planning,and other related activities.Most existing prediction methods do not fully consider the spatio-temporal characteristics of traffic flow,such as the need to model spatial correlations by constructing a pre-defined graph of the traffic road network(such as an adjacency matrix),making it difficult for the model to capture the potential spatial dependencies of the traffic road network(such as the mutual influence of one-way and two-way lanes,non-connected roads,etc.),and the accuracy of existing models in long-term traffic flow prediction tasks still needs improvement.To address the above problems,the main research contents of this dissertation are as follows:(1)This dissertation proposes a traffic flow prediction model based on improved graph convolution and spatio-temporal attention mechanism.To address the problem of existing models not fully capturing spatio-temporal correlations and potential spatial dependencies,a space module(GC Block)is constructed by using a graph learning matrix module and graph convolution to extract fixed and potential spatial correlations.A time module(TC Block)is composed of expanded causal time convolution and gating mechanism to extract temporal features.The spatio-temporal attention mechanism(STAtt Block)is combined to dynamically capture spatio-temporal correlations and enhance the accuracy of traffic flow prediction.The proposed model is experimentally verified on two public datasets,PeMSD04 and PeMSD08,and the results demonstrate its effectiveness in traffic flow prediction tasks.(2)To further improve the prediction accuracy and long-term prediction performance of the model,this dissertation proposes a traffic flow prediction model based on dynamic attention and multi-gated time convolutions.The model introduces multi-gated time convolution layers to extract time features from different time ranges.Additionally,the model improves the convolution kernel size and gate output of time convolutions on the basis of the original time gating to achieve accurate prediction of different time pattern tasks and thereby improve the accuracy of long-term prediction.The dynamic attention mechanism is used to extract the spatial relationship between traffic nodes and dynamically capture the spatiotemporal changes between nodes.To validate the effectiveness of the model,experiments are conducted on two public datasets,METR-LA and PEMS-BAY,and the results show that the prediction performance is better than that of multiple baseline methods.