| Traffic prediction establishes a model through historical traffic data to predict the future traffic conditions.The excellent prediction model can predict the future traffic conditions in advance and provide reference for traffic control personnel,so as to alleviate the possible traffic congestion and the possible economic losses caused by traffic congestion through control measures such as traffic diversion.The prediction model can also help individuals save travel time and help cities improve operation efficiency.Based on the above background,the main work of this thesis includes the following three aspects:1.A traffic prediction model AGC-LSTM based on adaptive graph convolution and long short-term memory network(LSTM)is proposed.It first adds spatial coding to the original input data.Then it combines graph convolution operation with LSTM,and uses adaptive graph convolution operation to replace the linear transformation in LSTM to obtain a graph convolution memory network module,GLSTM for spatio-temporal feature extraction.Finally it outputs the prediction results through the full connection layer.The proposed model is evaluated on four public data sets,and its performance is generally better than other baseline models.2.A traffic prediction model,MA-STGCN,based on multi-adjacency matrix and multihead time attention mechanism is proposed.Firstly,the model adds spatio-temporal coding to the original input data to make the data contain spatio-temporal characteristics.The spatialtemporal features are then fully captured by multiple stacked spatial-temporal feature blocks,and residual connections are used at each layer to improve the model generalization ability.Finally,the prediction results are output through the full connection layer.Here,the spatiotemporal feature block consists of three parts: 1)The temporal feature is extracted by expanding the time convolution block;2)The spatial features are extracted through the multigraph convolution layer.The multi-graph convolution includes three graph convolution blocks.The original adjacency matrix,adaptive adjacency matrix and the global similar adjacency matrix containing the spatial information of the road network which are constructed through the graph embedding algorithm are used for graph convolution to extract more abundant spatial features;3)The temporal features are further extracted through the multi head temporal attention mechanism.Through experimental verification,the performance of MA-STGCN exceeds AGC-LSTM in all data sets.3.A traffic visualization system is designed and developed.The platform is developed based on the Flask framework,and the database adopts My SQL.The system calculates the road condition classification threshold through the historical vehicle speed information and the traffic condition is visualized through Baidu map API.The MA-STGCN model is integrated to provide real-time vehicle speed prediction function.ECharts is applied to display the historical data of observation points. |
