Research On The Prediction And Mining Of Spatio-temporal Propagation Patterns Of Urban Traffic Congestion Based On Knowledge Graph

With the rapid improvement of the economic strength of our country and the acceleration of the process of urbanization,urban transportation participants have increased significantly,and the urban traffic demand increases dramatically,resulting in traffic congestion in a lot of major cities.How to induce the phenomenon of urban traffic congestion,dispatch urban traffic resources rationally and improve the efficiency of urban traffic is an urgent problem.The research for predicting congestion and mining the dynamic spreading patterns can capture the propagation patterns of urban traffic congestion in the temporal and spatial dimensions and explore the temporal and spatial factors that affect the propagation of urban traffic congestion,so as to predict the future urban traffic congestion accurately,and provide strong support for rational dispatch in future urban traffic resources,urban traffic congestion diversion and maintenance of urban traffic stability.Stable and accurate urban traffic congestion state prediction and urban traffic congestion spatio-temporal patterns mining have important research significance and practical application value in urban traffic.This paper predicts the congestion state of the main roads in the center district of Changchun city and mines the spatio-temporal propagation patterns of traffic congestion in road network.The research mainly uses the taxi trajectory data of Changchun City and the road network structure data of Changchun City as the research objects.We first map the taxi trajectory data on the Changchun City Road Network,and map the trajectory points of the taxis to specific road segments in the city.Then,we use the taxi trajectory data to calculate the average speed of each road segment at each timeslot in the road network.Finally,we utilize the average traffic speed to evaluate the current congestion state of each road.Because each road in road network is different from other roads,for example,road lanes and road types,we need to utilize the evaluation metric which takes the differences into consideration to evaluate the congestion state of traffic road network.In our work,we select the road congestion index to evaluate the traffic congestion status of specific road.In order to predict the congestion state of the urban traffic transportation system accurately in the future and mine the dynamic spreading patterns of urban traffic congestion propagation effectively,we proposes two novel models: Dynamic traffic correlations based spatio-Temporal graph convolutional network model(DTC-STGCN)and Path-aware based recurrent neural network model(PARN).The dynamic traffic correlations based spatio-temporal graph convolutional network model maps the congestion state of the urban traffic network in continuous time slices into continuous graph structure.In each graph,the nodes denote the road sections in the road network,the edges describe the connectivity of the connected roads,and the attributes of the nodes represent the congestion levels of the road segment in each time slice.In order to describe the dynamic temporal and spatial correlation of urban traffic,we use the dynamic flow transfer matrix to calculate the dynamic adjacency matrix in different time slices.Then,we execute the graph convolutions on each graph.At the same time,we propose a novel spatio-temporal attention mechanism to learn the dynamic spatio-temporal correlation of urban traffic.Finally,we utilize a unified graph convolutional network module to simultaneously utilize dynamic spatio-temporal features,road network structure and road features(for example,road types,and lanes)to predict the future road congestion state.The dynamic traffic correlations based spatio-temporal graph convolutional network model can obtain high traffic congestion state prediction accuracy,but it is lacking in the mining of urban traffic congestion spatio-temporal propagation patterns.Therefore,we propose a path-aware based recurrent neural network model.Similarly,the path-aware based recurrent neural network model maps the congestion state of the urban traffic network in continuous time slices into the same and continuous graph structure data Then,we link the connected roads in adjacent time slices according to the connectivity of the roads in the road network and we utilize a path selection algorithm to choose and unify the number of congestion propagation paths for each road.Next,we use a multi-layer LSTM network to obtain the vector representation of the congestion propagation paths of each road,and use a soft attention mechanism to learn the degree of influence of different congestion propagation paths on the congestion state of the target road.Finally,we use graph convolutional networks to integrate vector representations of different congestion paths,and then implement congestion prediction through a fully connected layer.The paper uses the taxi trajectory data in Changchun City for three weeks as the research object,which is used in the above methods to conduct experiments.The experimental results proved that our proposed DTC-STGCN model and PARN model can forecast the future urban traffic congestion state prediction effectively and precisely.In addition,we verify the effectiveness of the PARN model in the mining spatio-temporal of urban traffic congestion by experiments and analysis.