Research On Traffic Congestion Prediction With Low Communication Overhead

"With the help of technologies such as artificial intelligence,big data,and the Internet of Vehicles,intelligent transportation systems could realize all-round and real-time monitoring of vehicle operating states and the overall conditions of roads.In order to alleviate the increasingly serious traffic congestion problem,help people avoid congestion and plan travel routes,it is necessary to study the urban traffic congestion prediction.However,a large number of traffic data between server and terminals should be transmitted in real time for congestion prediction.Therefore,the problem of communication overhead in traffic prediction needs to be further investigated.To balance the tradeoff between real-time prediction performance and communication overhead,a traffic congestion prediction method based on feature importance is first proposed.Then,a deep traffic congestion prediction method based on node degree is studied.The specific contents include:A traffic congestion prediction method based on feature importance.This problem is modeled as a time series regression problem,and a regression random forest algorithm with multi-output is adopted as the prediction model for importance extraction with low communication overhead.The whole method is divided into three stages.First,the feature importance of road segments is calculated based on random forest.Then in the model pre-training stage,the communication overhead is reduced from two aspects.On the one hand,based on the spatial correlation of the road data,the important data is selected for uploading according to the feature importance,which directly reduces the number of road segments.On the other hand,based on the generalization ability of the model,the number of quantization bits is reduced by designing different quantization methods,and therefore the amount of transmitted data reduced.In the real-time prediction stage,the terminals send the road segment data according to the optimal number of transmitted road segments and the optimal quantization method feedback from the server.The server receives the data to predict congestion.The simulation results indicate that the traffic congestion prediction method based on feature importance can improve the prediction accuracy while reducing the system communication overhead.A deep prediction method based on node degree for traffic congestion.Taking advantage of the spatial structure of urban traffic road network and the knowledge of graph theory,this problem is modeled as a time series regression problem on topological graphs.At the same time,a Weights Learnable Time Graph Convolutional Network is proposed as our prediction model.The whole method is divided into two stages.First,in the model pre-training stage,the actual traffic network is converted into a topology map.Then,the communication overhead is reduced from two aspects.On the one hand,important road segment data is selected for uploading based on the node degree,and the data of road segments that are not transmitted is completed by the sever,therefore the number of uploaded road segments is reduced.On the other hand,based on the generalization ability of the model,the quantization bits of the transmitted data could be reduced so that reducing the amount of transmitted data.In the real-time prediction stage,the terminals send the road segment data according to the optimal number of transmitted road segments and the optimal quantization bits feedback from the server.The server receives the data to predict congestion.The simulation results verify the proposed road selection method and data completion method,as well as the advantages of prediction network,and indicate that the deep prediction method based on node degree for traffic congestion can improve the prediction accuracy and reduce the communication overhead.