Research On Key Technologies Of Traffic Prediction For Device-edge Collaboration In The Internet Of Vehicles Environmen

As a type of new information technology,the Internet of Vehicles enables interconnectedness between vehicles and between vehicles and infrastructure,providing a foundation for intelligent services.In the Internet of Vehicles,edge-side collaboration is an important means to achieve efficient cooperation among vehicles,devices,and users,providing powerful support for traffic management and travel decision-making.Traffic prediction can help traffic management departments to proactively understand future traffic conditions and develop effective measures to improve road utilization and reduce traffic congestion.In the context of the Internet of Vehicles and edge-side collaboration,traffic prediction needs to consider both prediction methods and resource optimization.In the Internet of Vehicles scenario,which relies on edge-side nodes such as mixed on-board computing units,roadside computing units,and road weather stations,traffic prediction methods mainly focus on the accuracy of prediction results,while ignoring the multi-source and heterogeneous features of traffic data and the training cost optimization for traffic prediction models.In addition,resource optimization for traffic prediction tasks focuses on computing latency and accuracy,but does not consider the computing and communication loads of edge-side nodes.Therefore,there are several challenges for traffic prediction in the context of edge-side collaboration in the Internet of Vehicles: 1)how to use multi-source and heterogeneous data to improve the accuracy of traffic prediction,and reduce the computational cost of traffic prediction models to establish lightweight models and edge-side collaboration models applicable to the Internet of Vehicles scenario;2)how to achieve edge-side collaborative computing and dynamic allocation of computing tasks in the Internet of Vehicles scenario to achieve system optimization in terms of execution latency,energy consumption,and accuracy.Therefore,this article focuses on researching traffic prediction technology for edge-side collaboration in the context of the Internet of Vehicles,with the following main objectives:(1)The traffic prediction method based on multi-source heterogeneous data in the V2 X environment proposes a deep learning approach for multiple traffic prediction tasks to comprehensively consider the feature extraction and training costs of multi-source heterogeneous data in the V2 X environment.Specifically,based on data analysis of the multiuser V2 X scenario,a self-attention mechanism-driven model is established,utilizing graph neural networks,location encoding techniques,and kernel function methods to optimize prediction accuracy.(2)The resource optimization method for edge-cloud collaboration oriented towards traffic prediction in the V2 X environment considers the trade-off between accuracy,time,and resource costs of traffic prediction.Specifically,the communication and resource cost model is established by simultaneously considering the model parameter size and node status during the traffic prediction process.Based on the model described in(1),a model with an early-exit structure is constructed by using an early-exit strategy.The action vector and reward function based on cost and accuracy are established,and deep Q-learning is used to decide the exit point of the model,balancing the accuracy and resource cost of the traffic prediction task.