Study On Travel Time Prediction Based On Convolutional Neural Network For Graph

Travel time prediction is very important for traffic managers and individuals.Especially in the increasingly serious traffic jams,travel time is not only the basic problem of traffic operation and management,but also the core issue of personal travel planning.The comprehensive and dynamic prediction of travel time can not only provide basic support for traffic control and management,but also alleviate traffic congestion at a large level.The existing travel time prediction methods either cannot effectively extract the spatial characteristics of the travel time,leading to low prediction accuracy or unable to expand on large-scale networks.In order to effectively solve the above problems,a travel time prediction method based on Convolutional Neural Network for Graph(GCN)is proposed in this paper.This method can effectively extract the spatial characteristics of travel time by spectral filtering the road network structure.Thus,the prediction accuracy rate is greatly improved;and the semi-supervised learning method is used to make the model have good scalability,so it is easy to extend the trained road network.First of all,in order to make the prediction model able to effectively extract the spatial characteristics of travel time,this study makes a preliminary transformation of the road network topology according to the convolutional neural network(CNN)structure,and converts the irregular data structure after conversion into a spectral filter.The data structure of the rules,resulting in a non-parametric filter that can extract spatial characteristics of travel time.However,the time complexity of the non-parametric filter obtained through spectrum filtering is O(n2).In the research,the time complexity of the filter is reduced to O(n)through polynomial fast filtering,thereby establishing a parameterized basis.Fig.Travel time prediction model of convolutional neural network.Secondly,based on the data features,a map convolutional neural network for travel time prediction is designed and implemented.After the network prediction results are analyzed,the average absolute error for 10 km TRIP prediction is 2 min,which shows that the model has a good accuracy..Comparing the prediction effect of the model with other algorithms,it is found that the accuracy of this model is greatly improved compared to other algorithms,which further proves the validity and necessity of the model establishment.Finally,the study uses a semi-supervised learning method to optimize the graph convolutional neural network,and analyzes the error and expansibility of the optimized model.The study found that when the proportion of labeled sections was reduced from 100%to 10%,the absolute error of the TRIP travel time only increased by 25%in the original basis,indicating that the model has a good scalability.The travel time prediction method based on graph convolutional neural network proposed in the paper has certain practical significance,and has good reference value for individual travel and traffic management.