Research On Travel Time Prediction Algorithm Based On Fusion Of Attention Mechanism And Graph Convolution Method

Travel time in a city can be estimated by the analysis of urban road traffic status.Thereby,providing travelers travel routes based on the operating status of the road network estimation.However,the current research on this field has the problems of low utilization rate of low-frequency data,poor rationality of the model of travel time prediction,weak sensitivity of spatio-temporal correlation,singleness of model,etc.The specialties are not fully described by the problems in the field of traffic operation status,which makes the algorithm spatical-temporal complexity higher.Based on the induction and analysis of the existing travel time prediction model and deep learning mining model,this paper proposes a path travel time prediction algorithm with significant advantages.The main research results are as follows.Firstly,the paper takes Nangang District of Harbin City as the main research area,carrying out cleaning data and doing map matching for the trajectory data of floating vehicles with low frequency specialties.Then,the paper randomly selects 64 trajectory data from the area as the research object,and puts forward the nonparametric method which is used to predict the path travel time,so as to dynamically predict the travel time of each trajectory.Secondly,this paper introduces the spatical-temporal correlation to improve the shortcomings of prediction of path travel time in view of the problems of single input factors and poor explanatory power of the existing travel time prediction model.Dig deeping into the floating car data to discover the related data for travel time,the paper discovers its underlying regularity,and discovers spatial correlations for adjacent roads or trajectories.Eventually,a travel time prediction model based on deep learning method is established by combining attention mechanism and graph convolution method.In the prediction process,for one thing,the paper uses the path as the basic unit to model through the attention mechanism,and establishes a submodel of travel time prediction based on the LSTM tree to achieve correlation prediction in the time dimension;for another,the paper forms a graph convolution network sub-model based on the spatiotemporal attention mechanism by learning spatial correlation between the paths through the graph convolution network.The two sub-model is finally integrated through Bagging integration theory,which further improves the prediction accuracy.Finally,the training process of the above model was optimized by the gradient descent algorithm.The optimal network structure hyperparameters were selected through multiple trainings.After training,the optimal number of LSAM units is 7,and the optimal value of K in ASTGCN is 3.In the experiment,this model and the benchmark model respectively predicted the travel time of 64 randomly selected paths between 8:45 and 9:00 on January 17,2017.The results show that this model is superior to other benchmark models.The model has a more stable result of prediction.At the same time,after analyzing the influence factors of the error,it is found that the the path length is longer and the number of urban intersections is greater,the error is larger,but even in the face of this situation,the model error in this paper is still controllable and superior to other models.