Study On Vehicles Flow Prediction Based On Data Augmentation And Spatial-temporal Factor

With the concept of green and safe travel spreading and the rapid development of vehicles networks technology,spatial-temporal vehicle flow data prediction gradually becomes a hot issue.Accurate and efficient vehicle flow data prediction is beneficial to traffic managers for the control of road networks and reduces economic losses and cost time of citizens.However,the sensors that monitor data will cause severe data missing if they break down.Traditional data augmentation methods can alleviate data missing and fill in data according to fixed distribution assumptions,such as Gaussian distribution,but these fixed data distributions cannot fit the real flow changes;On the other hand,vehicle flow data presents highly complex variability,because of the complex design of road network and the influence of real-time traffic status.Existed flow prediction algorithms can consider spatial-temporal information at the same time,but in the spatial dimension,only considers the spatial node’s flow change(content information)and neighbor information,ignoring the influence of distance information and direction information between nodes;in the time dimension,deep learning models used for information extraction can not avoid the loss of information,due to their structural design.Facing these problems,we focus on data level and technology level to explore data enhancement and prediction algorithms,aiming to solve these challenges: A.Simulate the vehicle flow data distribution that presents highly complex variability,and generate reliable data to fill in missing data;B.Integrate the content,neighborhood,distance and direction information of the spatial nodes of vehicle flow data;C.Design a deep learning model framework that can prevent the loss of time information.In view of the above challenges,we explore technical research,design experiments,and develop a system for data visualization display.All contents are mainly divided into four parts:(1)This paper analyzes the effects of lacking data for vehicle flow data prediction,and for challenge A,designs a game mechanism based on generative adversarial networks to fit the complex distribution of vehicle flow data.This paper proposes the enhanced model technology to data complement.(2)This paper summarizes the shortcomings of existing deep learning research in spatial feature extraction and temporal feature extraction and proposes a model framework of spatial-temporal vehicle flow data prediction.For challenge B,we design a spatial feature extraction based on graph convolutional neural network to integrate the content,neighborhood,distance,and direction information.For challenge C,we design a time feature extractor based on self-attention mechanism,to solve the loss problem of time information.(3)This paper designs and conducts experiments on two open data sets.We compare the experimental results of the classical and the latest technology methods,to prove the validity of technology studies.(4)This paper develops a website system for vehicle flow data prediction,visualizing the vehicle flow data,to prove the practicability of the theoretical research in this paper.