A Research And Implementation Of Traffic Flow Prediction Algorithm Based On Convolutional Neural Network And Graph Convolutional Network

In recent years,with the rapid development of urbanization,urban traffic flow has increased dramatically.To a certain extent,the prediction of urban traffic flow can improve the operation efficiency of vehicles,reasonably dispatch vehicles to ease traffic congestion and facilitate people’s travel.However,spatiotemporal data has complex spatiotemporal dependence and periodicity.For example,in urban road network,there are direct or indirect links between road sections,so the flow change of one road section will inevitably have a direct or potential impact on other road sections.At the same time,due to the regularity of human activity mode,the traffic flow has the similarity of long and short periods,for example,the trend of flow between working days is similar,there will be obvious morning peak and evening peak.Therefore,how to model the temporal and spatial relationship of spatiotemporal data and other influencing factors becomes the key to accurately predict traffic flow.This research proposes a multi task deep learning method based on CNN.A well-designed attention based multi-scale multi task net(AMSMT-Net)is used to jointly predict the arrival bus service times,line level on-board passenger flow and line level up and down passenger flow.Multi task learning framework can strengthen the interaction between each type of flow,and finally integrate the output to achieve fine-grained service level prediction;multi-scale module can mine different spatial ranges through a variety of convolution cores with different sizes to model more extensive spatial relationships in detail;attention module can give different time slices and regions different important information through compression and re incentive mechanism By encoding internal and external factors(such as geographic location information,weather information,holiday information),the impact of these factors on traffic flow is integrated.Based on GCN,this research proposes a multi-embedding spatial-temporal GCN(ME-STGCN).The core of the algorithm framework is to construct a fixed embedded space based on the point of interest(POI)information to store the topology information based on geographical location,and then learn and construct a dynamic topology spatial structure that changes with time based on the traffic flow historical observation data,so as to adaptively adjust the connection relationship between nodes.At the same time,an improved geometric aggregator is further used to update the information of each node in the graph more accurately by defining neighbors with different geometric relationships and making full use of the graph topology to give different importance to different neighbors.Aiming at the multi-scale attention multi task convolutional neural network,a large number of experiments are carried out based on a large-scale real bus operation data set.The results show that AMSMT-Net is superior to the ten latest models and improves the accuracy by 22.39%.Through the comparison of the models,the effectiveness and reliability of multi task module,multi-scale module and attention mechanism are verified.In this paper,a large number of experiments are carried out on the actual traffic flow data set.The results show that ME-STGCN can improve the relative error of 12%and 15%respectively for the two tasks in the bus data set.The effectiveness and reliability of various embedded spaces are verified through the comparison of the models.