Analysis And Research On Key Technologies Of Traffic Flow Prediction

With the continuous advancement of urbanization and the continuous improvement of people’s living standards,a large number of transportation modes and vehicles have emerged in the city.On the one hand,these complex and numerous traffic modes improve people’s life.On the other hand,the contradiction between people’s increasing traffic demand and imperfect traffic management mode leads to the difficulty of maintaining traffic order and ensuring social security.Therefore,the construction of intelligent transportation system and the study of efficient,accurate and universal urban regional traffic flow prediction model are of more and more significance in maintaining smooth roads,formulating traffic policies,maintaining social stability and avoiding the tragedy of pedestrian congestion.Due to the problems of large amount of data,high real-time requirements,no obvious law of data,noise in data,many interference items,poor generalization ability of model and so on,how to predict urban traffic flow efficiently and accurately has become a hot issue in recent years.Although the existing deep learning and graph neural network have been widely used in urban traffic flow prediction,these methods often can not solve the above problems well.How to accurately model the complex spatial-temporal data of traffic flow,how to improve the model performance from the semantic relationship within the model,and how to reduce the impact of multiple external information on the accuracy and generalization ability of traffic prediction model have become the top priority of research in this field.In view of the above limitations and challenges,in the process of studying the traffic flow prediction problem,this paper aims to solve the problems of improving the prediction accuracy of the model,capturing the temporal and spatial dependence of data,poor generalization ability and low anti-interference ability in the traffic flow prediction problem by designing the model from four aspects of time,space,semantics and contrastive learning.Specifically,the contributions of this paper are summarized as follows:(1)In this paper,a multi-level spatial dependence encoder across regions is proposed to improve the prediction accuracy of the model from a spatial perspective.As for the regional flow in the city,the spatial proximity between regions constitutes a local spatial dependence.With the increase of distance,this proximity relationship is also weakening.However,even if the two geographical regions are not adjacent in space,their flow change trends can be very similar.For example,the urban functions of the two regions are the same,and they are all business districts.Therefore,it is necessary to enhance cross regional traffic dependency modeling from local to global environment.In this paper,a hierarchical graph neural network is designed,and a stacked graph attention network and a graph diffusion mechanism based on convolution are combined to maintain the local similarity and global dependence of traffic flow,so as to realize multi-level spatial relationship extraction across regions,which significantly enhances the ability of traffic flow prediction model to capture high non-linear geospatial structure.(2)This paper proposes a hierarchical framework with multi-level time periodicity.Traffic flow data is a kind of time series data with multi-resolution periodicity,which is specifically reflected in the interaction between hourly,daily and weekly time instances.In this paper,a multi-scale self attention structure and cross-resolution aggregation module are designed to explicitly embed multi-level time context signals into the potential representation of resolution perception.And through a hierarchical time self attention structure and multi view cooperation module,this model captures multi-layer dependence in time.A large number of comparative experiments and visualization experiments show that the multi-level time periodic hierarchical structure is very effective for the model to describe the highly complex time similarity.(3)This paper proposes a channel aware recalibration network for traffic flow prediction,which improves the model from the perspective of spatial semantic context.When studying the traffic prediction problem,the convolution neural core with different potential channels can learn the location dependence between geographical neighborhoods.The existing solutions usually treat the feature representation learned from different channels equally,and perform cross-channel feature aggregation with the same weight.Since the different channel embeddings may reflect different types of spatial semantics,the importance of different potential channel views may vary greatly.This paper designs a channel aware recalibration network in the spatial pattern integration paradigm,explicitly encodes the importance of different representation subspaces,and explicitly embeds the traffic flow signal into the channel aware relational encoder.(4)In this paper,a contrastive learning network for traffic flow prediction is proposed to enhance the anti-interference ability and generalization ability of the model.The incompleteness of traffic flow data and the noise often have a negative impact on the final prediction results.To solve the above problems,starting from the spatial and time dependence of traffic flow data,this paper designs four data augmentation methods for graph structure to improve the anti-interference ability of the model.In space,while maintaining the spatial structure of the graph,we can use the addition / deletion of points and edges to construct more data.In the time domain,the time continuity of traffic data is used for interpolation,and the mask method is used to make the designed module have a certain ability of self error correction.At the same time,this paper uses a semantic aware top-k strategy to filter the most difficult negative cases,improves the previous learning method of contrastive loss,and effectively avoids the interference of ”false negative” samples on the prediction accuracy of the model,so as to make the overall model more robust.A large number of experimental results in this paper show that the traffic flow prediction model with high efficiency,high accuracy and high generalization ability can be designed from the four aspects of time,space,semantics and contrastive learning,which is of great significance to the research and development of urban traffic flow prediction and the construction of a more intelligent and efficient intelligent transportation system.At the same time,this paper also provides ideas for designing urban service applications and solving various spatial-temporal prediction problems.