Research On Traffic Prediction And Situation Generation Based On Deep Network

Traffic prediction is the basic task of the urban intelligent transportation systems,and its core problem is how to accurately mine the complex spatio-temporal correlations inherent in traffic data.At present,in terms of normal short-term traffic prediction,traditional sequence-based learning methods,machine learning methods,and deep neural network methods can all achieve good prediction performance and have been widely used.However,with the acceleration of urbanization and the continuous improvement of transportation infrastructure,the number of nodes and edges in transportation-related networks such as road networks,bus networks,and rail networks is increasing,the connection forms are diverse,and the structures have become complex.Previous structured data representation methods are unable to fully characterize transportation networks.When a transportation system undergoes significant changes,that is,when the balance of the transportation system is disrupted,the transportation system may experience abrupt changes.For instance,traffic conditions can be influenced by various factors such as sports events,large-scale city activities,emergencies,and extreme weather.Under such abnormal circumstances,predicting traffic and generating traffic situations is difficult to achieve ideal performance.However,predicting traffic and generating traffic situations under abnormal conditions have more guiding significance for public travel,enterprise operation,and government regulation.To address these issues,this study focuses on traffic prediction and traffic situation generation based on graph convolutional networks and cross-modal data fusion.This thesis delves deeply into the representation of graph structures for traffic road network data,prediction models based on graph convolutional networks,and cross-modal data fusion.The following innovative achievements have been made:First,the hierarchical traffic prediction model based on spatiotemporal graph convolution and attention mechanism.Traditional traffic flow prediction models are mainly oriented to short-term traffic flow prediction tasks of standardized traffic data.It is difficult to simultaneously capture the spatio-temporal relationship carried out by city-level traffic data.In response to the above problems,this thesis establishes a hierarchical traffic prediction model based on spatio-temporal graph convolution and attention mechanism.The model mainly includes three modules: the long-term Transformer network module mines the long-term temporal correlation of traffic data;the spatio-temporal graph convolution module explores the short-term spatio-temporal correlation of traffic data;the attention fusion module of long-term and short-term spatio-temporal information is used to coordinate the temporal information of longterm and short-term traffic flow and the spatial information of traffic road network.This hierarchical network architecture also alleviates the over-smoothing problem of the graph convolutional neural network.Second,traffic prediction model based on dynamic graph structure learning.Traditional traffic flow prediction models based on graph neural networks usually adopt static graph representations based on road distance or road network structure,but such graph representations overlook the feature correlations hidden in the traffic sequence data.To solve this problem,this thesis proposes a graph convolutional network model based on dynamic graph structure learning.The model includes three main modules:DTW-based global graph generator mines the long-term similarity contained in traffic flow sequences and constructs a global similarity graph of the road network;the gradient-based local graph generator finds the short-term similarity between traffic flow sequences and constructs a dynamic road network similarity graph;the multivariate time series prediction module uses the fused road network dynamic graph to mine the spatio-temporal relationship of traffic data and realize accurate traffic flow prediction.Third,the travel demand forecasting model based on the bidirectional graph convolutional network.Unlike traffic flow forecasting,travel demand prediction mainly focuses on predicting the travel behavior of individuals or groups,such as when and where they will travel.Travel prediction needs to consider people’s travel decisions and behaviors,and these decisions are influenced by many factors,including individual characteristics,socioeconomic background,traffic network status,travel purposes,etc.,making it difficult to reveal the complex correlations in individual travel.To address this issue,this thesis proposes a travel demand prediction model based on bidirectional graph convolutional neural networks,which includes two sub-modules: the bidirectional graph convolutional network module and the gated temporal convolutional network module.The bidirectional graph convolutional network captures the bidirectional correlations between origins and destinations in travel data,while the gated temporal convolutional network dynamically captures the temporal correlation between travel data.The proposed model achieves state-of-the-art results on two public datasets.Fourth,the traffic situation generation model based on generative adversarial networks.Considering that the traffic system is affected by multiple additional factors,such as weather,traffic accidents,and large-scale events,the traffic situation under these non-regular factors is difficult to predict and analyze.Unlike traffic data collected by sensors,social media data can directly show the causes of traffic congestion,which is excellent guidance for predicting and analyzing traffic situations.This thesis focuses on the generation of macro traffic situations under emergencies and proposes a traffic situation generation model based on generative adversarial networks.The text embedding module normalizes the traffic-related information in social media,which can help machines understand and process text data for subsequent analysis and mining.The concept of "adversarial generation" is introduced to analyze the impact of individual or multiple emergencies at the macroscopic level,thereby achieving interactive traffic situation evolution analysis.In summary,this thesis focuses on tasks such as urban traffic flow prediction,travel demand prediction,and traffic situation generation.This thesis introduces graph convolutional network models to extract spatial information from road network sequence data.It proposes a fusion attention mechanism to jointly capture the implicit long-term and short-term temporal correlation features and spatial features in traffic data.The discrete sampling approach is proposed to learn the dynamic relationships between road networks in the traffic system.The proposed bidirectional graph convolutional neural network model addresses the bidirectional correlation of traffic trips,characterizes the inherent connections among individual traffic trips,and significantly improves the accuracy of traffic demand prediction.Based on the polysemy,heterogeneity,massiveness,and sociality of traffic data,this thesis researches cross-modal traffic information fusion and builds a spatio-temporal deep neural network model for cross-modal data.To address unexpected situations in complex traffic systems,a traffic situation generation model based on generative adversarial networks is studied in depth,breaking through the technical bottleneck of multi-source cross-domain data fusion and achieving accurate prediction and situation generation of traffic situations under normal and abnormal conditions.These findings provide a solid foundation for the construction of intelligent traffic systems.