Research On Traffic Flow Prediction Based On Adaptive Spatiotemporal Graph Network

With the continuous development of domestic and international socioeconomic conditions,the demand for transportation among residents continues to increase.Given the contradiction between the limited transportation infrastructure and the continuous growth of transportation demand,transportation managers have proposed intelligent transportation system to cope with increasingly complex traffic scenarios.Traffic flow prediction is a fundamental task in achieving intelligent transportation systems.By predicting traffic flow information,transportation managers can optimize urban traffic in a timely manner,improve traffic efficiency,and reduce traffic congestion problems.Traffic flow prediction problems encompass various aspects such as traffic flow parameter prediction,congestion prediction,demand prediction,and accident risk prediction,among which traffic flow parameters include volume,speed,and density.This dissertation aims to conduct in-depth research on traffic flow parameter prediction and congestion prediction in the field of traffic flow prediction.Existing researches have difficulty in modeling the spatiotemporal correlation of complex and dynamic traffic flow data when dealing with increasingly complex traffic scenarios and a huge road network scale.The adaptive spatiotemporal graph network adopts a graph structure to model the spatial and temporal correlation of input data and dynamically processes non-Euclidean data,which has a wide range of application prospects in the field of traffic flow prediction.Therefore,this dissertation proposes two adaptive spatiotemporal graph network models,one for solving the parameter prediction problem and the other for solving the congestion prediction problem in the field of traffic flow prediction.For the traffic flow parameter prediction problem,this dissertation proposes a traffic volume prediction model based on a spatiotemporal deep network to solve the problem of ignoring the cross-regional functional dependencies of nodes in existing traffic flow parameter prediction models.The model uses a decoupling framework to decouple the traffic flow signal into instantaneous fusion signal and long-term dependence signal,and designs instantaneous fusion module and long-term dependence module to handle these two signals.In the instantaneous fusion module,local graph convolutional long short-term memory networks are used to model the instantaneous local spatiotemporal correlations between nodes.In the long-term dependence module,gate time convolutional layers and graph attention layers are combined and external feature information is integrated to model the long-term dependence exhibited by the functional characteristics of nodes themselves.Finally,the output features extracted by the two modules are fused to achieve prediction of multiple future time steps.Experiments show that the proposed model performs better than the majority of baseline models on three real traffic volume datasets and has robustness in long-term prediction.For the congestion prediction problem,this dissertation proposes a traffic congestion prediction model based on an autoencoder to solve the limitation of existing autoencoder algorithms in handling the complex dynamic spatiotemporal correlations in traffic flow data.The model is based on an autoencoder architecture and proposes an adaptive spatiotemporal dynamic fusion layer to build the encoder and decoder.To address the problem of dynamic feature changes between congestion and non-congestion states in traffic congestion prediction,the model introduces a data-driven dynamic graph generation module to model the dynamic feature relationships between nodes.The multigraph fusion module fuses the node spatiotemporal features extracted from the dynamic graph based on node similarity and the static graph based on node spatial relationships to achieve congestion prediction.In the model training part,a class-inter distance loss function is introduced to ensure the effectiveness of the features extracted by the autoencoder for downstream tasks.Extensive experiments are conducted on two real traffic congestion datasets,and the results demonstrate the effectiveness of the proposed model in the traffic congestion prediction problem.