A Traffic Flow Prediction Method Based On Spatial-temporal Graph Neural Network

Traffic flow prediction task is an important part of Intelligent Transportation System.Accurate and efficient traffic prediction is of great significance to traffic planning and dispatching.However,traffic prediction task is very challenging due to the complex and non-linear spatial-temporal correlations of traffic flow data.Existing traffic flow prediction methods do not make full use of spatial-temporal features,and cannot explore deep-level spatial-temporal correlations of traffic flow.In addition,the shortage of data is very common in actual traffic scenarios.The lack of sufficient data samples for optimizing prediction model also increases the challenge of traffic flow prediction task.Existing small-sample traffic flow prediction methods only model grid structure data,and cannot model the non-Euclidean structure data which is closer to actual traffic scenarios.Aiming at the insufficient utilization of spatial-temporal features,this thesis proposed an adaptive spatial-temporal graph neural network model for traffic flow prediction task.A temporal self-attention mechanism is used to capture the temporal correlations of different historical time steps,and a gated temporal convolutional neural network is used to capture the multi-level temporal dependencies.Moreover,a diffusion graph convolutional neural network is used to simulate the local spatial correlations between the nodes,and parameterized adjacency matrices are used to adaptively capture the hidden global deep spatial dependencies in the road network.The multi-layer spatial-temporal convolution module is used for feature extraction,and spatial-temporal features are fully utilized for traffic flow prediction.Aiming at the shortage of data samples,this thesis proposed a transfer learning spatial-temporal graph neural network model for traffic flow prediction task under small-sample condition.First,an internal attention mechanism is used to capture the correlations between the internal feature sequences of traffic sensors.After that,wavelet decomposition and reconstruction methods are used to extract temporal core components and detail components of traffic flow data.Gated temporal convolutional neural network and graph convolutional neural network are used to extract spatial-temporal features of the core components,and a spatial-temporal attention mechanism is used to extract spatial-temporal pattern features.Moreover,a linear prediction module is used to extract spatial-temporal features of the temporal detail components and fuse them with the prediction results of the core components.This modeling method makes full use of similar patterns in different data sources,and can reduce overfitting problem caused by differences in data patterns.This thesis conducted performance evaluation experiments on four real-world traffic datasets,and used evaluation metrics to compare the prediction performance of the model proposed in this thesis with other benchmark models.The experiment results respectively proved the effectiveness of the adaptive spatial-temporal graph neural network prediction model and the transfer learning spatial-temporal graph neural network prediction model proposed in this thesis.And this thesis verified the effectiveness of different modules in the prediction models through ablation analysis.