Traffic Flow Prediction Based On Spatial-Temporal Graph Neural Network

At present,traffic prediction is a hot research direction of spatial-temporal data mining,aiming at extracting valuable information from the traffic spatial-temporal data.Traffic prediction utilizes deep learning methods to predict the traffic state of each place in the city in the future,which is of great significance to alleviating urban traffic congestion,vehicle operation scheduling,and daily travel planning.This paper predicts urban traffic flow from two aspects: urban traffic flow prediction and origin-destination demand prediction.On the one hand,this paper proposes a spatial-temporal data fusion-based multi-view model.Firstly,to extract the data’s spatial-temporal properties,a gated fusion network is designed in this paper.The data of traffic flow,time,and point of interest are cross-fused to capture the synergies between different data.Secondly,a spatial-temporal multi-view model is designed,which uses a one-dimensional convolution network,graph convolution network,and deep neural network to extract features from three perspectives of time,space,and space-time.Finally,a multi-task learning mechanism is designed to predict the traffic flow in urban regions.On the other hand,this paper proposes an OD matrix prediction method integrating traffic context and bidirectional semantic information.Aiming at the problem of the high sparsity of OD matrix brings difficulties to model learning.Firstly,the Node2 vec algorithm is used to train the embedded features of the region as the static traffic context,and the traffic flow of the region is counted as the dynamic traffic context.Secondly,two adjacency matrices representing origin-destination and destination-origin travel demands between urban areas are constructed based on the OD matrix of each time slice.Graph convolution network is used to aggregate the traffic context information of semantic neighbors in two directions.Then,the combination of GRU and graph convolution based on the recurrent neural network architecture captures the spatial-temporal correlation of input data and finally predicts the OD matrix of the next time slice.Experiments on real datasets show that the prediction accuracy of the proposed traffic flow prediction method and OD matrix prediction method is better than the existing baseline method.