Dynamic Spatio-Temporal Graph Convolutional Neural Networks For Traffic Flow Forecasting

The rapid increase in motor vehicle ownership has aggravated the traffic congestion phenomena,which seriously hinders the healthy development of cities.Traffic flow forecasting is one of the effective methods to alleviate the pressure of urban traffic.Accurate traffic flow prediction models for road networks provide theoretical bases for the transportation department to formulate reasonable management measures and reduce the time cost of public travel,so as to relieve the urban traffic congestion.However,most of the current traffic flow forecasting methods for road network lack abilities of adequately capturing the spatio-temporal interdependency concealed in traffic data,and hardly yield satisfactory results when prediction duration varies.Therefore,this paper concentrates on the road network traffic flow forecasting.In order to adaptively learn the spatio-temporal dependencies of traffic flow data and road network topology,this paper has proposed two novel road network traffic flow prediction models based on graph convolutional neural networks which effectively promote the prediction performance.The detailed research work could be summarized as follows:(1)Road network and traffic feature modelling.Based on the real highway traffic flow data,this paper deeply studies the spatio-temporal variation features of the road network traffic flow,models the road network topology based on the graph theory,and puts forward a traffic flow feature tensor construction method which combines the local and periodic correlation to provide a support for the design of the road network traffic flow prediction model.(2)Graph convolutional neural networks for spatio-temporal traffic flow forecasting(GCST).In order to adequately mine the spatio-temporal dependencies of road network traffic flow,a learning mechanism for traffic flow features is proposed.In the spatial learning mechanism,a global spatial topology learning layer is designed to study the road network matrix,then a gated mechanism is applied to the graph convolutional neural network,which enhances the learning ability for spatial correlation.In the temporal learning mechanism,the long shortterm memory neural network is utilized to capture the temporal features,then a vertical temporal convolution is proposed to improve computation efficiency while preserving local temporal correlation of traffic flow.The experimental results have demonstrated that the long-term forecasting accuracies of GCST are obviously superior to the other baseline forecasting models.(3)Dynamic spatio-temporal Chebyshev graph convolutional neural networks for traffic flow forecasting(DCGCN).Fix Laplacian matrix could not precisely express the dynamic variation features of road network topology.To solve this problem,a dynamic Laplacian matrix learning mechanism is designed,which adaptively learns the spatio-temporal changing tendencies of traffic flow and dynamically updates the weights of Laplacian matrix.In the meantime,because the prediction results of GCST are less satisfactory in short-term traffic flow forecasting,the proposed method combines local temporal convolution and temporal attention mechanism to automatically learn the long short-term correlation of road network traffic flow.The experimental results have indicated that the forecasting accuracies of DCGCN are increased by 6.7% compared with the other baseline models,and the proposed method is robust enough to resist the influence of common noises.