Traffic Prediction Based On Spatio-Temporal Graph Neural Network

With the development of urbanization,traffic forecasting plays an important role in traffic planning and urban management.In the task of traffic forecasting,it is still a great challenge to model the complex dynamic spatio-temporal correlation.Therefore,it is of great practical significance to explore algorithms that can effectively capture the spatio-temporal correlation of traffic data.In recent years,the use of spatio-temporal graph neural networks to model the spatio-temporal correlation of traffic data has become popular and achieved certain results.However,most of the current spatio-temporal graph neural networks used for traffic speed prediction focus on extracting local spatial features,ignoring global spatial information,and it is difficult to take into account in both short-term and long-term prediction tasks.Therefore,A kind of combined attention mechanism of spatio-temporal graph convolutional network is designed and used for integrating global spatial information,and improving the ability of handling time series at the same time.And that time series processing capability is improved.In addition,considering that the predefined graph structure does not contain complete spatial information and the existing models fail to perform long-term traffic flow prediction well,an adaptive spatio-temporal graph neural network is designed to extract more complete spatial features and further improve the performance of long-term traffic flow prediction.The work and research results of this paper are as follows:(1)Design an attention spatio-temporal graph neural network to improve traffic speed forecasting performance.Aiming at the problems that the current spatio-temporal convolutional network ignores the global characteristics of space and is difficult to take into account in both long-term and short-term speed prediction,an attention spatio-temporal graph neural network is designed for traffic speed prediction.Firstly,the attention mechanism is introduced to adjust the importance of adjacent roads and non-adjacent roads,and integrate the global spatial information.Then,Graph Convolutional Networks(GCNs)and Gated Linear Unit(GLU)with extended causal convolution are used to capture the spatio-temporal correlation of traffic data.At the same time,the residual learning framework is used to improve the convergence speed and enhances the model fitting ability.Experiments on real data sets show that the proposed method has better prediction effect than graphwavenet and other methods.(2)Design an adaptive spatio-temporal graph neural network to improve the traffic flow forecasting performance.Aiming at the problem that the predefined graph structure used in the previous complex neural network framework does not contain complete spatial information in the spatial dimension,and cannot well capture the long-term time dependence of traffic data in the time dimension.An adaptive spatio-temporal neural network is designed for traffic flow prediction.Firstly,the Adaptive Graph Convolutional Network(AGCN)is used to automatically capture the specific state of nodes and automatically infer the interdependence between different nodes,to extract more complete spatial features.Then,the time memory module in the Spatio-Temporal Long-Short Term Memory(ST-LSTM)is used to enhance the ability of long sequence processing while capturing the short-term,medium-term and long-term time dependence of data.The experimental results on real data sets show that the proposed network has better prediction performance than ASTGCN,STSGCN and AGCRN.To sum up,this paper is mainly rooted in the relevant algorithms of spatio-temporal graph neural networks for traffic prediction tasks,focusing on capturing the spatio-temporal correlation of traffic data to improve speed prediction performance and flow prediction performance.