Research On Traffic Flow Prediction Based On Graph Neural Networks

Intelligent transportation system is a project that countries around the world have been committed to building in recent years,which can effectively manage and control traffic conditions.As the core component of intelligent transportation system,traffic flow prediction can provide timely and accurate data analysis and theoretical basis for transportation departments,so as to optimize the operation efficiency of urban road network and improve the quality of citizens’ travel.However,traffic flow data has complex spatiotemporal correlations and nonlinear dynamic changes(such as sudden events or changes in road conditions),which brings great challenges to traffic flow prediction.Most of the current traffic flow forecasting models ignore the learning ability of dynamic spatial correlations of traffic data,and the modeling of periodic and trendy characteristics is insufficient.To address these challenges,this paper proposes two new traffic flow prediction models based on graph neural networks and recurrent neural networks,and demonstrates the effectiveness of the proposed models through experiments on real-world traffic datasets.The main research content and contributions of this paper are as follows:(1)To address the shortcomings of existing methods in learning dynamic spatial correlation and ignoring long-distance dependency modeling,this paper proposes a new spatiotemporal attention-based graph convolutional recurrent neural network(AGCRNN)model.The model designs a spatiotemporal attention graph convolutional gated recurrent unit(STGCGRU),which combines spatiotemporal attention mechanism and uses graph convolutional network to modify the GRU module,which can more effectively model the long-distance and local dynamic spatiotemporal dependencies.Then,the model adopts a sequence-to-sequence architecture and stacks recurrent neural networks with STGCGRU to perform multi-step traffic prediction.Experiments on two real-world public datasets show that the proposed AGCRNN model performs better than other similar baseline models in terms of prediction performance.(2)To address the problems that AGCRNN model cannot handle directed graph-structured data and the existing models cannot model periodic and trend features in time series data well,this paper proposes a new traffic flow prediction model based on graph attention network and Auto-Correlation,which is named GAT-Auto Corr-LSTM.The model integrates GAT,Auto-Correlation and LSTM to capture the spatiotemporal correlation in traffic networks.GAT uses masked attention to incorporate graph structure into the attention mechanism to better extract the spatial correlation between road neighbor nodes in the spatial domain.Then,Auto-Correlation finds the periodicity and trend between different positions in the time series from the frequency domain by transplanting self-attention to temporal computation.LSTM network captures the long-term and short-term temporal dependency and nonlinear relationship of traffic data.Experiments on real-world dataset Pe MSD7 show that GAT-Auto Corr-LSTM has better prediction performance than other similar baseline models.