Research On Traffic Data Prediction Based On Graph Convolutional Network

As typical time-series data,road network traffic data is crucial for constructing intelligent cities.Accurate traffic flow prediction can enhance the road network’s operational efficiency,reduce accident rates and improve road capacity,which is of great significance in building a modern transport system.However,real-life traffic conditions are highly variable and uncertain and susceptible to fluctuations due to various factors,including traffic accidents.Therefore,it is still a great challenge to accurately model and predict the highly non-linear and complex traffic data and to grasp the spatiotemporal characteristics between traffic data.Although some methods already consider the spatiotemporal correlation between traffic data,using convolutional neural networks and recurrent neural networks to extract the spatiotemporal dependence of the data.However,most methods ignore the traffic network’s dynamic nature and cannot fully use the spatial features between neighboring sensors.In this paper to address this problem,this paper proposes a graph convolutional network-based approach to traffic flow prediction with the following main contributions:(1)In this paper,to address the limitations of convolutional neural networks in spatial feature extraction,graph convolutional networks are introduced to capture spatial dependencies in traffic data.In contrast,bi-directional long and short-term memory networks are fused to process temporal features in the data.A traffic prediction model based on graph convolution and bi-directional LSTM is proposed for the spatiotemporal modeling of traffic data.Among them,the graph convolution layer includes two layers of graph convolution,which is responsible for extracting spatial features;the Bi LSTM layer feeds sequential and inverse queues into the forward and reversed LSTM layers,respectively,to extract temporal features.Finally,the outputs of the two modules are stitched together,and a fully connected layer obtains the prediction results to prevent overfitting.(2)In this paper,the advantages of temporal convolutional networks in processing serialized data are proposed and investigated to address the limitations of recurrent neural networks in extracting traffic data.A traffic prediction model based on graph convolution and temporal convolutional networks is proposed to parallelize the inputs with fewer parameters and faster training.The model consists of two spatiotemporal convolutional blocks and a fully connected output layer.Each spatiotemporal convolutional block consists of two temporal gated layers and one spatial map convolutional layer.Within each block,residual connectivity and bottlenecking strategies are employed to evenly process the input time series data,explore spatiotemporal dependencies,and integrate the integrated features through the output layer to output the final prediction results.(3)A spatiotemporal attention mechanism is introduced to address the limitation that convolutional graph networks can hardly consider global features.A spatiotemporal attention-based graph convolutional and temporal convolutional traffic prediction model is designed.The model consists of a spatiotemporal attention block,a spatiotemporal convolution block,and a fully connected layer.The spatiotemporal attention mechanism is used to capture the dynamic spatiotemporal correlations in the traffic data,and then the spatiotemporal features of the traffic data are jointly extracted by graph convolution and temporal convolution,and finally,the prediction results are output using the fully connected layer.Experimental results on the real Pe MSD7 dataset show that the evaluation metrics in terms of MAPE for the three models mentioned above are 10.63,9.65,and 8.05,respectively.Compared to traditional convolutional neural methods,the predictive performance of these models has improved by 2.0%,11.1%,and 25.9%,respectively.Furthermore,the continuous improvements made among the three models have also led to a gradual enhancement in their predictive performance.