Research On Traffic Flow Prediction Based On Improved Graph Attention Network

As urbanization accelerates and traffic load pressure becomes greater,scientific management and scheduling are urgently needed,and the development of intelligent transportation systems is imperative.The current research on traffic flow prediction mainly focuses on capturing the time series features of historical data,and the analysis of the dynamic spatial correlation features in the road network is insufficient,which leads to a large deviation of the prediction model from the real value.To address these problems,a traffic flow prediction model based on an improved graph attention network is proposed in this paper,and the specific work is as follows.The model uses an Encoder-Decoder system architecture,where both the encoder and decoder consist of multiple spatio-temporal attention modules to model the effects of spatial and temporal factors on traffic conditions.Each spatio-temporal attention module consists of a spatial attention mechanism and a temporal attention mechanism,and the spatio-temporal features are fused through a gated recursive unit.The traffic flow data input to the model will be encoded by the encoder,predicted by the decoder and output the prediction sequence.A spatio-temporal embedding module is proposed that will incorporate road network structural features and temporal information in the modeling process of temporal and spatial attention.To address the error propagation problem in road network propagation,a conversion module is added to the encoder and decoder architectures to perform conversion coding to model directly between historical time and future time,and adaptively filter the feature information entering the decoder by attention coefficients to reduce the error impact.Then,in this paper,the model is trained and compared with existing typical traffic flow prediction models,ARIMA,SVR,LSTM,STGCN,DCRNN,etc.,on a real dataset,and the mean absolute error(MAE),root mean square error(RMSE),and mean absolute percentage error(MAPE)are used as evaluation criteria.The model achieves excellent performance in 15-minute,30-minute,and 60-minute predictions,with reduced error metrics,especially for long-time(60-minute)traffic flow prediction.Further analysis of the fault tolerance of the model verifies that the model has stable prediction performance on the disabled data set.In addition,the ablation experiments fully demonstrate that each sub-module of the model in this paper is truly effective in improving the prediction performance of the model.