Traffic Flow Prediction Based On Graph Attention Neural Network

With the steady development of the domestic economy and the continuous progress of vehicle manufacturing science,the number of motor vehicles has increased year by year,and the traffic problems in various cities have gradually become obvious.How to effectively combine communication technology and information technology to solve the problem of traffic congestion has become a problem.important issues of our time.Traffic flow forecasting has become a hot research topic,and more and more scholars at home and abroad are devoted to the research in this field.In the past ten years,researchers from all over the world have proposed many different traffic flow forecasting methods,but most of the methods are used to predict it at the time series level.There is still a big gap between the values.In response to these problems,we proposes a new spatiotemporal prediction model.The main work is as follows:（1）Analyze traffic flow data,use Lagrangian interpolation method to fill missing values,moving average method to remove noise,and use the maximum and minimum method to normalize the data.（2）Predict the traffic flow from the time level,and appropriately improve the DA-RNN model and apply it to the field of traffic flow prediction for the first time.The model is an encoder-decoder structure.The feature weight of time series data of traffic flow is extracted,and a second attention mechanism is introduced in the decoder part to obtain the most suitable hidden layer state,so as to obtain more accurate prediction results at the time series level.The experimental results show that the MSE value of the DA-RNN model is reduced by 29.25%compared with the two-layer LSTM model and 29.85% compared with the two-layer GRU model;compared with the two-layer LSTM and two-layer GRU models,the MAE value is reduced 15.92%.（3）To predict the traffic flow from the time and space level,a TSAGCN model is proposed.This model first builds the traffic flow sensor into a graph structure,inputs the traffic flow data into the graph neural network,and combines the spatial convolution,The temporal convolution and attention mechanism accurately extract the temporal and spatial features of the traffic flow data,so as to train a prediction model that is most similar to the real road traffic situation,thereby improving the accuracy of the final prediction result.The results show that the MAE value of the TSAGCN model is reduced by 16.22% compared with the GAMN model and 28.34% compared with the STGCN model.We conducts experiments based on PEMS04 data and compares the prediction results with the prediction results of mainstream deep learning models.It shows that the prediction model mentioned in this experiment has good performance and can provide an effective basis for traffic management and control.