Applied Research Of Spatial Temporal Graph Model In Traffic Flow Prediction

In daily life and travel,the impact of traffic congestion can not be ignored.For commuters,knowing the traffic jam in the morning and evening rush hours in advance can effectively avoid the situation of being late and affecting work efficiency.For people traveling,once encountered congestion,not only affect the mood of travel,but also easy to cause traffic accidents,bring unnecessary losses.Therefore,realtime and accurate traffic flow prediction can provide road information,provide a lot of convenience for people’s travel,but also help the traffic management department to better manage the traffic situation,traffic flow prediction is a core problem in intelligent transportation system,has a certain application value.The main research of traffic flow prediction problem is the traffic flow data on the road network.This kind of spatiotemporal data has time correlation and space correlation.Aiming at this spatiotemporal characteristics,this dissetation studies from these two aspects respectively,and designs a model to effectively improve the accuracy of traffic flow data prediction.First of all,in this dissertation,based on the graph convolution neural network combined with attention mechanism of the learning algorithm of traffic flow spatial characteristics,to the extent of interaction between different nodes assigned different weights,at the same time introducing node adaptive learning,sharing mode,changed the traditional parameters to better improve the expressing ability of graph model,extract the spatial characteristics.Secondly,based on temporal convolution of the network traffic flow characteristics of the learning algorithm,through causal convolution ensures that the dimension of the input and output data is consistent,and expansion of convolution by setting the sampling interval to receptive field,flexible control for the length of the long time series data,also is able to extract the time characteristics.Finally,a traffic flow prediction model based on spatio-temporal graph attention was established.The model extracted spatial features by combining graph convolutional neural network and attention mechanism,and learned parameters under different modes to improve the model effect.The temporal convolutional network is used to expand the receptive field to capture the temporal characteristics better,the residual connection is added to prevent the over-fitting caused by the excessive depth of network layers.In order to verify the validity of the model,this paper verifies the model on real traffic data sets Metr-La and PEMS-Bay,and conducts experiments with the current common prediction methods to analyze their performance.The experimental results show that compared with the current mainstream prediction methods,the method proposed in this paper has more accurate prediction results and smaller prediction errors in the next 15 minutes,30 minutes and 60 minutes.The validity of the model is proved,which provides a certain theoretical support for the development of a traffic flow forecasting system with better robustness and higher accuracy.