Research On Traffic Accident Prediction Based On Graph Neural Network

With the development of the world economy,the living standards of people in various countries are improving,and the number of private cars owned by citizens is also increasing.However,traffic accidents occur frequently,which seriously affects the safety of life and property of citizens.At present,the research of traffic accident prediction based on machine learning method,some ignore the temporal correlation of traffic accidents,some ignore the spatial correlation of traffic accidents.In this paper,a graph neural network model based on two channels(Bi-Merge Spatial-Temporal Graph Convolution Network,BMSTGCN)is proposed to solve the problem of traffic accident prediction.It can effectively solve the spatio-temporal correlation.BMSTGCN model predicts daily and weekly data through two channels,Both channels capture the spatio-temporal correlation of traffic accident data effectively by using time dimension convolution and space dimension convolution,and fuse the results of the two channels in the full connection layer to obtain the periodic law of traffic data.This paper conducts experiments on traffic accident data sets of Los Angeles freeway and ordinary highway in California.The results show that the prediction effect of BMSTGCN model is significantly better than other baseline models.This paper mainly proposes a method of traffic accident prediction based on graph neural network model.The specific research contents include:(1)based on graph theory,this paper introduces the basic elements of graph neural network and how graph neural network is extended from Fourier transform to graph convolution.(2)Define the important elements of graph neural network model.(3)In order to solve the problem of traffic accident data sparsity and map matching,this paper proposes a method of using grid to divide the region,which effectively solves the problem of traffic accident sparsity,improves the performance of graph neural network model,and optimizes the operation efficiency of the model;in order to solve the time correlation of traffic accident data,this paper proposes the method of using daily cycle This method effectively fuses the time correlation of traffic accident data;in order to ensure that the results of traffic accident prediction are meaningful,a mask layer is proposed to improve the accuracy of model prediction.(4)The graph neural network model based on deep learning is constructed to verify the feasibility and effectiveness of the graph neural network model in traffic accident prediction.