Research On Traffic Accident Risk Prediction Method Based On Neural Network

Traffic accidents have always been the focus of national and social attention.If traffic accidents can be effectively avoided,some urban traffic congestion can be avoided and the burden of urban traffic management can be reduced.Over the years,many scholars predict traffic accidents become a challenging research,if we can accurately predict the occurrence of traffic accidents,and according to the result of prediction for the occurrence of traffic accidents to programming,so will not only improve the efficiency of urban traffic,will reduce the number of traffic accidents accordingly brought about by the city,such as personal economic losses,at the same time to ensure the personal safety of people.The prediction of traffic accidents is based on the analysis of the traffic accidents occurred in the past section and the further prediction of the state of traffic accidents in the future section after considering the relevant factors of traffic accidents.The factors affecting traffic accidents are often complicated and uncertain,including traffic flow,weather factors,points of interest(POI),road complexity factors,pedestrian and driver behavior factors and so on.Most of existing studies,using the method is the more traditional machine learning method or a single deep learning model prediction method,however,most of the traditional machine learning method is to use the grid to determine the forecast spatial unit,and ignores the influence of traffic accidents related to external factors,such as weather,POI,ignoring the traffic problems of zero inflation data,This results in poor prediction performance and low accuracy.In view of the above problems,this research adopts deep learning theory to build an urban traffic accident risk prediction model considering spatio-temporal characteristics,integrating real-time weather factors,POI and traffic flow characteristics.The improved ISTGCN network is used in the model.The main research contents include: Firstly,pre-processing and correlation analysis of multi-source heterogeneous traffic big data are carried out.Firstly,different relevant data affecting traffic accidents are collected,including traffic accident data,traffic flow data,weather data and POI data.Secondly,a series of data pre-processing operations are carried out according to the space-time correlation of different data.Including missing value supplement,delete redundant value,and data normalization operation.Second,a deep learning network framework based on improved spatio-temporal graph convolution is proposed.Firstly,graph convolution network(GCN)is used to extract spatial related attributes,and standardized normalized layer(BN)is added to solve the problem of gradient disappearance and gradient explosion.Secondly,gated linear element(GLU)is used to implement one-dimensional convolution operation in time dimension,and time correlation features are extracted.Finally,GCN and GLU module added one-dimensional convolution are combined into a space-time convolution module(ST-Con V),and the spatio-temporal characteristics are extracted,and then input into the fusion prediction model for final prediction.The mean square MSE loss function is used to solve the problem of zero expansion of traffic big data samples.The ISTGCN model in the research can effectively capture the comprehensive spatio-temporal correlation of traffic accidents.Firstly,the model is trained based on the real urban traffic accident data set in the UK,and the accident risk is taken as the output of the model to conduct fusion prediction.Secondly,the proposed model was compared with the existing single neural network model,the traditional machine learning model and the combined prediction model.Compared with GLU model,SDCAE model improved by 28%,SDCAE model improved by 4.87%,and Conv LSTM model improved by 4.19%.The comparative experiments show that the proposed model has better comprehensive performance than the classical model.Finally,analysis of relevant variant models is made,including the influence of different model structures,activation functions and characteristics of different influencing factors on the overall prediction performance of the model,and the integrity of the model is verified successively.