DEEP LEARNING NEURAL NETWORK MODEL FOR PREDICTING RISKS AND TIMING OF ROAD INCIDENTS

Abstract

One of the ways to improve road safety is a predictive analysis of the risks of critical events depending on the dynamics of changes in accident rates, traffic violations, and the influence of objective and subjective factors. The analysis of accident rates and factors influencing the risks of road incidents is performed by studying the data presented in the form of time series using a deep learning neural network model. A type of recurrent neural network model with LSTM layers, a transformer, and a multi-headed attention mechanism is proposed as a tool for predictive analysis of time series. Time series are formed based on data extracted from photo radar systems for photo and video recording in the road transport environment. In the course of the study, a methodology for training a model for predicting the activation time of critical incidents in the road transport environment was developed through the analysis of time series of accident rates, traffic violations, and a number of objective influencing factors. The article presents the results of training and research of model variants for choosing its structure and selecting parameters in order to obtain an acceptable forecast error. The article presents studies of the influence of the number of training epochs, the number of neurons in the output LSTM layers, the batch size, the early stopping technique, and the number of heads in the attention layers on the forecast. The purpose of training and using a neural network is to predict the moments in time when the risks of critical events exceed the permissible level, as well as to determine the location of the incident activation taking into account the influencing factors, the combination and values of which can lead to accidents with visualization in the form of a heat map of hazardous areas.