The article considers the issues of predictive analysis of time series of accident rates and factors influencing the risks of road incidents using a recurrent neural network model with a transformer, multi-head attention mechanisms, and output LSTM layers. In the course of the study, a methodology for training the model and its application for predicting the activation time of critical incidents in the road transport environment was developed. The article presents the results of training and research of model variants, choice of structure and selection of training parameters in order to minimize the forecast error. As a result of the research, the following were determined: the number of training epochs, the number of neurons in the output LSTM layers, the batch size and the number of heads in the layers of internal attention of the transformer. The goal of the project is to predict the moments of time when the risks of critical events exceed the permissible level, as well as to determine the place of activation of the incident, taking into account the set of influencing factors. The results of risk and incident time forecasting are proposed to be displayed as a heat map of dangerous road sections.