DEVELOPMENT AND RESEARCH OF ALGORITHMS FOR FORECASTING FIRE HAZARDOUS SITUATIONS

Abstract

Early detection of fire hazard situations is a critical aspect of ensuring safety, as it helps to minimize the risk of material and human losses. Early detection of threats helps to preserve material assets, reduce the time for their restoration and, more importantly, save human lives. In this regard, a new approach to predicting fire hazard situations is proposed: an algorithm for training a model for predicting fire hazard situations, as well as an algorithm for predicting fire hazard situations, which are developed on machine learning models such as recurrent neural networks, random forest, optimization trees, autoregressive neural networks, etc. The study proposes to consider algorithms for predicting fire hazard situations developed on the basis of an analysis of existing forecasting algorithms, including methods based on machine learning, statistical models and simulation approaches, taking into account their advantages and disadvantages, accuracy indicators. The results of the study of the developed algorithms show that they are capable of predicting the outside temperature value of the sensor with an accuracy of 93.33% based on the test data from a complex of interconnected fire sensors, with errors of MAE = 1.72, MSE = 2.95 in the abnormal mode on the test data, and with an accuracy of 92.85% for the temperature inside the sensor, errors MAE = 1.66, MSE = 2.75. The accuracy on the test data in the normal mode for the outside temperature was 96.27%, errors MAE = 1.22, MSE = 1.48, and the accuracy of predicting the inside temperature was 96.16%, errors MAE = 1.24, MSE = 1.53. For the test sample of 500,000 readings, the errors of the predicted outside temperature were: MAE = 1.82, and MSE = 3.31, and the accuracy was 91.78%. The errors of the predicted temperature inside (temp2_inside) were: MAE = 1.89, and MSE = 3.57, and the accuracy was 91.35%.