BLEVE risk effect estimation using the Levenberg-Marquardt algorithm in an artificial neural network model

Abstract

With the advancement of sciences such as machine learning, deep learning, and artificial intelligence, various algorithms are designed and developed. Learning and application models based on data types such as sensor data and databases of computers are made. The BLEVE effects, one of the most common types of fire in industries, are predicted using the Levenberg-Marquardt algorithm, which has become increasingly popular in recent years. Here, the equations in the TNO (Toegepast Natuurwetenschappelijk Onderzoek-Netherlands Applied Scientific Research Organization) (YellowBook Static) model of BLEVE are used.The aim of this research is to use artificial neural networks to predict the risk size of the BLEVE event. All the results from the TNO model of BLEVE effects were estimated using an artificial neural network model. Without utilizing equations, outcomes that are close to true results could be estimated in this method. Furthermore, findings were acquired fast, with linear outcomes in many settings. For this reason, the study's necessity and significance have shifted in this direction. Sixteen TNO model equations from BLEVE are applied, and heat flux values are calculated as a result.As a result of the studies, the BLEVE effects were predicted by the artificial neural network model created using the Levenberg-Marquardt algorithm. It is seen that the estimated results and the actual results calculated are close to each other. The statistical values between the predicted results of the artificial neural network model created by the Levenberg-Marquardt algorithm and the actual results were examined. The average relative error between the last stage of the BLEVE model created with ANN and the actual values was 3.34%. In addition, it has been observed that when the training iteration is increased within the algorithm, it gets closer to the real results, and statistical values such as standard error decrease even more. In other words, the computation was done by increasing the iteration sequence processing cycle of the result estimations and repeated with the number of training iterations in order to better evaluate the network's performance. It has been observed that as the number of iterations increases, closer and more realistic results emerge.