Application of artificial neural networks in estimation of probable accident causes in nuclear power plants

An exploratory study was made to train an ANN to predict the probabilities of four causes of accidents (loss of coolant accident types), based on the time behaviour of three selected parameters (pressurizer pressure and level and containment humidity). 316 time scenarios have been generated, with the cause probabilities calculated by Bayesian procedures. Four ANN models were trained by the TURBO-NEURON 1.1 software package on a basis of 251 cases to predict the cause probabilities of a particular accident scenario. The generalization capacity of the models was tested by comparing the results of the remaining 65 cases. It was found that the ANN models were able to classify correctly the original cause in 88% of the test cases, while the probability of correct prediction by random guess is 25%. This result is considered quite encouraging for further work, in view of the possibility to increase easily the number of training cases, with a consequent increase in the generalization capability. 19 refs, 1 fig