Application of surrogate models for statistical safety evaluation. Investigation of applicability to uncertainty analysis for PWRs

Application of a surrogate model for statistical safety evaluation was discussed in the present paper on uncertainty analysis by the RELAP5 code for loss of coolant accidents in PWRs. Cross validation techniques were applied to verify the prediction capability of the surrogate model for the 95th percentile values of the safety evaluation parameters. Comparison with the RELAP5 uncertainty analysis results confirmed that the generalized performance of the surrogate model for the 95th percentile value prediction could be estimated using the cross validation on the training data set. From the results it was concluded that the surrogate model whose prediction accuracy for the 95th percentile values was confirmed by the cross validation could be used for the statistical safety evaluation. The effectiveness of the adaptive sampling technique was verified to improve the prediction accuracy for the 95th percentile values using a small number of training data. Comparison with the RELAP5 uncertainty analysis results confirmed that the prediction error of the 95th percentile values converged by repeating the adaptive sampling and the prediction errors were able to be estimated by the cross validation on the training data set. Therefore, the surrogate model with high accuracy for the 95th percentile value prediction was able to be effectively constructed by improving the surrogate model using the adaptive sampling procedure with the target of the 95th percentile values and confirming the model prediction accuracy using the cross validation on the training data set. (author)