Uncertainty Analysis Method Of Reactor Nuclear Design Software Based On Deep Learning

Uncertainty analysis is a science to quantitatively characterize and reduce uncertainty in computing applications.All computer programs,analysis methods and nuclear power plant models used in reactor safety analysis must fully consider various uncertainties.The traditional uncertainty analysis method needs to randomly sample the uncertain input parameters to generate multiple groups of input files,which can be used as input to repeatedly run the nuclear design software and analyze the uncertainty,which leads the process to be unacceptable time-consuming.In order to improve the efficiency of uncertainty analysis of reactor nuclear design software,this paper introduces the surrogate model method,in order to establish a high fidelity agent model based on Deep Gaussian Process regression to replace the step of repeated operation of nuclear design software.In this paper,Python is used as the programming language,and the advanced deep learning model frameworks are applied.Firstly,using datasets calculated by nuclear calculate software package CASMO as input data,the single-layer Gaussian process model and Deep Gaussian Process model are deeply studied and programmed.The model effects of the two models are evaluated by using the cross validation method and taking the root mean square error as the evaluation standard.The advantages and disadvantages of the model are analyzed,and the influence of the number of training sets on the model is studied.Comparing the results of the two models,the prediction ability of the deep learning model is significantly better than that of the single-layer model,especially for higher-dimensional data.In order to solve the existing problems of Deep Gaussian Process,based on the advanced scientific researches related to mathematics and computer science,this paper theoretically deduces and develops the models based on Doubly Stochastic Variational Inference for Deep Gaussian Process and Deep Sigma Point Process,which respectively solve the unreasonable assumption of interlayer independence and the degradation of model prediction ability.This paper also further optimizes the optimization algorithm and interlayer assumptions.The cross validation results show that the two models can obtain better results than the ordinary Deep Gaussian Process for both datasets.For low dimensional data,double random variational inference Gaussian process can significantly reduce the model error.Based on the trained model,the uncertainty analysis process framework is programmed to realize the functions of data input,model interface,sensitivity analysis and uncertainty analysis result output.The corresponding data conversion interface and model switching interface are developed.The framework is applied to analyze the uncertainty of the influence of four regional enrichment on the power peak factor and the influence of the whole influence factor on the power coefficient.The uncertainty calculation results of traditional design software is used as the benchmark to verify the availability of the framework.The results show that the high fidelity agent model based on Deep Gaussian Process can precisely predict the relationship between input and output,obtain more groups of data in a significantly shorter time,and calculate the uncertainty results.Surrogate model can significantly improves the uncertainty analysis efficiency of reactor nuclear design software.