INVESTIGATION OF SURROGATE MODEL FOR UNCERTAINTY QUANTIFICATION OF NUCLEAR SYSTEM

抄録

Uncertainty Quantification(UQ) is widely used in the design and safety assessment of nuclear reactor system. As for UQ process, including forward UQ and inverse UQ, large quantities of numerical simulations are required for the uncertainty propagation which is very time consuming. To reduce the time costs, surrogate model is normally used to replace the calculation of reactor programs and with the development of technology, high precision surrogate model is constantly optimized by recent researchers. In this paper, three classical surrogate models which are Radial Basis Function (RBF) artificial neural network, Kriging, Support Vector Machine (SVM) regression and two optimized models which are Ensemble Surrogate (ES) model and Particle Swarm Optimization (PSO) Kriging model are introduced. These models are applied to different numerical examples with various numbers and distribution of parameters and with different complexity and degree of nonlinearity. Accuracy, efficiency and robustness are analyzed and compared with each models. Then surrogate models are tested with thermal hydraulic case of reflooding based on FEBA experiment. The results indicate that with small samples, PSO-Kriging model has the highest accuracy and robustness and has broad applicability for different cases. Thus, it is adopted to build the surrogate model of a typical Small Modular Reactor(SMR) for future process of uncertainty propagation, sensitivity or reliability analysis to replace the expensive models and reduce the computational cost.