Bayesian Uncertainty Quantification of Reflooding Model With PSO–Kriging and PCA Approach

• Published in: Science and technology of nuclear installations Vol. 2025; no. 1
• Main Authors: Zhang, Ziyue, Li, Dong, Wang, Nianfeng, Lei, Meng
• Format: Journal Article
• Language: English
• Published: New York John Wiley & Sons, Inc 01.01.2025; Wiley
• Subjects: Algorithms; Bayesian analysis; Calibration; Error reduction; Experiments; Heat; Hydraulic models; Hydraulics; Markov chains; Nuclear power plants; Nuclear reactors; Nuclear safety; Parameters; Particle swarm optimization; Principal components analysis; Uncertainty
• ISSN: 1687-6075; 1687-6083
• DOI: 10.1155/stni/5416943

To improve the process of best estimate plus uncertainty (BEPU) for nuclear safety assessment and calibration of thermal–hydraulic models for error reduction, inverse uncertainty quantification (IUQ) is proposed in recent years to quantify the uncertainty of model parameters in reactor program. As reflooding is a vital stage to cool the core and prevent serious accidents and uncertainties exist in the important results of the program because of the complexity of the phenomena, IUQ is performed for reflooding models in this study based on Bayesian theory and Markov chain Monte Carlo (MCMC) algorithm. In order to solve the problem of large time costs in sampling and inefficient use of transient sample points, particle swarm optimization (PSO)–Kriging model and principal component analysis (PCA) are adopted in this paper. Measurement of peak cladding temperature (PCT) and quench time from FEBA and FLECHT SEASET experiments supply data for evaluation and validation. Results show that PSO–Kriging model could well represent the system program with R 2 (R‐squared coefficient of determination) close to 1 and uncertainties assessed by the method could cover most of the time sequential experiment data. By comparing the methods with and without PCA, it indicates that the IUQ method utilizing PCA not only reduces input parameter correlation but also provides more accurate estimates of input parameter posterior distributions. Furthermore, the validation outcomes of mean value calibration show enhanced agreement with the experimental data.