An advanced computational framework for the inverse uncertainty quantification of thermal-hydraulic code applications for the analysis of passive safety systems

Over the last decades, passive safety systems have been gathering a noticeable international interest playing an essential role in tackling the safety-related design goals of advanced Nuclear Power Plants (NPPs). A crucial challenge, still to be deal with, is the quantification of uncertainty in passive safety systems reliability assessments. In particular, uncertainty may be present in some input parameters of the Thermal-Hydraulic (T-H) models used to predict the passive system behavior, which is typically quantified by “expert judgment”. The purpose of the present thesis is to develop an advanced Inverse Uncertainty Quantification (IUQ) framework to support expert judgment in the assessment of the (epistemic) uncertainty affecting T-H model input parameters. IUQ is implemented to calibrate the Probability Density Functions (PDFs) of some selected input parameters, based on experimental field data and on the use of a Best Estimate (BE) code results, within a Bayesian framework. The proposed IUQ approach is applied to a RELAP5-3D model of the PERSEO (In-Pool Energy Removal System for Emergency Operation) experimental facility, located at SIET laboratory (Piacenza, Italy), for which time-series measurement data of the heat exchanger power are available. Principal Component Analysis (PCA) is applied for reducing the dimensionality of the problem, and fast-running Kriging metamodels are implemented for the emulation of the RELAP5-3D behavior at a lower computational cost. However, in the presence of nonlinear and noisy data, PCA may perform poorly. To address this issue, in this thesis we also propose a novel approach based on Stacked Sparse Autoencoders (SSAE), which are Artificial Neural Networks (ANNs) used for nonlinear dimensionality reduction. A comparison between PCA and SSAE for dimensionality reduction is carried out, providing some insights regarding the technical issues associated with the implementation and the use of both the techniques for the proposed IUQ approach. The comparison reveals that Stacked SAE might represent a promising alternative to PCA in case of noisy data. Finally, the empirical distributions obtained through both techniques are used to characterize the PDFs of the BE model uncertain input parameters and to compute the related summary statistics (i.e., mean values, modes, 5th percentiles, 95th percentiles, correlation matrices, and posterior marginals). Keywords: Inverse uncertainty quantification, Bayesian Inference, Surrogate modeling, Kriging, Principal component analysis, Stacked Sparse Autoencoders, Passive safety systems, PERSEO facility, RELAP5-3D