Surrogate Modelling of Risk Measures for Use in Probabilistic Safety Analysis Applications

Probabilistic Safety Analysis (PSA) is an efficient tool for assessing, maintaining and improving the Nuclear Power Plant (NPP) safety. In the literature, different PSA applications have been identified such as: PSA to support NPP testing and maintenance planning and optimization, PSA as a tool to monitor level of safety or PSA as a predictive evaluation of risk. In general, these applications require analyzing aging trends, updating reliability parameters and maintenance related to safety equipment. An indispensable tool in PSA is the software such as RiskSpectrum or CAFTA which are widely used in NPP. The main problem with the use of commercial software is its lack of flexibility in modelling. In this context, the use of tools like surrogate models, or metamodeling, emerges as a tool that can improve realism in probabilistic safety analysis. In this approach, the PSA code is substituted by a metamodel in order to obtain the risk measures of interest. In this paper, different metamodels have been considered which have been trained to predict the Core Damage Frequency (CDF). The performance of the different models is evaluated using three quality metrics (Root Mean Square Error, Mean Absolute Error and Mean Absolute Percentage Error) which have been evaluated using k-fold cross-validation technique. The results obtained demonstrate the capacity of the metamodels to provide accurate and computationally efficient estimates of CDF.

Keywords: Probabilistic Safety Analysis (PSA), Core Damage Frequency (CDF), Machine learning, Metamodel.

Download PDF