Machine learning of fire hazard model simulations for use in probabilistic safety assessments at nuclear power plants

Abstract

This study explored the application of machine learning to generate metamodel approximations of a physics-based fire hazard model. The motivation to generate accurate and efficient metamodels is to improve modeling realism in probabilistic safety assessments where computational burden has prevented broader application of high fidelity models. The process involved scenario definition, generating training data by iteratively running the fire hazard model called CFAST over a range of input space using the RAVEN software, exploratory data analysis and feature selection, an initial testing of a broad set of metamodel methods, and finally metamodel selection and tuning using the R software.Twenty-five metamodel methods ranging in class and complexity were investigated. Linear models struggled because the physics of fire are non-linear. A k-nearest neighbor (kNN) model fit the vast majority of calculations within ±10% for maximum upper layer temperature and its timing.The resulting kNN model was compared to an algebraic model typically used in fire probabilistic safety assessments. This comparison illustrated the potential of metamodels to improve modeling realism over simpler models selected for computational feasibility. While the kNN metamodel is a simplification of the higher fidelity model, the error introduced is quantifiable and can be explicitly considered.