Machine Learning of Hazard Model Simulations for use in Probabilistic Risk AssessmentsWorrell, Clarence (2018) Machine Learning of Hazard Model Simulations for use in Probabilistic Risk Assessments. Master's Thesis, University of Pittsburgh. (Unpublished)
AbstractThis study explored the use of machine learning to generate metamodel approximations of a physics-based fire hazard model called Consolidated Fire and Smoke Transport (CFAST). The motivation to generate accurate and efficient metamodels is to improve modeling realism in probabilistic risk assessments where computational burden has prevented broader application of high fidelity codes. The process involved scenario definition, generating training data by iteratively running the hazard model over a range of input space, exploratory data analysis and feature selection, an initial testing of a broad set of metamodel types, and finally metamodel selection and tuning. The study identified several factors that should be considered when metamodeling a physics-based computer code. First, the input space should be limited to a manageable scale and number of parameters; otherwise generating sufficient training data becomes infeasible. Second, there is a relationship between the physics being characterized and the metamodel types that will successfully mimic those physics. Finally, metamodel accuracy and efficiency must be balanced against initial development costs. Once developed, trained metamodels are portable and can be applied by many users over a range of modeling conditions. The Idaho National Laboratory software called RAVEN was used to facilitate the analysis. Twenty five (25) metamodel types were investigated for their potential to mimic CFAST-calculated maximum upper layer temperature and its timing. Linear metamodels struggled to predict with accuracy because the physics of fire are non-linear. k-nearest neighbor (kNN) model tuning generated a k =4 model that fit the vast majority of CFAST calculations within 10% for both maximum upper layer temperature and its timing. This model showed good generalization with use of 10-fold cross validation. The resulting kNN model was compared to algebraic models typically used in fire probabilistic risk assessments. The algebraic models were generally conservative relative to CFAST; whereas the kNN model closely mimicked CFAST. This illustrates the potential of metamodels to improve modeling realism over the simpler models often selected for computational feasibility. While the kNN metamodel is a simplification of the higher fidelity CFAST code, the error introduced is quantifiable and can be explicitly considered in applications of the metamodel. Share
Details
MetricsMonthly Views for the past 3 yearsPlum AnalyticsActions (login required)
|
![[feed]](/images/feed-icon-32x32.png){kind=icon}