Machine learning-driven probabilistic seismic demand model with multiple intensity measures and applicability in seismic fragility analysis for nuclear power plants

Abstract

A novel machine learning (ML) framework is applied in mapping intensity measures (IMs) to seismic engineering demand parameter (EDP). It is an efficient and accurate approach to establishing the probabilistic seismic demand model (PSDM) with multiple IMs for nuclear power plant (NPP). The optimal IM set on the performance of the prediction is further discussed herein. A total of 33 IMs are used to represent the characteristics of ground motions (GMs) comprehensively and significant IMs are identified by recursive random forest (RRF). 14 algorithms are used for determining the most optimal ML model. The generalization ability of the proposed ML model is empirically tested against traditional models based on scalar and vector IMs in a new dataset. ML is found to have absolute superiority in the construction of PSDM with inelastic correlation. Moreover, a promising methodology for deriving the fragility curve is presented based on multiple stripe analysis (MSA) driven by ML. By training the model, the workload of nonlinear time-history analysis (NLTHA) can be replaced, and the large inelastic behavior can be captured considering the stiffness degradation of NPP when it is subjected to high-intensity effects. It is meaningful for the fragility assessment of NPP beyond a design-basis event.