Abstract:
In order to accurately predict the seismic performance subjected to earthquakes beyond design basis and to ensure the safety and reliability of nuclear power plants, a novel machine learning framework is proposed to establish one probabilistic seismic demand model with multiple seismic intensity measures (IMs). 32 IMs are used to characterize earthquakes and a recursive random forest (RRF) method is used to determine the optimal subset of IMs. Machine learning algorithms are used to develop predictive models for the post-earthquake damage state and demand parameter response of nuclear power plants. The results show that the CatBoost algorithm could accurately assess the damage state, and that the GB and XGB algorithms are the most effective in predicting post-earthquake response of nuclear power plants. Machine learning has potential to achieve a high accuracy damage assessment and the response prediction for nuclear power plants, which can capture the inelastic behavior due to stiffness degradation under the effect beyond design basis earthquakes, and thus extends the power of the post-earthquake risk assessment system for nuclear power plants. Additionally, the results of the predictive model are interpreted by the Shapley additive explanations (SHAP) method, and the evolution of the impact of the dominant IMs is investigated with increasing seismic intensity. The research outcome could provide references for seismic design, for post-earthquake reinforcement and policy decision of actual nuclear power plants.