Dynamic probabilistic risk assessment of nuclear power plants using multi-fidelity simulations

Dynamic probabilistic risk assessment (PRA) more explicitly treats timing issues and stochastic elements of risk models. It extensively resorts to iterative simulations of accident progressions for the quantification of risk triplets including accident scenarios, probabilities and consequences. Dynamic PRA leverages the level of detail for risk modeling while intricately increases computational complexities, which result in heavy computational cost. This paper proposes to apply multi-fidelity simulations for a cost-effective dynamic PRA. It applies and improves the multi-fidelity importance sampling (MFIS) algorithm to generate cost-effective samples of nuclear reactor accident sequences. Sampled accident sequences are simulated in a parallel manner by using mechanistic code, which is treated as a high-fidelity model. Adaptively trained by using high-fidelity data, low-fidelity model is used to predicting simulation results. Interested predictions with reactor core damages are sorted out to build density functions of the biased distribution for importance sampling. After when collect enough number of high-fidelity data, risk triplets can be estimated. By solving a demonstration problem and a practical PRA problem by using MELCOR 2.2, the approach has been proven to be effective for risk assessment. Comparing with previous studies, the proposed multi-fidelity approach provides comparative estimation of risk triplets, while significantly reduces computational cost.