Reliability assessment of passive EHRS of IRIS based on the AM-SIS-ANN method

Research on the reliability estimation of passive systems in nuclear power plants has been pivotal to ensuring nuclear power generation safety. Numerous analytical methods have been proposed to calculate the failure probability of passive systems, with the most notable ones combining fast computational surrogate models with efficient variance reduction sampling techniques. Nevertheless, these methods are predominantly based on deterministic failure criteria, posing a significant challenge in managing the uncertainty of such criteria. To address this problem and enhance the accuracy and efficiency of passive system reliability estimation, this paper introduces a segmented method that integrates Adaptive Metamodel Subset Simulation Importance Sampling (AM-SIS) with neural networks, abbreviated as AM-SIS-ANN. This method harnesses the strengths of both techniques: employing the AM-SIS method to calculate failure probabilities in low-probability regions to mitigate high prediction errors from neural networks in sparse data areas, and using neural networks to compute output distributions in other regions, thereby eliminating the need for additional AM-SIS executions and substantially reducing computation time. The proposed method’s effectiveness is illustrated and validated through two numerical examples. Finally, this method is applied to calculate the failure probability of the Emergency Heat Removal System (EHRS) of the IRIS reactor. For the natural circulation loop example, considering four failure criteria, the AM-SIS-ANN method achieved the shortest computation time, reducing it by approximately 23.4% compared to the AM-SIS method and by about 57.5% compared to the neural network method. These results demonstrate that the AM-SIS-ANN method, under uncertain failure criteria, can significantly reduce computational time while ensuring high accuracy.