Online scenario labeling using a hidden Markov model for assessment of nuclear plant state

By taking into account both aleatory and epistemic uncertainties within the same probabilistic framework, dynamic event trees (DETs) provide more comprehensive and systematic coverage of possible scenarios following an initiating event compared to conventional event trees. When DET generation algorithms are applied to complex realistic systems, extremely large amounts of data can be produced due to both the large number of scenarios generated following a single initiating event and the large number of data channels that represent these scenarios. In addition, the computational time required for the simulation of each scenario can be very large (e.g. about 24h of serial run simulation time for a 4h station blackout scenario). Since scenarios leading to system failure are more of interest, a method is proposed for online labeling of scenarios as failure or non‐failure. The algorithm first trains a Hidden Markov Model, which represents the behavior of non‐failure scenarios, using a training set from previous simulations. Then, the maximum likelihoods of sample failure and non‐failure scenarios fitting this model are computed. These values are used to determine the timestamp at which the labeling of a certain scenario should be performed. Finally, during the succeeding timestamps, the likelihood of each scenario fitting the learned model is computed, and a dynamic thresholding based on the previously calculated likelihood values is applied. The scenarios whose likelihood is higher than the threshold are labeled as non‐failure. The proposed algorithm can further delay the non‐failure scenarios or discontinue them in order to redirect the computational resources toward the failure scenarios, and reduce computational time and complexity. Experiments using RELAP5/3D model of a fast reactor utilizing an Reactor Vessel Auxiliary Cooling System (RVACS) passive decay heat removal system and dynamic analysis of a station blackout (SBO) event show that the proposed method is capable of correctly labeling 100% of failure scenarios as failure and over 80% of non‐failure scenarios as non‐failure and provide significant simulation time savings.