A probabilistic model for online scenario labeling in dynamic event tree generation

Dynamic event trees provide a wide coverage of possible system evolution sequences (scenarios) and may lead to the simulation of thousands of scenarios following a single initiating event. The large number of scenarios can be a burden in terms of computational time and storage requirements. However, not all of the scenarios are equally significant. From a safety point, failure scenarios or the scenarios that lead to undesirable consequences are more important than the scenarios that represent normal system evolution (non-failure scenarios). A method is presented for online labeling of non-failure scenarios. Since the number of non-failure scenarios is usually much larger than that of failure scenarios, substantial computational savings could be obtained if the non-failure scenarios can be identified and not pursued by the simulator. First, the parameters of a hidden Markov model that represents the behavior of non-failure scenarios are learned using the examples of the non-failure scenarios. Next, the failure behavior with respect to the non-failure model is learned using sample failure scenarios. During the succeeding system simulations, a scenario is labeled as non-failure if its evolution path is more likely to fit the constructed model than the learned failure behavior. Experiments using RELAP5/3D model of a fast reactor utilizing an RVACS (Reactor Vessel Auxiliary Cooling System) passive decay heat removal system show that the proposed method is capable of correctly labeling over 85% of non-failure scenarios without mislabeling the failure scenarios and provide time savings of at least 55%. We also investigate the sensitivity of the proposed labeling scheme depending on the number of hidden states in HMM and the nature of the state variables used for scenario representation.