(A) study of reinforcement learning implementation and building reinforcement learning environment using supervised learning for generating nuclear power plant accidents scenario

Abstract

Accident scenarios for the study of existing nuclear power plant accident phenomena are deterministically selected through probabilistic safety analysis (PSA). Although it is possible to determine the sequence of failures of a particular component in this process, it is difficult to assess the exact timing of the component failure and the change and severity of the accident process resulting from it. MAAP, which is the existing nuclear accident analysis code, can also confirm the progress of the accident by specifying the malfunctioning component and the time for failure in the input value. In order to confirm the high-risk device failure time, it is necessary to set all the failure time of various components before the analysis and then interpret the analysis results. In this study, it is confirmed whether this process can be replaced by using reinforcement learning. Through the process of comparing whether the most serious accident scenario obtained via MAAP simulation is the same as the accident determined by the reinforcement learning agent, the possibility of generating a new nuclear accident scenario through reinforcement learning can be evaluated. However, because it is difficult to apply the reinforcement learning methodology to the existing MAAP code directly, a surrogate model was created through supervised learning and reinforcement learning was performed by interacting with the MAAP surrogate model. After evaluating how accurately the MAAP code can be represented with the surrogate model, it was evaluated whether the reinforcement learning agent can determine the most severe scenario identified from the MAAP data-driven model while various reward settings were tested. From this process, a methodology to apply reinforcement learning to the existing nuclear accident analysis code, which is difficult to use reinforcement learning, is newly presented, and factors to be considered in order to use this methodology are identified and validated.