Use of Dynamic Event Trees and Deep Learning for Real-Time Emergency Planning in Power Plant Operation

Abstract

An initiating event that disrupts regular nuclear power plant (NPP) operation can result in a variety of different scenarios as time progresses depending on the response of standby safety systems and operator actions to bring the plant to a safe, stable state, or the uncertainties in accident phenomenology. Depending on the severity of the accident and potential magnitude of release of radioactive material into the environment, off-site emergency response such as evacuation may be warranted. An approach that could be used for real-time emergency guidance to support the declaration of a site emergency and to guide off-site response is presented using observable plant data in the early stages of a severe accident. The approach is based on the simulation of the possible NPP behavior following an initiating event and projects the likelihood of different levels of off-site release of radionuclides from the plant using deep learning (DL) techniques. Training of the DL process is accomplished using results of a large number of scenarios generated with the Analysis of Dynamic Accident Progression Trees/MELCOR/Radiological Assessment System for Consequence Analysis (RASCAL) computer codes to simulate the variety of possible consequences following a station blackout event (similar to the Fukushima accident) for a large pressurized water reactor. The ability of the model to predict the likelihood of different levels of consequences is assessed using a separate test set of MELCOR/RASCAL calculations.