A Machine Learning Diagnosis of the Severe Accident Progression

Abstract

We propose a machine-learned platform for the diagnosis of a severe accident progression in a nuclear power plant, where some of the signals are lost due to a harsh environment. To predict the lost signal using other plant parameters, a long short term memory (LSTM) network is proposed, where multiple accident scenarios can be accumulated and used for training and prediction. Training and test data were produced by MELCOR analysis of the Fukushima Daiichi Nuclear Power Plant (FDNPP) accident at unit 3. The MELCOR simulation data is presented as time series at uniform intervals. Feature variables were selected among many plant parameters, where the importance ranking was determined by a recursive feature elimination technique using RandonForestRegressor. We performed a sensitivity study of the different choices of training and test scenarios, the number of feature variables, and target variables on the prediction accuracy of the ML platform. Among eight MELCOR simulation results, different combinations of training and test data are selected in terms of similarity between them. It is found that when the number of feature variables is more than five, the proposed ML platform was able to predict not only similar test data but also unseen test data with reasonable accuracy when the number of training data increases. It is also found that the proposed ML platform can consistently predict any of the lost signals with reasonable accuracy. These findings suggest that the proposed ML platform will be quite helpful in diagnosing severe accident progression when some of the signal is corrupted.

Keywords: Machine Learning, LSTM, Recursive Feature Elimination, Fukushima Accident, Time Series