Using machine learning to forecast and assess the uncertainty in the response of a typical PWR undergoing a steam generator tube rupture accident
Tran Canh Hai Nguyen, Aya Diab
Abstract
In this work, a multivariate time series machine learning (ML) meta-model is developed to forecast the transient response of a typical nuclear power plant (NPP) undergoing a steam generator tube rupture (SGTR). The model is based on Recurrent Neural Networks (RNNs); specifically, the Long-Short-Term-Memory (LSTM), Gated Recurrent Unit (GRU), and a hybrid model combining a convolutional neural network (CNN) with LSTM. To train the model, a database of the plant response is generated using the Best Estimate Plus Uncertainty (BEPU) methodology. The best estimate system code, RELAP5/MOD3.4, coupled with the statistical tool, DAKOTA is used to predict the thermal-hydraulic response under various operational uncertainties. The developed ML models successfully captured the inherent characteristics of the data with reasonable accuracy as evidenced by the performance metrics (MSE, MAE, RMSE and R2). Overall, LSTM outperforms GRU; while the hybrid CNN-LSTM model is computationally the most efficient. This work serves as a step towards gaining better understanding of the capabilities and limitations of ML models applied to nuclear safety in the hopes of expanding their implementations to more severe accident scenarios where operators are subjected to extreme stress and are therefore prone to error. If successful, ML models can offer support to operators or even act as expert systems to aid in the decision-making process while simultaneously minimizing the chances of human error.