Safety Evaluation of a Multiple Steam Generator Tube Rupture Accident in a Typical Nuclear Power Plant using Explainable AI

In this study, three machine learning (ML) models were developed to predict the progression of a Multiple Steam Generator Tube Rupture (MSGTR) accident in the APR1400 reactor. To train the models, a database of system responses was generated by coupling, a thermal hydraulic code (RELAP5/SCDAP/MOD3.4) to the DAKOTA software using a Python script to enable accident analysis using the Best Estimate Plus Uncertainty (BEPU) approach. The uncertainty analysis focused on parameters critical to safety, including Rector Coolant System (RCS) pressure, RCS temperature, Reactor Vessel Upper Head (RVUH) void fraction, MSSVs mass flow, as well as a set of operator mitigation actions. The BEPU analysis demonstrated that the applied mitigation strategy is sufficient to ensure the plant's safety during the accident. Among the three models: Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), and Convolutional Neural Network combined with Long Short-Term Memory (CNN+LSTM), the hybrid CNN+LSTM achieved the best performance in predicting key safety parameters for the accident with an average accuracy of 99.27% and a training time of 7.0 ms/step. To further enhance the model credibility, the integrated gradients (IG) method was deployed to explain the AI prediction process.