Uncertainty-aware prediction of Peak Cladding Temperature during extended station blackout using Transformer-based machine learning

Abstract

Accurate prediction of the Peak Clad Temperature (PCT) may be used to evaluate the efficacy of operator mitigation actions during extended Station Blackout (SBO) scenarios. In this study, we propose a two-stage machine learning (ML) framework that integrates classification and regression to forecast PCT. While the classification stage identifies whether mitigation efforts succeed or fail, the regression stage provides precise multi-step PCT predictions. Our framework leverages advanced ML models, including Transformer architectures, Attention mechanism, and Long Short-Term Memory (LSTM) networks, alongside the Best Estimate Plus Uncertainty (BEPU) approach. To account for the underlying uncertainty and generate confidence intervals, we incorporate Monte Carlo (MC) Dropout. By integrating BEPU with machine learning and uncertainty quantification, our model produces reliable temperature forecasts despite the system’s inherent complexity and nonlinearity with R² values exceeding 0.98 for 60-, 120-, and 240-step time frames. Notably, the LSTM-Transformer model proves to be the most effective, even for longer prediction horizons. The developed framework serves as a powerful real-time decision support tool for operators, for accurate prediction and effective mitigation of critical conditions like extended SBO events.