Interpretable time series forecasting of NPP parameters in accident scenarios

Abstract

Nuclear power plant operators are limited in their capacity to analyze large amount of data in a short period, during a transient or accident scenario. Due to recent advances in machine learning methods, advanced algorithms have become available, which can support the operator with real-time analysis and forecast of important NPP parameters. The aim of this study was to provide short-term forecasts of key NPP parameters, which operators use to diagnose reactor states. The study focused on forecasting pressurizer level and subcooling margin, measured between the pressurizer and RCS hot leg, by using a state-of-the-art supervised learning algorithm, the Temporal Fusion Transformer. The model was trained with data from various transient and accident scenarios. The data was generated using an efficient basic principles PWR simulator called the BasicSim-PWR. The models demonstrated the capacity to differentiate between the small break LOCA and SGTR by predicting the safety trip of the SGTR scenario. The interpretable nature of Temporal Fusion Transformer gave insight into model behavior and potentially useful correlations for managing accident scenarios by observing few key inputs. The future research is directed towards advancing the scenario identification capability and efficiency evaluation of operator actions by forecasting the power plant response.