Machine Learning-Based Classification and Regression Approach for Early Detection of Large Break Loss-of-Coolant Accident Conditions

Data-driven predictive approaches have significant potential for enhancing accident management and operator support in nuclear power plants (NPPs). Machine learning (ML)-based models can effectively monitor early signs of accidents and define initial conditions, allowing for better responses and reducing the need for manual intervention. This study explores using ML algorithms to identify the simulation-based initial conditions of accidents in a generic pressurized water reactor (GPWR), focusing on large loss-of-coolant accidents (LLOCAs) in primary cooling systems. Fuel temperature datasets were generated from the GPWR simulator for different LLOCA conditions, including 6 locations and break sizes of 0.11 m² to 0.22 m². The dataset includes 713,196 observations of fuel temperature from 12 zones of the reactor vessel for 36 LLOCA conditions within a 30-min interval. The classification and regression tree (CART) models were then used to predict the location and size of the break. Following post-learning and validation, CART model performed reasonably well with R² value of ~ 0.9 for break size prediction and a misclassification error of ~ 4% for break location. For performance evaluation, 12 new datasets with varying interpolated break sizes for the 6 locations were introduced, targeting the early detection during the time intervals of up to 10 min. The CART model in this study demonstrated acceptable performance in early identifying the break size and location of LLOCA conditions, with accuracy rates > ~ 70% and relative errors < ~ 6%. This approach aims to enhance emergency management and operator support to ensure timely responses and safe operations in NPPs.