Development of a Machine Learning Method to Predict the Break Diameter during PWR Loss-of-Coolant Accident

Abstract

This paper describes a method of predicting the break diameter in a loss-of-coolant accident of a pressurized water reactor with machine learning using the data obtained by the safety parameter display system. From the variation in reactor coolant pressure, two feature data, the time difference between LOCA and the time when pressure shows the minimum rate, and the mean pressure decrease rate are determined. The programming language MATLAB is used to extract these feature data and learn them by Gaussian process regression. There are some dispersion factors, such as the break location on the reactor coolant pipe and 1 min sampling timing. The learning results by GPR show relative errors of 7.8% for a 4-loop plant and 9.5% for a 3-loop plant. The learning results are valid for the delay times of 30 and 60 min from the reactor shutdown to LOCA. If the break diameter expands during LOCA, the predicted diameter is useful as input for the plant simulation.