Prediction of LOCA break sizes using LSTM architecture for pressurized water reactors

Abstract Neural networks (NNs) and deep learning have revolutionized several fields, and nuclear safety analysis is no exception. The proper operation of nuclear reactor safety systems is crucial and is designed to meet strict safety requirements. Such systems are expected to withstand certain postulated accidents known as design basis accidents (DBAs), which include the loss of coolant accident (LOCA). As the LOCA involves complex fluid mechanics and heat transfer, the pattern recognition abilities of NNs allow for excellent prediction capabilities and the bypassing of otherwise tedious conventional analysis methods. This work investigates the deep learning techniques through long short-term memory (LSTM) architecture, for its ability to deal with time-series problems which include LOCAs. The utilized model is taught to estimate the size of pipe breaks within the cooling system based off of the corresponding pressure drops. A range of 0.5 %–100 % break-size-time-variant parameters were collected using the WSC Inc. 1,400 MWe generic pressurized water reactor (GPWR) simulator, using two circulation loops. Neural networks were trained on parameters such as loop temperature, pressure, containment pressure and Boron concentration. The performance of the LSTM model showed a mean absolute error (MAE) of 5.185, mean squared error (MSE) of 76.50, root mean squared error (RMSE) of 7.953, R 2 of 0.888 and Accuracy of 80.684 % within a tolerance of 15 % across 100 runs.