Investigation of Long Short-Term Memory-Based Models for the Forecasting and Diagnosis of PWR LOCA Progression

Abstract In the aftermath of the Fukushima-Daiichi accident, there has been a continuous effort to improve the existing LOCA management tools. One of the ways to improve this is to help the operators assess the situation more accurately by providing them with real-time LOCA diagnostic tools. The potential of Long Short-Term Memory (LSTM) models in LOCA forecasting and real-time diagnosis was investigated. The LSTM models were trained and evaluated using simulated LOCA scenarios from RELAP5/SCDAPSIM. LSTM forecasting was first verified through the accurate prediction of the temporal progression of the water level, primary coolant inlet flow rate, and hot leg temperature after the onset of LOCA. Then, the study established the capability of LSTM to diagnose the pipe break location and extent of pipe damage using small portions of data near the onset of LOCA. Lastly, the study demonstrated the potential of realtime LOCA diagnosis by combining the LSTM-based forecasting and predictive models.