Roger Boza (Computer Science)

Roger Boza (Computer Science)

DOE Fellow Roger Boza contributed to the AI work/implementation of Long Short-Term Memory (LSTM) for the analysis of drywell cooling fan failures during his 2019 summer internship at Idaho National Laboratory. This work was published as a spinoff from his previous effort which focused on predicting cooling fan failures. This most recent work was published at the Progress in Nuclear Energy journal Volume 130, December 2020, 103540 under the title “A RELAP5-3D/LSTM model for the analysis of drywell cooling fan failure”

Journal: Progress in Nuclear Energy
Volume 130, December 2020, 103540
Title: A RELAP5-3D/LSTM model for the analysis of drywell cooling fan failure

Link: https://www.sciencedirect.com/science/article/pii/S0149197020302882

Abstract

A RELAP5-3D/LSTM model was created to analyze the failures of two drywell cooling fans at a nuclear power plant. A total of four fan coil units (FCUs) each comprised of a water-cooled heat exchanger and a centrifugal fan located in the drywell provide cooling via a closed nitrogen loop to the primary containment of the boiling water reactor. A Reactor Excursion and Leak Analysis Program (RELAP5-3D) thermal hydraulic model was created to simulate the steady-state normal operation of the FCUs. Historical data from the plant Process Information (PI) system was synchronized in time for a total of 33 Plant Management Information System (PMIS) tags per FCU representing: (1) the temperatures at various locations within the drywell, (2) inlet, outlet, and dewpoint temperatures at the FCUs, (3) reactor power, and (4) water coolant flowrate and temperature. Because the inlet temperature sensor for the two fans that failed did not provide consistent data prior to the failures, a long short-term memory (LSTM) recurrent neural network was trained to predict the FCU inlet temperature history based upon the states of the other valid PMIS points. RELAP5-3D simulations were performed using the measured FCU inlet temperatures, as well as the LSTM-generated temperatures, and the resulting FCU outlet temperatures were compared. The simulation results using the measured and predicted FCU inlet temperature were shown to be within 7.35% and 5.16%, respectively, of the values reported by the PI system. Thus, a viable approach has been demonstrated to predict the expected FCU outlet temperature. By comparing real-time measurements of FCU outlet temperature with predictions such as those presented here, off-normal operation can be readily detected. The use of RELAP5-3D with the LSTM results was successfully implemented to prototype a physics-based anomaly detection model for the drywell FCUs.