TY - GEN
T1 - Predicting PWR Fuel Assembly CIPS Susceptibility with Convolutional Neural Networks
T2 - 2024 International Conference on Physics of Reactors, PHYSOR 2024
AU - Furlong, Aidan
AU - Alsafadi, Farah
AU - Palmtag, Scott
AU - Godfrey, Andrew
AU - Hayes, Stanley
AU - Wu, Xu
N1 - Publisher Copyright:
© 2024 AMERICAN NUCLEAR SOCIETY. All rights reserved.
PY - 2024
Y1 - 2024
N2 - Crud-induced power shift (CIPS) is an important operational concern in Pressurized Water Reactors (PWRs), as it creates a negative axial offset due to the deposition and precipitation of metal species and boron on fuel rod cladding.Previous studies have been primarily focused on modeling the CIPS phenomenon using mechanistic, bottom-up approaches that are computationally expensive and do not involve measurement data from operating cores.This paper describes a top-down Convolutional Neural Network (ConvNet)-based machine learning approach that has been trained with measurements from Unit 1 of the Catawba Nuclear Station in addition to calculated core design information.Using this classification method, the model obtained an overall accuracy of 92.1% with a correlation coefficient of 0.737 and has been shown to effectively predict the onset and peak of CIPS instances in a CIPS-prevalent cycle while avoiding false positive predictions in a clean cycle.Monte Carlo dropout was then applied to quantify the prediction uncertainties for selected time steps.The presented ML model can serve as a fast and accurate approach that could be of potential use in core design to assist in predicting and avoiding the CIPS phenomenon.
AB - Crud-induced power shift (CIPS) is an important operational concern in Pressurized Water Reactors (PWRs), as it creates a negative axial offset due to the deposition and precipitation of metal species and boron on fuel rod cladding.Previous studies have been primarily focused on modeling the CIPS phenomenon using mechanistic, bottom-up approaches that are computationally expensive and do not involve measurement data from operating cores.This paper describes a top-down Convolutional Neural Network (ConvNet)-based machine learning approach that has been trained with measurements from Unit 1 of the Catawba Nuclear Station in addition to calculated core design information.Using this classification method, the model obtained an overall accuracy of 92.1% with a correlation coefficient of 0.737 and has been shown to effectively predict the onset and peak of CIPS instances in a CIPS-prevalent cycle while avoiding false positive predictions in a clean cycle.Monte Carlo dropout was then applied to quantify the prediction uncertainties for selected time steps.The presented ML model can serve as a fast and accurate approach that could be of potential use in core design to assist in predicting and avoiding the CIPS phenomenon.
KW - Convolutional Neural Networks
KW - Crud
KW - Machine Learning
KW - Uncertainty Quantification
UR - http://www.scopus.com/inward/record.url?scp=85200268836&partnerID=8YFLogxK
U2 - 10.13182/PHYSOR24-43523
DO - 10.13182/PHYSOR24-43523
M3 - Conference contribution
AN - SCOPUS:85200268836
T3 - Proceedings of the International Conference on Physics of Reactors, PHYSOR 2024
SP - 1684
EP - 1693
BT - Proceedings of the International Conference on Physics of Reactors, PHYSOR 2024
PB - American Nuclear Society
Y2 - 21 April 2024 through 24 April 2024
ER -