Clustering and Uncertainty Analysis to Improve the Machine Learning-based Predictions of SAFARI-1 Control Follower Assembly Axial Neutron Flux Profiles
Authors:
Lesego Moloko,
Pavel Bokov,
Xu Wu,
Kostadin Ivanov
Abstract:
The goal of this work is to develop accurate Machine Learning (ML) models for predicting the assembly axial neutron flux profiles in the SAFARI-1 research reactor, trained by measurement data from historical cycles. The data-driven nature of ML models makes them susceptible to uncertainties which are introduced by sources such as noise in training data, incomplete coverage of the domain, extrapola…
▽ More
The goal of this work is to develop accurate Machine Learning (ML) models for predicting the assembly axial neutron flux profiles in the SAFARI-1 research reactor, trained by measurement data from historical cycles. The data-driven nature of ML models makes them susceptible to uncertainties which are introduced by sources such as noise in training data, incomplete coverage of the domain, extrapolation and imperfect model architectures. To this end, we also aim at quantifying the approximation uncertainties of the ML model predictions. Previous work using Deep Neural Networks (DNNs) has been successful for fuel assemblies in SAFARI-1, however, not as accurate for control follower assemblies. The aim of this work is to improve the ML models for the control assemblies by a combination of supervised and unsupervised ML algorithms. The $k$-means and Affinity Propagation unsupervised ML algorithms are employed to identify clusters in the set of the measured axial neutron flux profiles. Then, regression-based supervised ML models using DNN (with prediction uncertainties quantified with Monte Carlo dropout) and Gaussian Process (GP) are trained for different clusters and the prediction uncertainty is estimated. It was found that applying the proposed procedure improves the prediction accuracy for the control assemblies and reduces the prediction uncertainty. Flux shapes predicted by DNN and GP are very close, and the overall accuracy became comparable to the fuel assemblies. The prediction uncertainty is however smaller for GP models.
△ Less
Submitted 20 December, 2023;
originally announced December 2023.
Prediction and Uncertainty Quantification of SAFARI-1 Axial Neutron Flux Profiles with Neural Networks
Authors:
Lesego E. Moloko,
Pavel M. Bokov,
Xu Wu,
Kostadin N. Ivanov
Abstract:
Artificial Neural Networks (ANNs) have been successfully used in various nuclear engineering applications, such as predicting reactor physics parameters within reasonable time and with a high level of accuracy. Despite this success, they cannot provide information about the model prediction uncertainties, making it difficult to assess ANN prediction credibility, especially in extrapolated domains.…
▽ More
Artificial Neural Networks (ANNs) have been successfully used in various nuclear engineering applications, such as predicting reactor physics parameters within reasonable time and with a high level of accuracy. Despite this success, they cannot provide information about the model prediction uncertainties, making it difficult to assess ANN prediction credibility, especially in extrapolated domains. In this study, Deep Neural Networks (DNNs) are used to predict the assembly axial neutron flux profiles in the SAFARI-1 research reactor, with quantified uncertainties in the ANN predictions and extrapolation to cycles not used in the training process. The training dataset consists of copper-wire activation measurements, the axial measurement locations and the measured control bank positions obtained from the reactor's historical cycles. Uncertainty Quantification of the regular DNN models' predictions is performed using Monte Carlo Dropout (MCD) and Bayesian Neural Networks solved by Variational Inference (BNN VI). The regular DNNs, DNNs solved with MCD and BNN VI results agree very well among each other as well as with the new measured dataset not used in the training process, thus indicating good prediction and generalization capability. The uncertainty bands produced by MCD and BNN VI agree very well, and in general, they can fully envelop the noisy measurement data points. The developed ANNs are useful in supporting the experimental measurements campaign and neutronics code Verification and Validation (V&V).
△ Less
Submitted 15 November, 2022;
originally announced November 2022.