Session: 07-04: Experiments and Analyses - III

Paper Number: 135404

135404 - Prediction of Critical Heat Flux of Wire-Wrapped Rod Bundle Based on Artificial Neural Network 

Abstract: 

Accurate prediction for the critical heat flux of rod bundle assemblies is important for reactor safety assessment. However, the existing prediction methods fail to unify their generalization and accuracy. In detail, the understanding of critical heat transfer is still limited, and the theoretical prediction models are unable to provide a satisfactory estimation. Otherwise, although the empirical methods which has been proposed for several decades could give result precisely, their suitable conditions are strictly limited, causing the endless development of different correlations for every specific rod bundle design. Since the mentioned problems, new methods have been investigated to overcome the difficulty that lies between prediction range and accuracy. Among them, the artificial neural network caught attention recently due to its excellent performance in predicting complex problems. For the reactor core safety analysis, the regulations require an evaluation of the critical prediction model, and it is asked to provide a penalized limit to cover the possible error uncertainty. When the theoretical or empirical models are replaced by an artificial neural network, the prediction accuracy is easily satisfied, but the error characteristic was seldom analyzed in the literature. Therefore, this study discusses the prediction performance of a traditional empirical method, a neural network method, and a hybrid method combined with traditional methods and machine learning algorithms. The hybird method takes an empirical prediction method as the basis result and uses the neural network method to modify the bias between the basis result and true value. In conclusion, the neural network method could be better than traditional methods in accuracy, but it fails to provide suitable results as it may sacrifice the estimating mean to improve the variance, causing the outcome lower overall. The statistic mean of the ratio (measurement/prediction) tends to be larger than 1, which means an underestimation of the critical heat flux. In the meanwhile, the hybrid method was proved as a better way to estimate the critical heat flux. It is found that the hybrid model can effectively improve the prediction performance and has consistent prediction ability in different steam quality regions. What is more, the statistical average ratio of predicted value and true value maintains well, indicating this suggested method could provide a result with a better error characteristic, and as a result a better critical heat flux estimation. The research will be helpful in improving the feasibility of the application of machine learning algorithms such as neural networks in the nuclear energy industry.

Presenting Author: Wei Zhang Shanghai Jiao Tong University

Presenting Author Biography: Wei Zhang is a 2021 doctoral student of The College of Smart Energy of Shanghai Jiao Tong University, majoring in Energy and Power Engineering. He graduated from Shanghai Jiao Tong University with a master's degree in nuclear energy and nuclear technology engineering, and his main research interests include steam jet condensation, pressure vibration, critical heat transfer, and machine learning.

Authors:

Wei Zhang Shanghai Jiao Tong University
Yao Xiao Shanghai Jiao Tong Unversity
Lijun Yu Shanghai Jiao Tong Unversity
Hanyang Gu Shanghai Jiao Tong Unversity