Robustness of a Neural Network Model for Power Peak Factor Estimation in Protection Systems
- Centro Tecnologico da Marinha em Sao Paulo - CTMSP, avenida Prof. Lineu Prestes, 2468 - Butanta, Sao Paulo (Brazil)
This work presents results of robustness verification of artificial neural network correlations that improve the real time prediction of the power peak factor for reactor protection systems. The input variables considered in the correlation are those available in the reactor protection systems, namely, the axial power differences obtained from measured ex-core detectors, and the position of control rods. The correlations, based on radial basis function (RBF) and multilayer perceptron (MLP) neural networks, estimate the power peak factor, without faulty signals, with average errors between 0.13%, 0.19% and 0.15%, and maximum relative error of 2.35%. The robustness verification was performed for three different neural network correlations. The results show that they are robust against signal degradation, producing results with faulty signals with a maximum error of 6.90%. The average error associated to faulty signals for the MLP network is about half of that of the RBF network, and the maximum error is about 1% smaller. These results demonstrate that MLP neural network correlation is more robust than the RBF neural network correlation. The results also show that the input variables present redundant information. The axial power difference signals compensate the faulty signal for the position of a given control rod, and improves the results by about 10%. The results show that the errors in the power peak factor estimation by these neural network correlations, even in faulty conditions, are smaller than the current PWR schemes which may have uncertainties as high as 8%. Considering the maximum relative error of 2.35%, these neural network correlations would allow decreasing the power peak factor safety margin by about 5%. Such a reduction could be used for operating the reactor with a higher power level or with more flexibility. The neural network correlation has to meet requirements of high integrity software that performs safety grade actions. It is shown that the correlation is a very simple algorithm that can be easily codified in software. Due to its simplicity, it facilitates the necessary process of validation and verification. (authors)
- Research Organization:
- The ASME Foundation, Inc., Three Park Avenue, New York, NY 10016-5990 (United States)
- OSTI ID:
- 20995648
- Resource Relation:
- Conference: 14. international conference on nuclear engineering (ICONE 14), Miami, FL (United States), 17-20 Jul 2006; Other Information: Country of input: France
- Country of Publication:
- United States
- Language:
- English
Similar Records
The potential of different artificial neural network (ANN) techniques in daily global solar radiation modeling based on meteorological data
Hybrid Improved Empirical Mode Decomposition and Artificial Neural Network Model for the Prediction of Critical Heat Flux (CHF)