Thermal power estimation based on a neural network and principal component analysis in nuclear power plants

Abstract

This study proposes an improved signal correction strategy for accurate thermal power estimation in nuclear power plants. Generally ~2% thermal efficiency can be overestimated because of fouling phenomena in a feedwater flowmeter. In the proposed strategy the noises in feedwater flowrate signals are classified into three kinds according to their properties such as fouling degradations, statistical fluctuations, and rapid distortions. Each noise is removed by a multi-step de-noising technique that consists of signal preprocessing, signal aggregation, and signal estimation. Signal preprocessing is achieved for the removal of rapid distortions on the basis of low pass filtering using a wavelet analysis. In signal aggregation a principal component analysis is applied as a linear fitting method to cover statistical fluctuations in fouling phenomena. To estimate real signals from the signals including fouling degradations, an auto-associative neural network is used as another principal component analysis methodology to apply to a nonlinear system. Additionally thermal power deviation estimators are proposed to recognize the effects of deviations in each parameter for thermal power calculation. Because this estimator can provide error information according to the parameters, it may be useful for maintenance decision-making. The proposed methodologies were validated using the signals from a micro simulator and noise modeling. 2.3% root mean square (RMS) errors in noisy signals were reduced to 0.3% RMS errors after the multi-step de-noising technique. And calculation error, which is estimated by the thermal power deviation estimator and the results of the multi-step de-noising technique, is nearly approximated to the results by analytical calculation.