Accuracy Improvement of Axial Power Shape Reconstruction of GMDH Algorithm Applying Pattern Recognition Technique

Abstract

A monitoring of the status of Nuclear Power Plants (NPPs) is essential to support operators of NPPs to avoid the initial events which eventually lead to severe accidents. In order to monitor the axial power distribution of nuclear reactor, the reconstruction of axial power shape using in-core detector signals is important. Group Method of Data Handling (GMDH) algorithm is a robust neural network algorithm to determine the relationship between in-core detector signals and axial power distribution. In the previous investigations, GMDH algorithm was examined for replacing Fourier series expansion method adopted to reconstruct the axial power distribution by commercially available Core Operating Limit Supervisory System (COLSS), and it was confirmed that GMDH algorithm could produce higher accuracy than the Fourier series expansion method. In this paper, training data sets have been categorized depending on the axial power shapes, i.e., cosine, saddle, top-skewed, and bottom-skewed shapes. Each category of the axial shape can be pattern-recognized by examining the in-core detector signals. The results show that recognizing data sets in different categories of the axial shapes and then applying the GMDH produces 1.37% higher accuracy in average, and 2.84% at maximum.