Neural network correlation for power peak factor estimation

Abstract

This paper proposes a method, based on the artificial neural network technique, to predict accurately and in real time the power peak factor in a form that can be implemented in reactor protection systems. The neural network inputs are the position of control rods and signals of ex-core detectors. The data used to train the networks were obtained in the IPEN/MB-01 zero-power reactor from especially designed experiments. The relative error for the power peak factor estimation ranged from 0.19% to 0.67%, an accuracy better than what is obtained performing a power density distribution map with in-core detectors. The networks were able to identify classes and interpolate the power peak factor values. It was observed that the positions of control rods bear the detailed and localised information about the power density distribution, and that the axial and the quadrant power differences, obtained from signals of ex-core detectors, describe its global variations in the axial and radial directions. In the power reactor environment, the neural networks would require in the input vector the position of control rods, and axial and quadrant power differences. The results showed that the RBF networks produced slightly better results than the MLP networks, but, for practical purposes, both can be considered of similar accuracy. The results indicate that they may allow decreasing the power peak factor safety margin by as much as 5%.