Application of feedforward neural networks for virtual sensor implementation in nuclear reactors

Abstract

This study builds on a line of research developed at the Nuclear and Energy Research Institute (IPEN), focused on the use of Artificial Neural Networks (ANNs) to provide sensor redundancy in nuclear reactors. The proposed methodology consists of training a Feedforward Neural Network (FFNN) to estimate the value of one reactor variable based on measurements from other variables. The feasibility was demonstrated using data from IPEN’s nuclear research reactor and a scaled model. While previous works relied on data from a single operating cycle of the reactor, this study tested network performance using data from multiple cycles to assess accuracy and generalization, and to identify deficiencies in the method. One fault-free cycle was selected for training and validation, while five others were used for testing, featuring different events. Three temperature variables, three radiation variables, one power measurement, and one safety rod position were chosen as output sensors for the ANNs to estimate. Performance was evaluated using the Mean Absolute Percentage Error (MAPE). All ANNs performed well during training. However, only two temperature variables were estimated across all the test cycles with similar accuracy as in the training and validation process, with MAPE values below 3%. Other ANNs performed poorly, with instances of persistent offsets or failure to track the general shape of the measured signals. The results underscore the challenges posed by strong input-output correlations and the difficulty in capturing the full complexity of functional relationships within reactor variables.