Development of hybrid artificial neural network system for the prediction and validation of nuclear reactor parameters

Abstract

Safety and reliability of nuclear power plant are largely dependent upon the validity and accuracy of sensor signals that indicate the unclear reactor operating states. In this thesis, a new signal processing method named hybrid artificial neural network(HANN) system is developed for the pridiction and validation of nuclear reator parameters. The developed system consist of a two step signal processing submodels, one for the input signal prevalidation using the parity space representation model(PSRM) and second, for the parameter premeter prediction through the artificial neural network(ANN) model. To construct and demonstrate the applicability of the system, in part I, a hybrid multi-layer network(HMLN) model which consists of a general multi-layer network(GMLN) combined with PSRM is developed and that model is used for the prediction and validation of thermal power prarmeters on the nuclear power plant(NPP). The back-propagation learning rule, one of supervised learning algo-rithms, is applied to train the network in the system. A number of case studies are performed with emphasis on the applicability of the model in a steady state high power level. The studies reveal that the developed model can precisely predict the thermal power of an NPP. It also shows that defected signals resulting from instrumentation problems, even when the signals comprising various patterns are nosiy or incomplete, can be properly handled. In part II, as an improved single layer ANN, a hybrid functional-link network(HFLN) model is developed by using the weighted non-linear signal combination method. This model is applied in the case study of load follow operation simulation in which four major operating parameters are predicted. The prediction results are compared with the results of GMLN case and also with that of actual test data. The overall prediction results agreed well with the plant data except for the minor discrepancies at the near upper or lower bounds of the trained band....