Development of real-time core monitoring system models with accuracy-enhanced neural network and its application

In a complicated system like pressurized water reactor, a number of key safety parameters need to be selected to represent the reactor systems safety. It could be more effective for the reactor safety to make the key safety parameters in real-time available directly to the reactor operator. Direct representation of key safety parameters is also desirable in the view of reactor core design since it could reduce unnecessary margins for various components of uncertainties. In this thesis, real-time core monitoring system models have been developed with use of artificial neural networks for the prediction of nuclear hot channel factor (HCF) and core departure from nucleate boiling ratio (DNBR) which are known to be the fundamental core safety parameters for pressurized water reactors. Artificial neural network algorithm, has been shown to be successful for the conservative and accurate prediction of the HCF and DNBR. For the development of system models, training patterns were generated using the FLAIR and COBRAIV-i computer codes for the HCF and DNBR. The selected input variables were the core power, reactor coolant flow, temperature, pressure, power distribution, boron concentration, and rod position. The developed system models could replace the existing core monitoring systems and then afford a better efficiency by using the additional margin which otherwise needs to be reserved for various unidentified uncertainties. Several variations of the neural network technique have been proposed and compared based on numerical experiments. The neural network can be augmented by use of a functional link to improve the performance of the network model. The functional link is found to be very effective especially when the relationship between the input parameters and the output parameters is overly complicated such as in the core HCF and DNBR. For the further enhancement of DNBR accuracy, two-fold weight sets were used. The coarse weight set can provide a quick and conservative prediction for the wide range of core DNBR while the fine weight set can provide a more accurate yet conservative prediction of core DNBR over a specified DNBR range. With this scheme, the DNBR can be predicted more accurately when the core DNBR approaches the safety setpoint. The weighted system error backpropagation method has also been found to be very effective for the accuracy enhancement and can have a variety of applications if proper weighted function is chosen. For actual applications the uncertainty factor as a function of output was introduced to provide the conservative predictions. A sensitivity analysis was then peformed by taking the partial derivative of the DNBR with respect to the core power to secure a sensitivity coefficient. This coefficient was used to retrieve the available power margin which means how far the core power is away from the DNB occurrence. Based on this analysis, it can be concluded that the backpropagation network training algorithm can be used as a tool for the prediction of the core safety parameters and that the developed system models using the neural network can be used for the accurate yet conservative prediction of the core HCF and DNBR on a real-time basis