Monitoring changes in important parameters has been suggested as a potentially useful condition monitoring (CM) method for the accidents occurred in nuclear power plants (NPPs). The reactor core power is believed to be an important parameter governing the performance of reactor during transient response. The accurate prediction of reactor behavior and power is very important for nuclear power plant operators, especially during the major severe accident scenarios following an initiating event such as rod ejection accident (REA) or rod drop accident (RDA). REA and RDA together are found to be the worst postulated power transients in reactor licensing and are referred to as reactivity-initiated accidents (RIAs). RIAs is a postulated event of very low probability and involves inadvertent removal of a control element from an operating reactor, leading to a rapid power excursion in nearby fuel elements. On the basis of our previous work, in which only REA scenario is analyzed, the primary objective of this study is to develop and implement fuzzy weighted support vector regression (FWSVR) for condition monitoring under REA and RDA in NPPs. FWSVR is an extension of support vector regression (SVR) which introduces fuzzy weights in traditional SVR formulation. The accidents simulated in this study are based on the same model used in our previous work. This model can be accomplished by two procedures. First, the neutron flux and enthalpy distributions of the core can be obtained from a solution to the three-dimensional nodal space time kinetics equations and energy equations for both single and two-phase flows respectively. Second, the reactivity effects of the moderator temperature, boron concentration, fuel temperature, coolant void, xenon worth, samarium worth, control element positions (CEAs) and core burnup status can be calculated and determined. For the purpose of condition monitoring, it’s a fundamental issue to acquire condition monitoring data for useful data representation. Otherwise, it’s hard to predict the power accurately if enough data is not available. The data used in this study is collected from computer generated accident scenarios. Then the obtained data is split into two subgroups, training and test. The training subgroup data is utilized to train the FWSVR model and the fuzzy weights introduced in FWSVR are employed to extract multiple linear structures in a training dataset and assign to each data point a cluster index determined by its trained kernel radius function. The test is employed to validate the model. Finally the results of FWSVR model are compared with that of other models such as traditional SVR and back propagation network (BPN) which is one type of artificial neural networks (ANNs). Comparison of results among the three methods indicates that FWSVR model not only outperforms traditional SVR and BPN, but also has a better agreement with the general understanding than them.

This content is only available via PDF.
You do not currently have access to this content.