Artificial neural network prediction of PWR critical boron concentration

The direct calculation of core parameters such as k{sub eff} and pin power peaks for light water reactors is ordinarily accomplished by numerically solving the neutron diffusion equation. Despite the rapid advances in computer architecture and algorithm development, further calculational speedups are always in great demand. One example of such an application is nuclear fuel management optimization, where the core attributes of tens of thousands of loading pattern candidates must typically be evaluated over the fuel cycle. If an artificial neural network (ANN) could be trained to accurately model the neutronic behavior of a core, a substantial time savings could be realized in the prediction of core parameters. Such an ANN could be exploited in at least two ways: 1. The a priori training of an ANN model could be tailored to address a specific plant and its corresponding licensing core neutronics software. 2. Once trained to within acceptable accuracy guidelines, an ANN model could provide the luxury of nearly instantaneous evaluations of core parameters. Recent publications by Kim et al. on core parameter prediction via ANNs have revealed a variety of promising results, which, in part, motivated our studies. Kim proved that a solution was possible; however, the largemore » size and complexity of such a model can lead to memorization instead of generalization of the problem`s solution. Thus, the purpose of this work was to show that a much smaller ANN could predict a global core parameter such as the critical boron concentration over a wide range of training and validation data. The successful modeling of this problem with a much smaller ANN is considered to be a significant highlight of this study. This work employed Studsvik of America`s SOA1 Database, which proved to be useful for ANN training and validation.« less