A hybrid accident simulation methodology for nuclear power plant by combining thermal-hydraulic program and artificial neural networks

Compact simulators for nuclear power plants can be used as cost-effective training or analysis tools; generally, they demonstrate overall responses of transients or accidents in real time or faster. In the thermal-hydraulic models of compact simulators, governing equations are simplified with reasonable assumptions and empirical correlations, and approximate solutions are obtained by using appropriate numerical schemes. Moreover, many physical control volumes in plant modeling are lumped to reduce the computing time. The simplification of equations and reduction of control volume numbers usually degrade the accuracy of solutions. A hybrid accident simulation methodology is proposed to enhance the capabilities of a compact simulator by introducing artificial neural networks. A simplified thermal-hydraulic program, playing the role of compact simulator, is designed to calculate the overall responses of transients and accidents. Two neural networks are designed and trained with the target values obtained from the analyses of detailed computer codes and trained results are combined with the simplified thermal-hydraulic program to perform the following roles: (I) compensation for inaccuracy of a simplified thermal-hydraulic program occurring from simplified governing equation and small number of physical control volumes: the auto-associative neural network (AANN), trained with the target values obtained from RELAP5/MOD3 code analyses, improves the calculated results of the simplified thermal-hydraulic program, and (II) prediction of the critical parameter usually calculated from the sophisticated computer code: the back propagation neural network (BPN), trained with the target values obtained from COBRA-IV code analyses, predicts the minimum departure from nucleate boiling ratio (DNBR) which is not calculated in simplified thermal-hydraulic program. Simulations for the several accidents are carried out to verify the applicability of the proposed methodology. The verification results show that more accurate computational results can be obtained from the simplified thermal-hydraulic computer code while maintaining its fast simulation capability: the AANN improves the accuracy of results from the simplified thermal-hydraulic program up to the accuracy level of detailed computer code, and multi-calculation stages to obtain the minimum DNBR can be integrated into one stage with reasonable accuracy: the minimum DNBR is calculated by the BPN without any additional algorithmic calculation processes. It is concluded that the neural network can be used as a complementary tool to improve the capabilities of a simplified thermal-hydraulic computer code