Artificial Neural Network (ANN)-Based Response Surface Approach for Passive System Reliability Assessment

Abstract

Advanced designs of nuclear reactors deploy passive safety systems, in combination with active systems in order to enhance safety, to improve reliability, and to reduce human intervention. Substantive efforts are underway worldwide to estimate passive system reliability using different approaches, however, consensus of these approaches is not achieved so far. In most of the methods developed for passive system reliability assessment, the fundamental approach is to analyze thermal-hydraulic behavior of system for limited number of scenarios and develop mathematical relationship between set of input parameters and parameter of interest, normally known as ‘Response Surface’. Multivariate linear/nonlinear regression is used for developing such relationships. This assumption may not be always true for complex systems, and hence, such assumptions for problem simplification could introduce uncertainty in reliability assessment. Furthermore, the input variables may be interrelated and that could introduce additional uncertainties. Artificial neural network (ANN) can eliminate need for such assumptions in predicting the relationship between the output and input parameters and thereby reducing the uncertainties in passive system reliability analysis. In this paper, the ANN approach is used for generating the response surface, which can then be used for estimating passive system reliability. A typical isolation condenser system, which operates in two-phase natural circulation principle, is considered in this analysis. The system performance is analyzed for different combinations of input parameters. Latin hypercube method is used to select a random set of input parameters. For this set of input parameters, system performance is analysis using thermal-hydraulic code RELAP5/MOD3.2. Multivariate linear regression analysis is carried out for generating response surface. Similarly, response surface is generated using the ANN model developed on MATLAB tool. The comparison is made between two approaches and their effect on the response surface, which eventually have significant impact on the reliability estimates of passive system.