Approaching system evolution in dynamic PSA by neural networks

Abstract

Conventional Probabilistic Safety Assessment (PSA) methodologies are static in nature and present potential weaknesses when applied to dynamic process systems. For this reason, the scientific community is considering the possibility of including the plant dynamic behaviour within the risk analysis. This entails that the systems under analysis be characterized in terms of their hardware stochastic features, of the values that the internal process variables take during the evolution and also in terms of their mutual interactions. The action of the control systems is also to be integrated in the analysis. The development of a dynamic approach to PSA presents the problem of setting and solving the mathematical models describing the plant's behaviour. The corresponding equations need to be integrated with time steps which are related to the evolution of the physical processes and therefore much smaller than the characteristic times of stochastic transitions. Consequently there is a substantial increase in the computational time required by the analysis. In this framework investigations are being performed in the development of fast methods for the approximate solution of the dynamic equations, with a degree of accuracy acceptable within the scope of the analysis. In this paper, we consider the application of a set of multilayered supervised artificial neural networks trained by the error back-propagation to follow the evolution of the process variables of the pressurizer in a nuclear power plant, under transient conditions. The results obtained show a satisfactory accuracy, in light of the inherent uncertainties in the stochastic data in input to the analysis, whereas the computer times required are substantially reduced, as compared to the times required for a numerical solution of the model. This reduction is expected to increase when the future parallel computers become available.