Multiobjective optimization method dedicated to nuclear power plant operation: application to a sodium-cooled fast neutron reactor using a Brayton cycle

Defining the reactor operation allows the nuclear power plant to achieve objectives in terms of thermodynamic performance and to meet safety requirements. This work develops a method to define the reactor operation as the solution to a multiobjective optimization and constrained problem. The decision variables selected to solve this kind of problems are the actuators and the descriptive parameters related to regulations implemented in the reactor operation. The decision variables number is potentially high and induced a high number of simulations to solve the multiobjective problem. The reactor operation is modeled using the CATHARE2 code and is characterized by a long computation runtime, which makes the multiobjective problem resolve impossible. To overstep this problem, the developed method reduces the dimension size of the research space and builds surrogate models (metamodels) to replace CATHARE2 code simulations in order to mimic objectives and constraints depending on the decision variables. These metamodels use the conditioned Gaussian processes structure on a learning base of the variable to mimic. A coupling of these substitution models to a genetic algorithm enables the definition of a set of reactor optimal operations homogeneously spread in the solutions space. The low prediction errors of the metamodels provide an accurate estimate of the Pareto Front. The method is used to optimize the ASTRID operation for the loss of off-site power and the frequency setting transients. (author)