Neural network modeling: application to fuel management in a reactor.

Summary Every year, a quarter of the nuclear fuel in a reactor core is replaced by new fuel. Because of their history. The other fuels have very different physical properties. This raises the problem of how to distribute the fuel within the core while ensuring that the proposed loading plan meets safety constraints: this is the problem of fuel repositioning. If an exhaustive search is illusory, it requires the calculation of neutron characteristics, especially the power peak, for thousands of loading plans. In an automatic optimization code such as Formosa, the calculation of these characteristics represents 90% of the ten hours necessary for the optimization. In order to reduce this computation time, we propose an original neural architecture adapted to the physical phenomenon to model. A data analysis has allowed us to characterize the nuclear fuel more precisely. The introduction of an a priori knowledge of the neutron phenomena allowed to reduce the number of free parameters of the model. We then implemented this neural model in Formosa, and we showed a considerable time saving on real EDF problems. Finally, we also propose a hybrid method combining the best of the local linear approximator GPT (generalized perturbation theory) and neural networks.