Local power peaking factor estimation in nuclear fuel by artificial neural networks

This paper presents the training of an artificial neural network (ANN) to accurately predict, in very short time, a physical parameter used in nuclear fuel reactor optimization: the local power peaking factor (LPPF) in a typical boiling water reactor (BWR) fuel lattice. The ANN training patterns are distribution of fissile and burnable poison materials in the fuel lattice and their associated LPPF. These data were obtained by modeling the fuel lattices with a neutronic simulator: the HELIOS transport code. The combination of the pin U235 enrichment and the Gd2O3 (gadolinia) concentration, inside the 10 × 10 fuel lattice array, was encoded by three different methods. However, the only encoding method that was able to give a good prediction of the LPPF was the method which added the U235 enrichment and the gadolinia concentration. The results show that the relative error in the estimation of the LPPF, obtained by the trained ANN, ranged from 0.022% to 0.045%, with respect to the HELIOS results.