Otimização da recarga de um núcleo PWR usando redes neurais artificiais MLP e PSO

Abstract

A recharge is generally required every 18 months in pressurized water reactors, such as those in Brazil. This process would be simple if it were similar to changing parts in a given machine. However, recharging requires knowledge due to several factors, such as a high amount of fuel elements in the core; in the case of Angra 2 there are 193. The exchanges are not for all positions, generally the task is carried out in only 1/3 of the total. The amount withdrawn is the one that burns the most fuel. When inserting new fuel elements, changes occur throughout the core. This happens due to the fact that each fuel element generates power corresponding to the type of enrichment used and position arranged. Therefore, such factors influence the distribution of power in the reactor, which may alter the expected and/or planned performance for the next operating cycle, even altering important safety parameters. Several studies that use computational algorithms in order to optimize the reactor core are proposed in the literature and such research is still an open field for new proposals. This work intends to find results that can contribute to obtain better fuel distributions in the core using ANN (Artificial Neural Networks) MLP (Multi_Layer Perceptron) with Backpropagation and PSO (Particle Swarm Optimization) training algorithms. The data used in the ANN training were obtained through the PARCS v code. 2.4 (Purdue Advanced Reactor Core Simulator), using the Angra 2 reactor modeling whose data were obtained from its respective Final Safety Analysis Report. In the tests performed, the ANN achieved an accuracy of 98%.