Estudo do padrão de recarga de um reator PWR com a arquitetura transformers

Abstract

The pressurized water reactor (PWR) core recharge optimization problem, while easily explained, is far from being easily solved. The designer’s task is to identify the layout of fresh fuel, partially burned fuel, and burnable poisons in the reactor core. This methodology optimizes the performance of the reactor during the operational cycle, until it needs to be refilled again, while ensuring that various safety constraints are always met. In this work, the study of the recharge pattern is approached, with the objective of developing a possible methodology to solve the optimization problem, making use of Maching Learn algorithms such as the transformers architecture. The objective is to obtain a satisfactory recharge pattern, with respect to the efficiency of the fuel assemblies, within the imposed parameters. For the study, the BEAVRS benchmark was considered, which specifies the details of operation, geometry, composition, etc. from a Westinghouse PWR reactor. To achieve the objective, the adopted methodology consists of simulating the reactor core with the WIMS-ANL and PARCS codes. These two codes are used together, where the WIMS-ANL code is used to obtain the macroscopic cross-sections, which are later used in the PARCS code. The neutronic simulation was performed for two purposes. The first was to validate the methodology used in the neutronic simulation. This validation allows verifying if the model is consistent with the real data, helping the development of other simulations with reliable data. The second purpose was to generate a training dataset for the model used in this work. With the neutron simulation, 40000 datasets of recharge patterns and power distributions were obtained, which were used in the training of the model. The methodology used consisted of proposing a power distribution considered “ideal” with the requirements of the theory and then using it as an input in the model to return the recharge pattern corresponding to this proposed power distribution. The main idea was to develop a model based on the transformer architecture that receives a power distribution as an input and returns the recharge pattern as an output. The model was trained with 40000 datasets, using the GoogleColab platform, which took approximately 13 hours to complete the process. The model results offered satisfactory results for power distributions that were generated randomly. And for the power distribution considered “ideal”, a power distribution that meets the restrictions used in this work was achieved.