Metodologia para a extração automática de regras para a modelagem de um sistema de refrigeração de água utilizando abordagem digital Twin

• Ano de defesa: 2021
• Autor(a) principal: Araújo Junior, Carlos Antonio Alves de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Engenharia Elétrica; Programa de Pós-Graduação em Engenharia Elétrica; UFPB
• Palavras-chave em Português: Setor elétrico; Planta térmica; Gêmeo digital; Inteligência artificial; Lógica fuzzy; Extração automática de regras; Energy sector; Thermo powerplant; Digital twin; Artificial intelligence; Fuzzy logic; Automatic rule extraction; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/20749

In the search of higher productivity, industry has developed technological strategies to achieve this goal, Industry 4.0. In the energy sector, thermal power plants are looking for tools capable of optimizing energy production and reducing unscheduled downtime. This work proposed to model the water cooling system from the Digital Twin approach, using artificial intelligence techniques such as fuzzy logic and automatic extraction of fuzzy rules. The water cooling system is a system that has many fans to perform the thermal exchange between water and ambient air, however, the strong non-linearity characteristic of the system means that conventional PID controllers do not have a good performance, inducing the operator to activate all the fans to guarantee the water temperature below the necessary and, consequently, increasing the energy expenditure unnecessarily. This approach uses data from a thermal power plant located in the capital of Paraíba, João Pessoa, to create a knowledge database through the history data of the system and, thus, create a digital twin capable of helping to optimize the energy consumption of this system, which arrives about 3% of all energy generated at the plant. An algorithm for online updating the model rules was also proposed. The update aims to ensure that the system has a low instantaneous percentage error, as well as to acquire new knowledge of situations in which the model has not been trained. An optimization algorithm was used, using evolutionary strategies to search for a smaller number of fuzzy sets, achieving less computational effort without major impacts on the result. The results presented by the model from tests, in three different scenarios, showed the robustness of the proposed model, and in all scenarios, the average percentage error was below 5% and the average absolute error below 3°C.