Metodologia para previsão de carga no horizonte de curto prazo utilizando redes neurais

• Ano de defesa: 2019
• Autor(a) principal: Milke, Tafarel Franco
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; Engenharia Elétrica; UFSM; Programa de Pós-Graduação em Engenharia Elétrica; Centro de Tecnologia
• Palavras-chave em Português: Previsão de carga; Redes neurais artificiais; Curto prazo; Load forecast; Artificial neural networks; Short term; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufsm.br/handle/1/19743

Competitiveness and the insertion of new technologies in the electricity sector now condition companies to find ways to improve the quality of their services and ensure profitability. The short-term load forecasting activity is indispensable to support the planning and operation of electrical systems, aiming to make the energy supply stable and reliable. To perform load prediction using Artificial Neural Networks (ANN), it is necessary to evaluate the variables involved in the behavior of the daily load curve. By evaluating and obtaining the most available variables influencing the load behavior, it is then possible to use them as input to the adopted ANN model. Artificial neural networks are computational models inspired by the simplification of the functioning of biological neurons, with the ability to learn from experience with system inputs. They are similar to the brain due to the characteristics of knowledge acquired by a learning process and connections between its neurons used to store the acquired knowledge. A neural network has high power to generalize information after a learning phase, allowing to capture functional relationships between data producing output close to the expected. The process of learning or training the network consists in the application of ordered steps necessary for the tuning of the synaptic weights and thresholds of their neurons, aiming to produce the generalization of solutions by their outputs. The goal of network training is to make the application of a set of inputs a set of desired outputs. The tools using artificial intelligence techniques have been improved, allowing their application in various areas of knowledge, standing out among the main techniques used to perform short-term load forecasting, and are currently widely researched and employed for this purpose. Thus, its use has been showing more accurate results compared to traditional methods, since they can better develop the required mathematical processing. This paper presents a proposal for the prediction of the daily load curve for one day ahead applied to real energy, demand and temperature data, since it is the variables that best represent the short-term load behavior; For this, a model developed with multilayer perceptron neural networks using the Levenberg-Marquardt learning algorithm was implemented. The results found were satisfactory and acceptable compared to those presented in the literature review, being sufficient for practical application meeting the proposal of this work.