Redes neurais randomizadas para determinação da curva de potência de aerogeradores

• Ano de defesa: 2022
• Autor(a) principal: Silva, Tarsis Lima da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Curva de potência; Ajuste de curvas; Geração eólica; Redes neurais randomizadas; Regressão não-linear; Power curve; Curves fitting; Wind generation; Randomized neural networks; Non-linear regression
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/74025

Modeling the wind turbine power curve (WTPC) from measured data is critical to predict the electricity actually generated by wind farms. Polynomial regression is usually the first choice for WTPC modeling due to its wide availability in well-known software packages, but there are several powerful alternatives based, for example, on neural networks and fuzzy algorithms. With regarding the neural paradigm, the rediscovery of a class of supervised algorithms, generically called randomized neural networks, has aroused the interest of its use in practical applications, mainly due to the fast training process. However, the application of random networks in the modeling of power curves is not immediate and commonly presents incorrect solutions, here called pathological solutions, despite the good values achieved by the standard performance numerical indices, such as determining the Rˆ2 coefficient. In this thesis, a comprehensive critical study is carried out in order to evaluate the feasibility of using randomized networks in power curve modeling. Three architectures will be evaluated, namely, Random Vector Functional Link (RVFL), Extreme Learning Machine (ELM) and NoPropagation Network (No-Prop) on two real data sets. The adaptations are proposed and evaluated for the occurrence of pathological solutions. The performances of the aforementioned randomized networks are compared with the polynomial regression model and with the MLP (Multilayer Perceptron) network. The study carried out concludes that randomized networks are viable in the treatment of task of interest and present superior performance to the aforementioned models, both in numerical performance and in speed in the construction of the regression model, without generating pathological solutions