Modelo para estimação do estado de carga de baterias de lítio-íon baseado em redes neurais auto regressivas não-lineares com entradas externas

• Autor(a) principal: Moura, Jonathan Jefferson Pereira
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFPB
• Texto Completo: https://repositorio.ufpb.br/jspui/handle/123456789/19985
• topic: Redes neurais artificiais; Estado de carga; Baterias de lítio-íon; Redes neurais não-lineares auto regressivas com entradas externas; Matlab/Simulink®; Estruturas de rede; Artificial neural networks; State of charge; Lithium-ion batteries; Nonlinear autoregressive with external input neural networks; Network structures; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA

Nowadays,the research for use and enhancement in the electrical energy storage for use in isolated power systems (offgrid), mobile devices or electric vehicles applications was became a focus of research that aim a greater efficiency in the charge and discharge control of batteries. In this scenario, a method is proposed to State of Charge (SoC) estimation for lithium-ion batteries using as base structure a neural network of type nonlinear autoregressive with external inputs (NARX). The structure used to determine the SoC consists of a NARX network containing the battery terminal voltage and ambient Temperature as external inputs in addition to the feedback SoC in the previous instant, providing in the network output the SoC value at the current instant. The procedures for creating the database with discharge tests for different currents and temperatures were performed in the Simulink tool. The training of the neural network was done using the Neural Net Time Series toolbox of Matlab® software. A total of 12 neural networks were developed using hyperbolic tangent, sigmoid, linear saturated and purely linear activation functions so that their validations were also performed through Matlab®. The results of the validation tests were compared with other structures of neural networks based on techniques reported in the literature, so that the proposed structure obtained a maximum Mean Squared Error of 1.864% and a maximum Maximum Absolute Percent Error of 2.807%.