Predição do estado de carga de baterias via redes neurais LSTM
Arquivos
Data
2024-09-10
Autores
Pedro, Paulo Roberto Costa 
Orientadores
Quiles, Marcos Gonçalves
Tipo
Trabalho de conclusão de curso
Título da Revista
ISSN da Revista
Título de Volume
Resumo
Com o avanço da importância da energia elétrica globalmente, torna-se crucial desenvolver modelos eficazes para avaliar os indicadores e parâmetros de funcionamento das baterias ao longo do tempo, uma vez que essas baterias são responsáveis por alimentar sistemas que dependem dessa forma de energia e pode trazer prejuízos a eles caso não funcione adequadamente. Neste trabalho, uma arquitetura de rede neural LSTM foi construída para realizar a tarefa de predição da capacidade de uma bateria em função da série temporal de seus parâmetros coletados. Esse modelo foi aplicado a um conjunto de dados públicos de simulação de ciclos de carga e descarga da bateria e, posteriormente, duas análises foram feitas a respeito do modelo: a otimização dos hiperparâmetros das camadas LSTM e a aferição da influência da introdução ou não de camadas CNN ao modelo. Os resultados mostraram que para a rede neural LSTM construída neste trabalho realiza as predições com erro RMSE de 0,07 nos conjuntos de treino, validação e teste e a introdução de camadas CNN ao modelo prejudicou seu desempenho, não sendo, portanto, adequada no domínio deste trabalho.
With the growing importance of electrical energy globally, it has become crucial to develop effective models to evaluate the indicators and performance parameters of batteries over time, as these batteries are responsible for powering systems that depend on this form of energy and could cause harm if they do not function properly. In this work, an LSTM neural network architecture was constructed to perform the task of predicting a battery's capacity based on the time series of its collected parameters. This model was applied to a public dataset of battery charge and discharge cycle simulations, and subsequently, two analyses were conducted regarding the model: the optimization of the LSTM layer hyperparameters and the assessment of the influence of introducing CNN layers to the model. The results showed that the LSTM neural network built in this work performs predictions with an RMSE error of 0.07 on the training, validation, and test sets, and the introduction of CNN layers to the model worsened its performance, making it unsuitable for the domain of this work.
With the growing importance of electrical energy globally, it has become crucial to develop effective models to evaluate the indicators and performance parameters of batteries over time, as these batteries are responsible for powering systems that depend on this form of energy and could cause harm if they do not function properly. In this work, an LSTM neural network architecture was constructed to perform the task of predicting a battery's capacity based on the time series of its collected parameters. This model was applied to a public dataset of battery charge and discharge cycle simulations, and subsequently, two analyses were conducted regarding the model: the optimization of the LSTM layer hyperparameters and the assessment of the influence of introducing CNN layers to the model. The results showed that the LSTM neural network built in this work performs predictions with an RMSE error of 0.07 on the training, validation, and test sets, and the introduction of CNN layers to the model worsened its performance, making it unsuitable for the domain of this work.