Modelo para estimação do estado de carga e saúde de baterias de lítio-íon baseado em redes neurais com função de custo em correntropia

Abstract

The great interest in identifying non-linear dynamic systems is due, mainly, to the fact that a large number of real systems are complex and need to have their linearities considered so that their models can be successfully used in applications such as control, prediction, inference, among others. An example of a non-linear dynamic system is the battery. Understanding battery aging is a complex process, because many factors, from environmental conditions to operational conditions of use, interact in order to generate different effects of aging and degradation. This work analyzes the application of Artificial Neural Networks in the identification of performance and health indicators in batteries. Normally, the learning of these networks is done through some method based on gradient, having the average quadratic error as a cost function. This work analyzes the replacement of this traditional cost function by a measure of the similarity of Information Theory, the Correntropia. This measure of similarity allows that statistical moments of a higher order can be considered during the training process. Due to this fact, it becomes more appropriate for non-Gaussian error distributions and makes the training less sensitive to the presence of outliers. To evaluate this substitution, ANNs models are obtained in the identification of two case studies: the State of Charge (SoC) and State of Health (SoH) of batteries. The results show that the use of correntropy, as a cost function in the error backpropagation algorithm, makes the identification procedure using Neural Networks more robust to outliers, as well as minimizing model estimation errors. However, this can only be achieved by properly adjusting the width of kernel gaussian, either through techniques based on genetic algorithms or through estimates found in the scientific literature.