An impact of activation functions on CNN-Bi-LSTM SoC estimation for Li-Ion batteries

Abstract

Estimation of state of charge (SoC) in Li-ion batteries has been accomplished by many methods over couple of years. Every research in this domain majorly focusses on improving accuracy of estimation and some in exploring new algorithms. In this regard, we worked in analyzing the accuracy of estimation of SoC of sophisticatedly accurate techniques. A data-driven SoC estimation using deep neural network architectures (CNN) were collaborated with different activation functions like global pooling, exponential linear unit, Leaky rectified linear unit on convolutional neural network (CNN) and bidirectional-long short-term memory (Bi-LSTM) models. The performance of all these models was evaluated using quantitative error analysis and qualitative signal tracking studies to compare the estimated and actual SoC trajectory. This work consolidates the performances of different models and helps realize how best a fusion model works on the best accuracy in estimation.