Li-ion Battery Health State Estimation with Two Feed Forward Neural Networks

Abstract

Deep neural network (DNN) architectures have been extensively studied for estimating the state-of-health (SOH) of Li-ion batteries (LIBs), but the effects of hyperparameters on their performance have not been thoroughly analyzed. As a result, it is difficult to compare the effectiveness of different DNN architectures and to extend or adapt previous works. We investigate the impact of hyperparameters on the performance of feed forward neural networks for LIB SOH estimation. Specifically, we propose two feed-forward neural networks: an accuracy model optimized for predicting the SOH of a single cell with high accuracy and a generalized model optimized for predicting the SOH of various cells. We demonstrate the effects of hyperparameters on various models using the National Aeronautics and Space Administration (NASA) prognostic battery dataset, which we cleanse to eliminate anomalous data and analyze its features. Our experiments show that our accuracy model achieves a root mean square error of 0.33%, demonstrating its efficacy for predicting the health state of individual cells. We also compare our results with those of other studies, establishing a benchmark for future research. Our open resources, including code, cleaned data, feature analysis, and experimental reports, are publicly available on GitHub (https://github.com/jhrrlee/batterystates).