State-of-Charge Estimation of Li-ion Battery at Variable Ambient Temperature with Gated Recurrent Unit Network

The state of charge (SOC) is a crucial indicator of a Li-ion battery management system (BMS). A BMS with a good SOC assessment can dramatically improve the lifespan of the battery and ensure the safety of the end-user. With deep learning making tremendous strides in many other fields, this study aim...

Full description

Saved in:
Bibliographic Details
Main Authors: Hannan, M. A., How, D. N. T., Mansor, M., Lipu, M. S. Hossain, Ker, P. J., Muttaqi, K. M.
Format: Conference Proceeding
Language:English
Subjects:
Computational modeling
Deep learning
Estimation
gated recurrent unit
GRU
Li-ion
LIB
Lithium-ion batteries
Logic gates
SOC
State of charge
state-of-charge estimation
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The state of charge (SOC) is a crucial indicator of a Li-ion battery management system (BMS). A BMS with a good SOC assessment can dramatically improve the lifespan of the battery and ensure the safety of the end-user. With deep learning making tremendous strides in many other fields, this study aims to provide an empirical evaluation of commonly used deep learning methods on the task of SOC estimation. We propose the use of two-hidden-layer gated recurrent units (GRU) to estimate the SOC at various ambient temperatures. In this work, we conducted two experiment setups to showcase the capability of the proposed GRU model. In the first setup, the GRU was trained on the DST, BJDST and US06 drive cycle and evaluated the FUDS drive cycle upon convergence. The same procedure was repeated with the second setup except the GRU was trained on the DST, BJDST and FUDS drive cycle and evaluated on the US06 drive cycle. In both experiment setups, the proposed GRU was evaluated on a novel drive cycle that it has not encountered during the training phase. We show that a two-hidden-layer GRU with appropriate hyperparameter combination and training methodology can reliably estimate the SOC of novel drive cycles at various ambient temperatures in comparison with other deep learning methods such as simple recurrent network (SRNN), Long Short-Term Memory (LSTM), 1D Residual Network (Resnet), 1D Visual Geometry Group Network (VGG) and the Multilayer Perceptron (MLP). The proposed GRU achieves 2.3% RMSE on the FUDS drive cycle and 1.2% RMSE on the US06 drive cycle outperforming all other models.
ISSN:2576-702X
DOI:10.1109/IAS44978.2020.9334824