Deep learning enabled state-of-charge estimation of LiFePO4 batteries: A systematic validation on state-of-the-art charging protocols

Abstract

LiFePO4 batteries generally face a challenge of inaccurate state of charge (SoC) estimation due to the plateaus existing in the middle range of the open circuit voltage (OCV)-SoC curve. Generally, conventional SoC estimation methods are not capable of accurately estimating the SoC in this range. In this paper, a deep neural network (DNN) is constructed to estimate the SoC in the charging process. Battery data collected from five state-of-the-art charging protocols at 10 °C, 25 °C and 40 °C are used to train the DNN. The developed DNN can be used for online SoC estimation subsequently. This estimated SoC can serve as the initial SoC of the ampere-hour counting method to calculate the SoC of the discharging process. The overall maximum error and the root mean square error of the SoC estimation over charging process are within 2.5% and 0.8%, respectively. In addition, the input depth of time from 10 s to 100 s with a 10 s interval is investigated. The maximum error is less than 5% in the case of the depth time within 100 s and the error fall to 2% when the depth of time reaches 90 s.