@article{866beae908834d17a8f708c8c122a1e6,
title = "Adaptable capacity estimation of lithium-ion battery based on short-duration random constant-current charging voltages and convolutional neural networks",
abstract = "The accurate lithium-ion battery capacity estimation is vital for ensuring the safe and reliable operation of battery-powered systems. Existing data-driven methods heavily rely on fixed charging stages for feature extractions, posing significant limitations in real-world applications. This paper proposes an adaptable capacity estimation approach utilising short-duration random charging voltages during the constant-current charging stage and leveraging convolutional neural networks (CNNs). Based on the user-friendly “Vstart−tend” strategy, two health features including charging voltage and its increment are firstly extracted from random charging segments. Secondly, a feature evolution pattern analysis over the battery's lifespan is proposed to divide the charging voltage range for robust model development. An optimal combination of both the sampling interval and data length is determined for the feature extraction. Then, a two-dimensional CNN model is developed to effectively learn ageing-related knowledge from various random charging segments in a specific charging voltage range. The effectiveness of the proposed approach is ultimately verified using two distinct types of batteries across three operational temperatures. The results demonstrate that the proposed approach show much potential as a promising capacity estimation technique utilising a 600 s random charging segment sampled at a 20 s interval.",
keywords = "Capacity estimation, Convolutional neural network, Lithium-ion battery, Random and short-duration charging voltages",
author = "Zuolu Wang and Xiaoyu Zhao and Dong Zhen and Jo{\~a}o Pombo and Wenxian Yang and Fengshou Gu and Andrew Ball",
note = "Funding Information: In practical applications, the assessment of battery capacity can be performed during either the discharging or charging phases, with a preference for the charging profile due to its simplicity. During the charging process, the battery inevitably experiences the stable constant-current-constant-voltage (CC-CV) protocol and relaxation stage [31]. Based on the charging profile, numerous machine learning techniques have been created for estimating battery capacity. Deng et al. [32] and Xiao et al. [33] extracted various training features from the complete CC-CV charging data of the NASA battery dataset and mapped the battery SOH using least squares support vector machine (LSSVM) and least square support vector regression (LSSVR), respectively. Zheng et al. [34] used the entire CC charging voltages for the model training, and the obtained root mean squared error (RMSE) and mean absolute error (MAE) were 1.91 \% and 3.37 \%. However, a lot of inputs of charging voltages will bring large computations. Based on the complete CV charging current, Wang et al. [35] extracted the CV charging ageing factor for the battery SOH evaluation using nonlinear least squares. While the mentioned methods have been confirmed to provide an accurate estimation of battery SOH, the extraction of features from the extended CC-CV, CC, and CV phases will augment the estimation workload. In pursuit of a relatively efficient estimation of battery SOH or capacity, certain researchers have specialized in feature engineering from partial charging data. For example, Ruan et al. [36] extracted 500 s charging data from the CC-CV transition segment and estimated the battery SOH based on the transfer learning-based convolutional neural network (CNN) model. However, the identification of the CC-CV transition increases challenges in practical applications. Ruan et al. [37] also developed a novel SOH estimator based on Q\textbackslash{}u2212V modelling and open-circuit voltage (OCV) using 1000 s CV charging capacity data. The OCV under different levels of degradations and the initial CV point are required to be considered for the accurate battery SOH estimation. Guo et al. [31] investigated the CV charging process and selected three types of features in the current range of 0.7C and 0.2C. Combined with the Gaussian process regression (GPR) model, a satisfactory estimation was obtained with a 2.9 \% error. Chen et al. [26] extracted the charging current and differential current from the 1000 s CV process for the accurate estimation of the battery SOH using a developed CNN model. Additionally, Tian et al. [38] suggested extracting ageing-related features from the CC charging segment within the voltage range of 3.9\textbackslash{}u20134.1 V, utilising the ICA curve and parameter identification of the ECM. The utilisation of both ICA and model parameter identification causes high estimation complexity. Wu et al. [39] also analysed the partial CC charging voltage and extracted dQ/dV (the derivative of the battery capacity (Q) with respect to the cell voltage (V)) data from the voltage range of 3.8\textbackslash{}u20134.1 V for the model training and battery SOH characterization. Fu et al. [24] introduced an incremental slope (IS) aided feature extraction approach to acquire universal multidimensional health features from the CC charging stage and accomplished the battery SOH estimation with a multilayer perceptron (MLP) and transfer learning method. Lu et al. [40] explored various voltage ranges for sequence feature extraction and battery SOH estimation based on a developed CNN model. It depends on the specific charging voltage range, and the feature engineering, including voltage, dQdV, and the derivative of the capacity, is a complex process. Furthermore, some recent studies have attempted to analyse the relaxation voltage (RV) for feature extractions and battery capacity estimation. RV means the terminal voltage across the battery in the rest stage after charging. Particularly, Zhu et al. [41] curated a dataset comprising three distinct types of lithium-ion batteries and explored feature extraction, specifically focusing on various statistical parameters derived from the RV stage when the battery is fully charged. The effectiveness was validated through support vector regression (SVR) and XGBoost. Moreover, Fan et al. [42] verified the high correlation of the RV with the battery capacity throughout the entire battery life cycle and realised the accurate battery capacity estimation only using 10 s RV data and a CNN model. Nevertheless, the RV data has been extracted from a fully charged state of the battery. However, the above studies focus on the feature extraction in a fixed charging stage during CC-CV, CV, CC, and RV processes, which reduces adaptability for real-world applications.For further comparison, this paper compares the estimation performance of the proposed model with several typical models, such as feedforward neural network (FNN) and support vector machine (SVM) regression model, LSTM, 1D-CNN and attention mechanism based 2D-CNN (2D\textbackslash{}u2013CNN\textbackslash{}u2013Attention) in this section. The 1D-CNN model with the same network architecture as the proposed 2D-CNN model but different input sequences is also used to the battery capacity estimation.The authors wish to convey their gratitude to the Center for Efficiency and Performance Engineering (CEPE) at the University of Huddersfield for their support in the completion of this research. This work was also supported by FCT, through IDMEC, under LAETA, project UIDB/50022/2020. Funding Information: The authors wish to convey their gratitude to the Center for Efficiency and Performance Engineering (CEPE) at the University of Huddersfield for their support in the completion of this research. This work was also supported by FCT, through IDMEC, under LAETA, project UIDB/50022/2020. Publisher Copyright: {\textcopyright} 2024 The Authors",
year = "2024",
month = oct,
day = "15",
doi = "10.1016/j.energy.2024.132541",
language = "English",
volume = "306",
journal = "Energy",
issn = "0360-5442",
publisher = "Elsevier Limited",
}
