Elsevier

Journal of Power Sources

Volume 270, 15 December 2014, Pages 262-272
Journal of Power Sources

A support vector machine-based state-of-health estimation method for lithium-ion batteries under electric vehicle operation

https://doi.org/10.1016/j.jpowsour.2014.07.116Get rights and content

Highlights

  • We present a novel SOH estimation method based on on-board available battery data.

  • Standard performance tests are applied virtually to a data-driven battery model.

  • The method conveniently yields the same figures of merit as in standard tests.

  • Capacities and instantaneous resistances are estimated accurately.

Abstract

Capacity and resistance are state-of-health (SOH) indicators that are essential to monitor during the application of batteries on board electric vehicles. For state-of-health determination in laboratory environment, standard battery performance tests are established and well-functioning. Since standard performance tests are not available on-board a vehicle, we are developing a method where those standard tests are applied virtually to a support vector machine-based battery model. This data-driven model is solely based on variables available during ordinary electric vehicle (EV) operation such as battery current, voltage and temperature. This article contributes with a thorough experimental validation of this method, as well as the introduction of new features – capacity estimation and temperature dependence. Typical EV battery usage data is generated and exposed to the suggested method in order to estimate capacity and resistance. These estimations are compared to direct measurements of the SOH indicators with standard tests. The obtained estimations of capacities and instantaneous resistances demonstrate good accuracy over a temperature and state-of-charge range typical for EV operating conditions and allow thus for online detection of battery degradation. The proposed method is also found to be suitable for on-board application in respect of processing power and memory restrictions.

Introduction

Electric vehicle (EV) users are particularly impressed by the lack of local air and noise pollution and the low maintenance costs in comparison to conventional combustion-engine-driven cars. However, these users are seriously concerned about EV driving range [1]. Apart from range, the power available for driving the vehicle in different situations, e.g. for acceleration, is important for EV operation. The range of an electric vehicle corresponds mainly to the battery pack's capacity, and the available power relates to the resistance of the battery pack. With time and usage, capacity and resistance experience degradation, which results in another issue with EVs: limited battery lifetime.

An estimation of the battery's capacity and resistance on-board is thus essential in order to rate the battery's performance during operation, i.e. what peak power can be reached or what range to expect. Capacity and resistance estimations provide also a state-of-health (SOH) indication of the battery pack. A reliable SOH estimation not only ensures safe operation, it can also contribute to a smart optimization of battery usage resulting in an eventually extended lifetime.

Battery SOH estimation methods that are described in the scientific literature, however, suffer from one or several shortcomings with respect to on-board application such as battery-specificity, high computational effort, extensive preceding laboratory work (e.g. aging experiments or determination of physical properties), questionable validity for EV operating conditions (e.g. estimations deduced from usage history in laboratory studies) or need for operation interruptions and additional equipment [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16]. In a recent article [17], we have therefore introduced a novel method that accomplishes battery performance estimation without prior battery knowledge or experimental preparation. We applied an established machine learning method called support vector machines (SVM) to SOH estimation of a battery pack in a plug-in hybrid electric vehicle (PHEV). Without any experimental preparation or need for prior information about the specific battery, data readily available from the battery management system (BMS) collected in PHEV field tests was analyzed with SVM to build a battery model that captured the behavior of the battery at a reasonable computational load. More precisely, the voltage was modeled based on current and state-of-charge (SOC). The battery resistance was then estimated by running a virtual test on the battery model. This virtual test consisted of the current and SOC profiles from a real standard test and yielded a voltage estimation. In that way, the ease of statistical learning modeling from online available signals was combined with the advantage of the general validity of standard performance tests.

Purely data-driven methods for SOH estimation are rather uncommon in literature. There have, however, been several examples where statistical learning methods such as artificial neural networks (ANN) and SVM have successfully been applied for the related SOC estimation [18], [19], [20], [21]. Such black box models do not contribute to an improved understanding of the processes in the battery like physical models do. They have though the advantage of being battery-unspecific, as they – in contrast to physical models, empirical models based on aging experiments or models based on equivalent circuits – do not rely on predetermined system parameters or have any connection to physical properties. The aim of such statistical methods is to learn the behavior of the studied system from a large number of examples and find a mathematical system description. ANN suffer from several local minima, whereas SVM find one global solution. SVM, which also can handle non-linear systems, outperform ordinary regression due to its insensitivity to small changes. Popular application areas for SVM regression are e.g. bioinformatics, financial time series and electricity load forecasting [22], [23], [24]. However, a general limitation with data-driven models is that the methods are only valid within the trained data range.

Apart from Ref. [17], we know of one other paper that applied SVM in the context of on-board SOH estimation. That method, however, has a different focus. In Ref. [25], SVM were used to learn the capacity degradation behavior of a battery and to predict remaining useful life (RUL) whereas we use SVM to capture the battery behavior at a point in time by voltage modeling from current and SOC and then estimate SOH indicators via a virtual test of the battery model.

The present study constitutes a continuation of our previous paper [17]. We set here the PHEV field-tested, but non-validated method from that study into a controlled laboratory environment. In contrast to the proof-of-concept character of the previous study, this study aims at increasing understanding of the method's abilities and restrictions with the help of careful investigations enabled by the possibilities of an experimental study. Most decisively, estimation results are to be thoroughly compared to real standard performance tests. Moreover, as it was found in Ref. [17] that temperature is a crucial variable in real-life application, temperature dependence is introduced into the model. Also, the previously presented resistance estimations complemented with capacity estimation, which is an at least equally important SOH indicator for electric vehicle operation, which this study focuses on.

Fig. 1 visualizes the scope of this article. The study compares the SOH figures of merit, capacity and resistance, which are determined in two different ways. The conventional way, which serves as validation, is the derivation of performance values from direct measurements with standard performance tests. Then, the SOH estimation method from Ref. [17] with its two-step procedure of SVM training and virtual tests is applied to battery data generated with a typical EV current profile. In a first step, a SVM is trained with current, temperature and SOC as input and voltage as output. The resulting SVM model serves then as voltage look-up table for hypothetical current/temperature/SOC-input according to the measured current/temperature/SOC-profile from the respective real standard test. The virtual test result can subsequently be used to derive resistance and capacity as in real standard performance tests. The question to be answered is if the capacity and resistance estimations are sufficiently accurate to replace a direct measurement.

Section snippets

SOH figures of merit

Capacity and resistance are the two properties of a battery cell that are commonly used as measures of battery performance in conventional testing. They capture the most important characteristics of the behavior of the battery and allow therefore for documenting performance degradation with time.

The discharge capacity Qdischarge in Ah can be obtained by numerically integrating the current I of a full discharge between specified voltage limits over the discharge time t.Qdischarge=It

The

Results and discussion

This section treats at first exemplary results from experiments (Section 3.1) and SVM modeling (Section 3.2) separately to then turn to a more extensive side-by-side comparison of real and virtual test results including a qualitative evaluation of estimation performance for capacity and resistance (Section 3.3). The results section closes with an overview of the SVM training and test performance in terms of required computational effort (Section 3.4) and a description of an EV implementation

Conclusions

A SOH estimation method based on SVM models and virtual standard performance tests has been validated, developed further and evaluated. It requires neither preceding laboratory work, nor operation interruptions and additional equipment.

SVM are found to be a powerful method for handling large amounts of battery data. Battery models can conveniently be created from sets of current, voltage, SOC/Ah and temperature data. In real-life applications where a large range of operating temperatures can be

Acknowledgments

This work is supported by the Swedish Hybrid Vehicle Centre (SHC). We thank Anders Bern at Volvo Car Corporation for kindly providing us with field test data as well as the battery cells for laboratory testing. Thanks also to Pontus Svens and Scania for lending us equipment.

References (35)

  • J. Belt et al.

    J. Power Sources

    (2011)
  • K.L. Gering et al.

    J. Power Sources

    (2011)
  • M. Dubarry et al.

    J. Power Sources

    (2007)
  • A. Eddahech et al.

    Int. J. Electr. Power

    (2012)
  • U. Tröltzsch et al.

    Electrochim. Acta

    (2006)
  • G.L. Plett

    J. Power Sources

    (2004)
  • G.L. Plett

    J. Power Sources

    (2004)
  • G.L. Plett

    J. Power Sources

    (2004)
  • D. Andre et al.

    J. Power Sources

    (2013)
  • D. Andre et al.

    Eng. Appl. Artif. Intel.

    (2013)
  • Z. Chen et al.

    J. Power Sources

    (2013)
  • Y.-H. Chiang et al.

    J. Power Sources

    (2011)
  • T. Hansen et al.

    J. Power Sources

    (2005)
  • W. Junping et al.

    Energy Convers. Manag.

    (2006)
  • F.E.H. Tay et al.

    Omega

    (2001)
  • P. Pai et al.

    Electr. Power Syst. Res.

    (2005)
  • A. Nuhic et al.

    J. Power Sources

    (2013)
  • Cited by (258)

    View all citing articles on Scopus
    View full text