baseplett2004jps2-

Journal of Power Sources 134 (2004) 262276

Extended Kalman filtering for battery management systemsof LiPB-based HEV battery packsPart 2. Modeling and identification

Gregory L. Plett,1Department of Electrical and Computer Engineering, University of Colorado at Colorado Springs,

1420 Austin Bluffs Parkway, P.O. Box 7150, Colorado Springs, CO 80933-7150, USAReceived 27 January 2004; accepted 26 February 2004

Available online 28 May 2004

Abstract

Battery management systems in hybrid electric vehicle battery packs must estimate values descriptive of the packs present operatingcondition. These include: battery state of charge, power fade, capacity fade, and instantaneous available power. The estimation mechanismmust adapt to changing cell characteristics as cells age and therefore provide accurate estimates over the lifetime of the pack.

In a series of three papers, we propose a method, based on extended Kalman filtering (EKF), that is able to accomplish these goals ona lithium ion polymer battery pack. We expect that it will also work well on other battery chemistries. These papers cover the requiredmathematical background, cell modeling and system identification requirements, and the final solution, together with results.

In order to use EKF to estimate the desired quantities, we first require a mathematical model that can accurately capture the dynamics ofa cell. In this paper we evolve a suitable model from one that is very primitive to one that is more advanced and works well in practice.The final model includes terms that describe the dynamic contributions due to open-circuit voltage, ohmic loss, polarization time constants,electro-chemical hysteresis, and the effects of temperature. We also give a means, based on EKF, whereby the constant model parametersmay be determined from cell test data. Results are presented that demonstrate it is possible to achieve root-mean-squared modeling errorsmaller than the level of quantization error expected in an implementation. 2004 Elsevier B.V. All rights reserved.

Keywords: Battery management system (BMS); Hybrid electric vehicle (HEV); Extended Kalman filter (EKF); State of charge (SOC); State of health (SOH);Lithium-ion polymer battery (LiPB)

1. Introduction

This paper is the second in a series of three that de-scribe advanced algorithms for a battery management sys-tem (BMS) for hybrid electric vehicle (HEV) application.This BMS is able to estimate battery state of charge (SOC),power fade, capacity fade and instantaneous available power,and is able to adapt to changing cell characteristics over timeas the cells in the battery pack age. The algorithms havebeen implemented on a lithium-ion polymer battery (LiPB)

Tel.: +1-719-262-3468; fax: +1-719-262-3589.E-mail addresses: [email protected], [email protected](G.L. Plett).URL: http://mocha-java.uccs.edu.

1 The author is also consultant to Compact Power Inc., Monument, CO80132, USA. Tel.: +1-719-488-1600; fax: +1-719-487-9485.

pack, but we expect them to work well for other batterychemistries.

The method that we use to estimate these parameters isbased on Kalman filter theory. (There have been other re-ported methods for SOC estimation that use Kalman filter-ing [1,2], but the method in this series of papers expands onthese results and also differs in some important respects, aswill be outlined later.) Kalman filters are an intelligentandsometimes optimalmeans for estimating the state of a dy-namic system. By modeling our battery system to includethe wanted unknown quantities in the state, we may usethe Kalman filter to estimate their values. An additional ben-efit of the Kalman filter is that it automatically providesdynamic error-bounds on these estimates as well. We ex-ploit this fact to give aggressive performance from our bat-tery pack, without fear of causing damage by overcharge oroverdischarge.

0378-7753/$ see front matter 2004 Elsevier B.V. All rights reserved.doi:10.1016/j.jpowsour.2004.02.032

G.L. Plett / Journal of Power Sources 134 (2004) 262276 263

The first paper [3] is an introduction to the problem. Itdescribes the HEV environment and the requirement spec-ifications for a BMS. The remainder of the paper is a brieftutorial on the Kalman filter theory necessary to grasp thecontent of the remaining papers; additionally, a nonlinearextension called the extended Kalman filter (EKF) isdiscussed.

This second paper describes some mathematical cell mod-els that may be used with this method. The HEV applicationis a very harsh environment, with rate requirements up to andexceeding 20C and very dynamic rate profiles. This is incontrast to relatively benign portable-electronic applicationswith constant power output and fractional C rates. Methodsfor estimating SOC that work well in portable-electronic de-vices may not work well in the HEV application. If preciseSOC estimation is required by the HEV, then a very accuratecell model is necessary.

Results of lab tests on physical cells are presented andcompared with model prediction. The best modeling resultsobtained to date are so precise that the root-mean-squared(RMS) estimation error is less than the quantization noisefloor expected in our battery management system design.More importantly, the model allows very precise SOCestimation, therefore allowing the vehicle controller to con-fidently use the battery packs full operating range withoutfear of over- or under-charging cells. This paper also givesan overview of other modeling methods in the literatureand shows how an EKF may be used to adaptively identifyunknown parameters in a cell model, in real time, given testdata.

The third paper [4] covers the real-time parameter estima-tion problem; namely, how to dynamically estimate SOC,power fade, capacity fade, available power and so forth. AnEKF is used in conjunction with the cell model. The cellmodel may be fixed, or may itself have adaptable parame-ters so that the model tracks cell aging effects. Details for apractical implementation are discussed.

We now proceed by briefly reviewing cell models in theliterature that have been proposed for SOC estimation. Weexplain why these do not meet the requirements presentedin [3]. Several models from Refs. [5,6] do meet the require-ments, and they are described in detail here, together withsome new models and results. A method for identifyingmodel parameters using an extended Kalman filter is pre-sented, followed by conclusions.

2. Standard cell-modeling methods for SOC estimation

The literature documents a number of cell-modeling meth-ods for SOC estimation. An excellent summary, in greaterdetail than can be presented here, may be found in refer-ence [7]. Here, we investigate to see whether any of thesemethods meets our needs. Recall that our application is tomodel cell dynamics for the purpose of SOC estimation inan HEV battery pack.

For this application, the cell model must be accurate for alloperating conditions. These include: very high rates (up toabout20C, far exceeding the low rates considered by manypapers in the literature for portable electronic applications),temperature variation in the automotive range of 30 to50 C, very dynamic rates (unlike the more benign portableelectronic and battery electric vehicle application). Chargingmust be accounted for in the model.

We also require non-invasive methods using only readilyavailable signals. This requirement is imposed by the HEVenvironment where the BMS has no direct control over cur-rent and voltage experienced by the battery packthese arein the domain of the vehicle controller and inverter. We mustrely on such measurements as instantaneous cell terminalvoltage, cell current and cell external temperature.

Our cell chemistry also limits the range of approaches wemight consider. Techniques specific to lead-acid chemistries,for example, are not appropriate for LiPB cells.

2.1. Laboratory and chemistry-dependent methods

Several methods for direct SOC estimation simply cannotbe used in our application:

1. A laboratory method for determining SOC is tocompletely discharge a cell, recording dischargedampere-hours, to determine its present remaining capac-ity. This is the most accurate SOC measurement tech-nique, but is impractical in HEV as the battery energyis wasted by the test, and the test cannot dynamicallyestimate SOC.

2. Chemistry-dependent methods for other chemistries,such as Coup de Fouet measurement, or measurementof electrolyte physical properties for lead-acid batteries,are all inappropriate (as our application uses LiPB cells).

3. Open-circuit voltage (OCV) measurements: If the cell isallowed to rest for a long period, its terminal voltage de-cays to OCV, and OCV may be used to infer SOC (vialookup table, for example). However, long periods (some-times hours) of battery inactivity must occur before theterminal voltage approaches OCV. This method may notbe used for dynamic SOC estimation. (Other complica-tions with this method include the dependence of OCVon temperature, and presence of terminal voltage hys-teresis, especially at low temperatures.)

2.2. Electro-chemical modeling

One approach to modeling cell electrical dynamics isto carefully consider, at the molecular level, the variousprocesses that occur within the cell. Accurate terminalvoltage prediction may be achieved by these models (seeRef. [8], for example). However, it would be difficult (ifpossible) to measure the many required physical param-eters on a cell-by-cell basis in a high-volume consumerproduct. We have not pursued this approach, although

264 G.L. Plett / Journal of Power Sources 134 (2004) 262276

we employ many of the macroscopic concepts in ourmodels.

2.3. Impedance spectroscopy

Another broad category of cell modeling involves mea-suring cell impedances over a wide range of ac frequenciesat different states of charge [913]. Values of the modelparameters are found by least-squares fitting to measuredimpedance values. SOC may be indirectly inferred by mea-suring present cell impedance and correlating them withknown impedances at various SOC levels. We must also dis-count this method for our application, as we have no directmethod to inject signals into cells to measure impedances.We rely on the vehicle to generate and dissipate all energyflowing through the battery pack.

2.4. Circuit models

A number of papers present equivalent circuit models ofcells [1417]. Typically, a high-valued capacitor or voltagesource is used to represent the open-circuit voltage. The re-mainder of the circuit models the cells internal resistanceand more dynamic effects such as terminal voltage relax-ation. From the OCV estimate, SOC may be inferred viatable lookup. Both linear- and nonlinear-circuit models maybe used. Our model has many similarities to a circuit model,except that our fundamental state is SOC, not OCV.

2.5. Coulomb counting

The final method discussed in the literature involves SOCestimation directly via Coulomb counting. This may be doneopen-loop, which is very sensitive to current measurementerror, or closed-loop which can be much more accurate.The feedback mechanism may be empirically designed [18]or may use a mathematically optimized approach such asthe Kalman filtering method [1,2] to generate the feedback.All Kalman filtering-based methods by other authors in theliterature (with which we are familiar) use a circuit model ofthe cell with voltage sources and capacitor voltages repre-senting OCV and relaxation effects. OCV may be estimatedand SOC inferred from OCV.

Our approach is also based on the Kalman filteringmethod, but the fundamental aspect of our model that sets itapart from those reported in the literature is that SOC itselfis required to be a state of the system. The direct benefit ofthis approach is that the Kalman filter automatically givesa dynamic estimate of the SOC and its uncertainty (thisis discussed in greater detail in Ref. [4]). That is, insteadof reporting the SOC to the vehicle controller (at somepoint in time) to be about 55%, the algorithm is able toreport that the SOC is 55 3%, for example. This allowsthe vehicle controller to confidently use the battery packsfull operating range without fear of over- or under-chargingcells.

3. An evolution of cell model structures

In order to use Kalman-based methods for a batterymanagement system, we must first have a cell model in adiscrete-time state-space form. Specifically, we assume theform

xk+1 = f(xk, uk)+ wk, (1)yk = g(xk, uk)+ vk, (2)

where xk is the system state vector at discrete-time index k,where the state of a system comprises in summary formthe total effect of past inputs on the system operation sothat the present output may be predicted solely as a functionof the state and present input. Values of past inputs are notrequired. The vector uk is the measured exogenous systeminput at time k and wk is unmeasured process noise thataffects the system state. The system output is yk and vk ismeasurement noise, which does not affect the system state.Equation (1) is called the state equation, (2) is called theoutput equation, and f(, ) and g(, ) are (possibly non-linear) functions, specified by the particular cell model used.All of the system dynamics are represented in (1). Equation(2) is a static relationship. In the models to follow, wk isused to account for current-sensor error and inaccuracy ofthe state equation, and vk is used to account for voltage sen-sor error and inaccuracy of the output equation.

In the case where we wish to model a cells dynamicsusing (1) and (2), the vector uk contains the instantaneouscell current ik. It may also contain the cell temperature Tk,an estimate of the cells capacity C, and/or an estimate ofthe cells internal resistance Rk, for example. The systemoutput is typically a scalar but may be vector valued as well.Here we consider the output to be the cells loaded terminalvoltage (not its at-rest OCV). Our method constrains thestate vector xk to include SOC as one component.

There are many possible candidates for (1) and (2), and forthe choice of xk and uk. Here, we describe the developmentof some modeling equations in an evolutionary sense. Thatis, we start with a very simple model, and gradually addcomplexity to better represent the true cell dynamics. Inorder to justify the changes in the model, we compare modeldynamics to cell dynamics based on data collected from celltests. We first describe the cell tests, and then develop themodel structures.

3.1. Cell tests for model fitting

In order to compare the abilities of the proposed models tocapture a cells dynamics, we gathered data from a prototypeLiPB cell. The cell comprises a LiMn2O4 cathode, an artifi-cial graphite anode, is designed for high-power applications,has a nominal capacity of 7.5 Ah and a nominal voltage of3.8 V. For the tests, we used a Tenney thermal chamber set at25C and an Arbin BT2000 cell cycler. Each channel of theArbin was capable of 20 A current, and ten channels were


Fig. 1. Plots showing SOC vs. time and rate vs. time for pulsed-current cell tests. Discharge portion of test is shown in (a); charge portion of test isshown in (b). Dark line is SOC, gray line is current.

connected in parallel to achieve currents of up to 200 A. Thecyclers voltage measurement accuracy was 5 mV and itscurrent measurement accuracy was 200 mA.

Two types of cell tests were performed for the work re-ported in this paper. The first type comprised a sequence ofconstant-current discharge pulses and rests followed by a se-quence of constant-current charge pulses and rests. The cellstarted fully charged before the test began. Discharge currentpulses from 150 down to 1 A, and charge pulses from 150down to 1 A were used. The current and SOC profiles forthis test are shown in Fig. 1(a) and (b). Frame (a) shows thedischarge portion of the test and frame (b) shows the chargeportion of the test. Data points (including voltage, current,ampere-hours discharged and ampere-hours charged) werecollected once per second.

The second test was a sequence of 16 urban dynamome-ter driving schedule (UDDS) cycles, separated by 40 A dis-charge pulses and 5 min rests, and spread over the 9010%SOC range. The SOC as a function of time is plotted inFig. 2(a), and rate as a function of time for one of the UDDScycles is plotted in Fig. 2(b). We see that SOC increases byabout 5% during each UDDS cycle, but is brought downabout 10% during each discharge between cycles. The en-tire operating range for these cells (1090% SOC) is excitedduring the cell test.

0 50 100 150 200 250 300 350 400 4500

102030405060708090

100SOC as a function of time

Time (min.)

SOC

(perce

nt)

205 215 225 235 245-60

-40

-20

0

20

40

60

80Current for one UDDS cycle (zoom)

Time (min.)

Curre

nt (A

)

(b)(a)

Fig. 2. Plots showing SOC vs. time and rate vs. time for UDDS cell tests. SOC is shown in (a); rate for one UDDS cycle is shown in (b).

The data was used to identify parameters of the cell mod-els to be described in the next sections. The goal is to havethe cell model output resemble the cell terminal voltage un-der load as closely as possible, at all times, when the cellmodel input is equal to the cell current. Model fit was judgedby comparing root-mean-squared estimation error (estima-tion error equals cell voltage minus model voltage) over theportions of the cell tests where SOC was between 5 and95%. Model error outside that SOC range was not consid-ered as the HEV pack operation design limits are 1090%SOC.

3.2. SOC as a state-vector component

The one constant requirement for the cell model is thatwe constrain SOC, denoted as zk, to be a member of thestate vector xk. To be careful, we give a list of definitionsculminating in our understood definition of SOC.

Definition: A cell is fully charged when its voltage reachesv = vh after being charged at infinitesimal current levels.Here, we use vh = 4.2 V at room temperature (25 C).

Definition: A cell is fully discharged when its voltagereaches v = vl after being drained at infinitesimal currentlevels. Here, we use vl = 3.0 V at room temperature.


Definition: The capacity of a cell C is the maximum num-ber of ampere-hours that can be drawn from the cell be-fore it is fully discharged, at room temperature, startingwith the cell fully charged.

Definition: The nominal capacity Cn of the cell is thenumber of ampere-hours that can be drawn from the cellat room temperature at the C/30 rate, starting with the cellfully charged.

Definition: The SOC of the cell is the ratio of the remain-ing capacity to the nominal capacity of the cell, where theremaining capacity is the number of ampere-hours thatcan be drawn from the cell at room temperature at theC/30 rate before it is fully discharged.

With these definitions in place, we can then investigatesome mathematical relations involving SOC. Particularly,

z(t) = z(0) t

0

ii()

Cnd, (3)

where z(t) is the cell SOC, i(t) is instantaneous cell current(assumed positive for discharge, negative for charge), andCn is the cell nominal capacity. Cell Coulombic efficiencyi is i = 1 for discharge, and i = 1 for charge.

Using a rectangular approximation for integration and asuitably small sampling periodt, a discrete-time approx-imate recurrence may then be written as

zk+1 = zk (it

Cn

)ik. (4)

Eq. (4) is the basis for including SOC in the state vector ofthe cell model as it is in state equation format already, withSOC as the state and ik as the input.

The cell model may be completed by adding additionalstates, as necessary, and an output equation. Here, we firstrevisit an output equation from an earlier paper [5], and showhow it may be enhanced. Next, we add a state to the modelto account for cell hysteresis. Thirdly, we add dynamics tothe state to model cell terminal voltage relaxation. Finally,we discuss adding temperature dependence to the models.

3.3. Models with only SOC as a state

The first three model structures that we investigate havestate vector xk = zk. That is, the only state in the stateequation (1) is SOC. These models can estimate cell terminalvoltage in a limited way, and are improved upon later usingmultiple states.

3.3.1. The combined modelWith SOC available as part of the model state, terminal

voltage may be predicted in a number of different ways.Several different forms are adapted from reference [19]. Shepherd model: yk = E0 Rik Ki/zk. Unnewehr universal model: yk = E0 Rik Kizk. Nernst model: yk = E0Rik+K2 ln(zk)+K3 ln(1zk).

In these models, yk is the cell terminal voltage, R is thecell internal resistance (different values may be used forcharge/discharge and at different SOC levels if desired),Ki isthe polarization resistance and K1, K2 and K3 are constantschosen to make the model fit the data well. All of the terms ofthese models may be collected to make a combined modelthat performs better than any of the individual models alone.This model is

zk+1 = zk (it

Cn

)ik,

yk = K0 Rik K1zk

K2zk +K3 ln(zk)+K4 ln(1 zk).

The unknown quantities in the combined model may be es-timated using a system identification procedure. This modelhas the advantage of being linear in the parameters; that is,the unknowns occur linearly in the output equation. Givena set of N cell inputoutput three-tuples {yk, ik, zk}, the pa-rameters may be solved for in closed form using a resultfrom least-squares estimation. This simple off-line (batch)method is as follows: We first form the vector

Y = [y1, y2, . . . , yN]T,and the matrix

H = [h1, h2, . . . , hN]T.The rows of H are

hTj =[

1, i+j , ij ,

1zj, zj, ln(zj), ln(1 zj)

],

where i+j is equal to ij if ij > 0, ij is equal to ij if ij ,1, ik < ,sk1, |ik| .

M(zk) is half the difference between the two legs of thecharge/discharge curve, minus the Rik loss, and is plotted inFig. 5(b). Here, we use a constant value for M.

4 Successively smaller concentric minor loops may be obtained byalternating shorter and shorter charge and discharge pulses, eventuallyconverging on the mean of the two values of Fig. 5(a) at each SOC.Therefore, we compute OCV as a function of SOC as the mean of thetwo legs of the major hysteresis loop.


This model type is also linear in the parameters. Off-linesystem identification is done as follows: We first form thevector

Y= [y1 OCV(z1), y2 OCV(z2), . . . , yn OCV(zN)]T,

and the matrix

H = [h1, h2, . . . , hN]T.

The rows of H are

hTj = [i+j , ij , sj].

0 10 20 30 40 50 60 70

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling discharge: Combined model

Time (min.)

Volta

ge (V

)

True voltageEstimated voltage

0 10 20 30 40 50 60 70 80 90

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling charge: Combined model

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling discharge: Simple model

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70 80 90

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling charge: Simple model

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling discharge: Zero state hysteresis model

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70 80 90

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling charge: Zero state hysteresis model

Time (min.)

Volta

ge (V

)


(a) (d)

(b)

(c) (f)

(e)

Fig. 6. Results of cell modeling using models with only SOC as a state for the pulsed-current cell tests. Discharge portion of test is shown in (a)(c);charge portion of test is shown in (d)(f). The gray line is the measured cell voltage, and the black line is the model prediction.

Again, we see that Y = H, where T = [R+, R,M] is thevector of unknown parameters. We solve for the parameters using the known matrices Y and H as = (HTH)1HTY .

Results comparing the zero-state hysteresis model cellvoltage estimation with the cells true voltage for thepulsed-current test are shown in Figs. 6(c) and (f). Fig. 6(c),shows the comparison over discharge pulses, and Fig. 6(f)shows the comparison over charge pulses. The RMS cellmodel estimation error over the test shown in Fig. 6 islisted in Table 1. Performance of the zero-state hysteresismodel is consistently better than that of the simple model.Results comparing the zero-state hysteresis model cell volt-age estimation with the cells true voltage for one cycle of


the UDDS test are shown in Fig. 8(c). Similar commentsapply.

3.4. Models with SOC and additional states

In order to better estimate cell voltage effects that are cou-pled to the history of the cells input current, we must makemodifications to the model state equation (1). We examinetwo additions in the following sections.

3.4.1. The one-state hysteresis modelThe zero-state hysteresis model is an improvement over

the simple model, but only crudely approximates the under-lying phenomenon. Whereas the level of hysteresis slowlychanges as the cell is charged or discharged, the model es-timates hysteresis as immediately flipping between its max-imum positive and negative values when the sign of currentchanges.

The slow transition may be modeled by adding a hys-teresis state to the model state equation (1). The hystere-sis state is not a differential equation in time, but in SOC(or, ampere-hours). Let h(z, t) be the hysteresis voltage as afunction of SOC and time, and let z = dz/dt. Then,dh(z, t)

dz= sgn(z)(M(z, z) h(z, t)),

where M(z, z) is a function that gives the maximum po-larization due to hysteresis as a function of SOC and therate-of-change of SOC. Specifically, M(z, z) is positive forcharge (z > 0) and is negative for discharge (z < 0). TheM(z, z) h(z, t) term in the differential equation states thatthe rate-of-change of hysteresis voltage is proportional to thedistance away from the major hysteresis loop, leading to akind of exponential decay of voltage to the major loop. Theterm in front of this has a positive constant , which tunesthe rate of decay, and sgn(z), which forces the equation tobe stable for both charge and discharge.

In order to fit the differential equation for h(z, t) into ourmodel, we must manipulate it to be a differential equation intime, not in SOC. We accomplish this by multiplying bothsides of the equation by dz/dt

dh(z, t)dz

dzdt

= sgn(z)(M(z, z) h(z, t))dzdt.

Note that dz/dt = ii(t)/Cn, and that z sgn(z) = |z|. Thus,

h(t) = ii(t)Cn

h(t)+ii(t)Cn

M(z, z).This may be converted into a difference equation for ourdiscrete-time application using standard techniques (assum-ing that i(t) andM(z, z) are constant over the sample period):

hk+1 = exp(iiktCn

)hk

+(

1 exp(iiktCn

))

M(z, z).

Note that this is a linear-time-varying system as the factorsmultiplying the state and input change with ik and hencewith time. If we define F(ik) = exp(|iikt/Cn|), thenthe overall state-space equations for the one-state hysteresismodel are[hk+1zk+1

]=[F(ik) 0

0 1

][hk

zk

]

+ 0 (1 F(ik))it

Cn0

[ ik

M(z, z)

],

yk = OCV(zk) Rik + hk.

Results comparing the one-state hysteresis model cellvoltage estimation with the cells true voltage for thepulsed-current test are shown in Fig. 7(a) and (d). Fig. 7(a),shows the comparison over discharge pulses, and Fig. 7(d)shows the comparison over charge pulses. The RMS cellmodel estimation error over the test shown in Fig. 7 islisted in Table 1. Performance of the one-state hysteresismodel is consistently better than the simpler models. Re-sults comparing the one-state hysteresis model cell voltageestimation with the cells true voltage for one cycle of theUDDS test are shown in Fig. 8(d). Similar comments apply.

3.4.2. The enhanced self-correcting (ESC) modelA significant element missing from these models is a de-

scription of time constants during pulsed current events. Ifa cell is allowed to rest, it takes some time for the voltageto completely relax to its rest voltage. If a cell is pulsedwith current, it takes time for the voltage to converge to itssteady-state level. These time constants, which describe thephenomenon we henceforth refer to as the relaxation effect,may be implemented as a low-pass filter on ik. Since thecell model must accurately predict its behavior in a dynamicHEV environment, we find it is essential to include relax-ation effects.

Early attempts to model the relaxation effect includedfiltering the state-of-charge as well as the input current (cf.the filter state cell model [5]). While this model couldfit voltage data reasonably well, it had the unfortunate sideeffect that SOC estimation using an EKF was not reliable.The output equation had the form

yk = OCV(zk)+ filt(zk)+ filt(ik) Rik,

where filt() is some dynamic operation filtering its operand.The EKF had problems with this model because thecredit/blame portion of the algorithm could not reliably de-termine whether model error was due to bad SOC throughthe OCV() function or through the filt(zk) function, ordue to bad filter states in the filt(ik) function. In particular,some simple cell tests running the EKF showed a lack ofrobustness:


0 10 20 30 40 50 60 70

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling discharge: One state hysteresis

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70 80 90

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling charge: One state hysteresis

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling discharge: ESC, 2 filter states

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70 80 90

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling charge: ESC, 2 filter states

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling discharge: ESC, 4 filter states

Time (min.)

Volta

ge (V

)


0 10 20 30 40 50 60 70 80 90

3.0

3.2

3.4

3.6

3.8

4.0

4.2

Modeling charge: ESC, 4 filter states

Time (min.)

Volta

ge (V

)


(a) (d)

(b) (e)

(c) (f)

Fig. 7. Results of cell modeling using models with multiple states for the pulsed-current cell tests. Discharge portion of test is shown in (a)(c); chargeportion of test is shown in (d)(f). The gray line is the measured cell voltage, and the black line is the model prediction.

1. A constant-current discharge/charge should make theSOC ramp down/up at the slope i/Cn (A/Ah). In practice,the slope using the filter state model was often wrong.

2. During a rest period, cell terminal voltage converges toOCV (neglecting hysteresis effects) and estimated SOCshould converge to the SOC predicted by OCV. In theimplementation, we observed SOC to drift considerably,not converging to the correct value.

The model that we will develop in this section, called theenhanced self-correcting model, forces yk to converge toOCV after a rest period and it forces yk to converge to

OCVRik for a constant-current discharge/charge. To meetthese requirements, with hysteresis added, the output equa-tion needs to have the form

yk = OCV(zk) fn(zk)

+ hkfn(zk,ik)

+ filt(ik) Rik fn(ik)

.

In this equation, SOC and hysteresis contribute the long-termdc level (bias) to the output and ik and its history contributethe short-term variation around this level. SOC itself is nolonger filtered as in the filter state modelit makes nosense to have a moving bias point.


205 215 225 235 2453.7

3.8

3.9

4.0

4.1

4.2Modeling (zoom): Combined model

Time (min.)

Volta

ge (V

)


205 215 225 235 2453.7

3.8

3.9

4.0

4.1

4.2Modeling (zoom): Simple model

Time (min.)

Volta

ge (V

)


205 215 225 235 2453.7

3.8

3.9

4.0

4.1

4.2Modeling (zoom): Zero state hysteresis model

Time (min.)

Volta

ge (V

)


205 215 225 235 2453.7

3.8

3.9

4.0

4.1

4.2Modeling (zoom): One state hysteresis

Time (min.)

Volta

ge (V

)


205 215 225 235 2453.7

3.8

3.9

4.0

4.1

4.2Modeling (zoom): ESC, 2 filter states

Time (min.)

Volta

ge (V

)


205 215 225 235 2453.7

3.8

3.9

4.0

4.1

4.2Modeling (zoom): ESC, 4 filter states

Time (min.)

Volta

ge (V

)


(a) (d)

(b) (e)

(c) (f)

Fig. 8. Results of cell modeling for the UDDS cell tests. The gray line is the measured cell voltage, and the black line is the model prediction.

The filter filt() must satisfy two criteria: (1) after along rest period its output must be zero so that yk OCV + hk; (2) during a constant-current discharge/charge,its output must converge to zero so that yk OCV +hk Rik. The first criterion is satisfied by a stable linearfilter, and the second is satisfied by a linear filter with zerodc gain. Both of these may be enforced in the filter de-sign.

A linear filter may be implemented in a state-space formas

fk+1 = Affk + Bf ik,yfk = Gfk,

where Af is the state-transition matrix of the filter, Bf is theinput matrix of the filter, G is the output matrix of the filter,and fk is the filter state. The eigenvalues of the Af matrixare the poles of the filter and determine its stability. Thefilter is stable if max|eig(Af)| < 1. The location of poles de-termine the systems dynamic behavior: poles near +1 haveslowly decaying dynamics, poles near zero decay quickly,negative poles oscillate. Complex-conjugate poles also oscil-late, and do not appear to improve performance by their in-clusion. Therefore, it is sufficient to have an Af matrix of theform Af = diag(), where is a vector comprising the polelocations. Stability is ensured if all 1 < j < 1. The Bfmatrix may be chosen arbitrarily so long as no entry is zero.


Now that we have guaranteed stability, yk OCV + hkduring rest. By carefully selecting the G matrix, we canensure a zero dc-gain of the filter so that yk OCV+hkRik during constant-current profiles. The gain of the filter is(where nf is the number of filter states, determined duringsystem identification to fit cell data):

G(I A)1B = 0,G

[diag

(1

1 k

)]B = 0,

nfk=1

gk

1 k = 0.

If we let g1 through gnf1 be found by a system-identificationprocedure and assuming that Bf = [ 1 1 ]T, then thezero dc-gain constraint fixes gnf as

gnf = nfk=1

gk(1 nf )(1 k) .

So, the full self-correcting model is fk+1hk+1zk+1

=

diag() 0 00 F(ik) 0

0 0 1

fkhkzk

+

1 00 (1 F(ik))

itCn

0

[

ik

M(z, z)

],

yk = OCV(zk) Rik + hk + Gfk.Results comparing the ESC model cell voltage estimationwith the cells true voltage for the pulsed-current test areshown in Figs. 7(b) and (e) for nf = 2 and in Figs. 7(c)and (f) for nf = 4. Figs. 7(b) and (c), show the comparisonover discharge pulses, and Figs. 7(e) and (f) show the com-parison over charge pulses. The RMS cell model estimationerror over the tests shown in Fig. 7 are listed in Table 1. Per-formance is significantly improved by the addition of filter

30 25 20 15 10 5 0 5 10 15 20 25 30 35 40 450

102030405060708090

100

Temperature (C)

RM

S m

odel

ing

erro

r (mV

)

RMS modeling error versus temperatureCombined modelSimple modelZero state hysteresisOne state hysteresisESC, nf=2ESC, nf=4

0 10 20 30 40 50 60

11.8

19.1

22.9

35.0

53.8

32.0

Zero-state Hysteresis

Simple

Combined

ESC, nf=4

ESC, nf=2

One state Hysteresis

Average RMS modeling error (mV)

Mod

el s

truct

ure

Summary modeling performance over temperature

(a) (b)

Fig. 9. Results when modeling over a temperature range. Frame (a) shows the individual modeling results, and frame (b) compares the average modelingresults.

states. We do not see much improvement by increasing nf be-yond 4. Results comparing the ESC model cell voltage esti-mation with the cells true voltage for one cycle of the UDDStest are shown in Figs. 8(e) for nf = 2 and (f) for nf = 4.

3.5. Adding temperature dependence to the models

Thus far, we have discussed only how to model cell dy-namics at one specific temperature. We now embark on abrief discussion on how to incorporate temperature depen-dence into the models.

A very simple method, and the one we tried first, was touse a table of different models, where each model had pa-rameters optimized for a specific temperature. For example,we used sixteen models over the temperature range 30 to45 C in increments of 5 C. This worked well so long asthe cell under test had temperature equivalent to one of thesixteen stored models. If the temperature was between twostored model values, we linearly interpolated model param-eters between the parameters of the models in the table. Thisdid not work well.

We found that two adjacent models in the table did notnecessarily have similar parameters. Individually optimiz-ing model parameters at specific temperatures resulted invalues that were over-fit to the data and did not generalizewell to cases not previously seen. We remedied the problemby performing joint optimization over the entire temperaturerange, where every parameter was represented by a contin-uous polynomial of temperature (fourth order). This forcednearby models to have similar parameter values. Althoughjoint optimization did not result in modeling errors as lowas when individually optimized, the generalization perfor-mance was much better.

Fig. 9 shows the results for joint optimization overthe entire temperature range, and Table 1 lists some nu-meric values corresponding to the plots. The cell data wascollected from UDDS tests similar to those described inSection 3.1, but performed at 16 controlled temperaturesfrom 30 to 45 C, in increments of 5C. At lower temper-atures, the magnitude of the current had to be scaled downso as not to exceed voltage limits (due to increased cell


resistance at lower temperatures), and hence more UDDScycles had to be completed to cover the desired SOC range,but the tests were otherwise the same. In frame (a), theRMS modeling errors for the jointly optimized models areplotted versus temperature for the different cell models.In frame (b), the average RMS modeling errors over alltemperatures are presented as a bar graph.

We see that the combined model appears to perform well.However, this is an artifact. The {K0, . . . , K4} points werefound to over-fit the measured data, so that the resultingcurve plotted with their values did not resemble OCV to anydegree of fidelity. While the simple model has worse nu-meric indicators of performance, it generalized better thanthe combined model. By adding the crude model of hystere-sis, we see a significant performance jump, especially at coldtemperatures where hysteresis is most evident in our cellsdynamics. Adding the dynamics of a state to the hysteresismodel also improves performance, again with the greatestgains at low temperatures. Filter states also contribute to per-formance gains. The final model, enhanced-self-correctingwith nf = 4 gave the best performance in all cases. Byincreasing the number of filter states we would expect con-tinued performance gains, at the cost of greater complexity.Note that the cold-temperature performance is not improvedas much, in relative terms, by adding filter states, so it islikely that some cold-temperature phenomena is not yetbeing modeled well. Conceivably, a second hysteresis statecould be added to the model to improve performance here.We have not investigated this possibility as yet.

4. System identification

The first three system models introduced in this paper arelinear in the parameters. This makes identifying the valuesof the model parameters straightforward using least-squaresestimation, and has been discussed earlier. When the modelis not linear in the parameters, as in the remaining systemmodels, this method may not be used. We must turn to moreadvanced methods.

Here we look at one method in particular. We know thata Kalman filter or extended Kalman filter may be used toestimate the state of a dynamic system given noisy mea-surements; e.g., to estimate the cell SOC. We may also usean extended Kalman filter to perform system identificationgiven clean measurements. To do so, we require a state-spacemodel describing the dynamics of the parameters of thesystem model. We will use the Kalman filter as an optimumobserver of these parameter values, creating an estimate .In electro-chemical cells, the true parameters will changeonly very slowly, so we model them as constant with somesmall perturbation:

k+1 = k + rk.The small white noise input rk is fictitious, but modelsthe slow change in the parameters of the system plus the

Table 2Summary of the nonlinear extended Kalman filter for system identification[21]Nonlinear state-space modela

k+1 = k + rkdk = g(xk, uk, k)+ ek

Definition

Ck =dg(xk, uk, )

d

=

k

InitializationFor k = 0, set+0 = E[0]+,0 = E[(0

+0 )(0 +0 )T]

ComputationFor k = 1, 2, . . . compute

State estimate time update: k = +k1Error covariance time update:

,k= +

,k1 +rKalman gain matrix: Lk = ,k(C

k)

T[Ck,k(Ck)

T +e]1

State estimate measurement update:+k = k + Lk[yk g(xk, uk, k )]

Error covariance measurement update: +,k

= (I LkCk),ka rk and ek are independent, zero-mean, Gaussian noise processes of

covariance matrices r and e, respectively.

infidelity of the model structure to capture all of the celldynamics.

The output equation required for Kalman-filter systemidentification must be a measurable function of the systemparameters. We usedk = g(xk, uk, k)+ ek,where g(, , ) is the output equation of the system modelbeing identified, and ek models the sensor noise and model-ing error. We compare dk computed using k to the measuredcell output, and adapt k to minimize the difference.

We can create an extended Kalman filter using thisstate-space model and cell data to estimate the system pa-rameters as summarized in Table 2. We initialize the stateestimate with our best information re. the state value: +0 =E[0], and the state estimation error covariance matrix:+,0

= E[( +0 )( +0 )T].The time update propagates the state estimate as k =

+k1 since the parameters are assumed constant, and theerror covariance as

,k= +

,k1 +r to account for theadded uncertainty due to the fictitious noise input rk. Theeffect of adding r is to increase the estimates uncertainty,and to allow adaptation to .

The extended Kalman filter gain matrix is computed bylinearizing the state-space models output equation. We com-pute

Ck =dg(xk, uk, )

d

=

k

,

Lk = ,k(Ck)

T[Ck,k(Ck)

T +e]1.


Next, we measure the true cell output yk and compare it to themodel output g(xk, uk, k ). To do so, we need to simulatethe model in parallel with the real cell to have an appropriatevalue of xk available. The difference between cell outputand model output can be attributed to noises and modelingerror. The extended Kalman filter adapts to minimize thisdifference

+k = k + Lk[yk g(xk, uk, k )].Finally, the measurement update is applied to the state errorcovariance matrix

+,k

= (I LkCk),k.This has the effect of decreasing the modeling uncertaintydue to the measurement update. All terms are accounted for,and the algorithm is complete.

4.1. Extended Kalman-filter system identification for theone-state hysteresis model

The details for applying this method to any particular cellmodel are differentiated by the calculation of Ck . For theone-state hysteresis model, let the vector of parameters be

= [R+ R M ]T.R+ is the cell resistance when current is positive, and Ris the cell resistance when current is negative. M is themaximum hysteresis voltage, and is the hysteresis rateconstant, which is part of F(ik). To calculate Ck we requiredg(xk, uk, )

d= g(xk, uk, )

+ g(xk, uk, )

xk

dxkd

, (6)

dxkd

= f(xk1, uk1, )

+ f(xk1, uk1, )xk1

dxk1d

. (7)

The derivative calculations are recursive in nature, andevolve over time as the state evolves. The term dx0/d isinitialized to zero unless side information gives a better es-timate of its value. We see that in order to calculate Ck forany specific model structure, we require methods to calcu-late the partial derivatives in (6) and (7). For the one-statehysteresis model we have

g(xk, uk, )

= [ i+ i 0 0 ], g(xk, uk, )

xk=[

1OCV(zk)

zk

],

f(xk1, uk1, )

= 0 0 (1 Fk1) sgn(ik1) (M hk1)

iik1tCFk1

0 0 0 0

, f(xk1, uk1, )

xk1=[Fk1 0

0 1

].

Note that the OCV(zk)/zk term is never needed, as italways multiplies zero. For this particular model, we cansimplify the calculations by removing the multiplies by zero:

dg(xk, uk, )d

= [ i+ i 0 0 ] + dhkd

,dhkd

=[

0 0 (1 Fk1) sgn(ik1)(M hk1)iik1tC

Fk1]+ Fk1 dhk1d .

4.2. Extended Kalman-filter system identification for theenhanced self-correcting model

For the enhanced self-correcting model, let the vector ofparameters be

= [R+ R g1 gnf1 1 nf M ]T,where = tanh () and is the vector of filter pole lo-cations. We use the tanh () function during system identi-fication because it forces filter poles to remain within 1(i.e., stable) regardless of the value of . When calculatingthe partial derivatives we must remember that since gnf =nf1i=1 gi(1 nf )/(1 i), it is not independently iden-tified but is computed from the g1, . . . , gnf1 terms. Thisalso forces the derivatives to be more complicated than aquick glance would indicate. That is,

g(xk, uk, )

g1= fk,1 1 nf1 1 fk,nf ,

and so forth. Also, since i = tanh (i), it can never beunity, so division by zero is impossible in the derivativecomputation. With this in mind, the partial derivativesin (6) and (7) may be computed in a straightforwardway.

5. Conclusions

This paper has proposed five mathematical state-spacestructures for the purpose of modeling LiPB HEV cell dy-namics for their eventual role in HEV BMS algorithms.Models with a single-state are very simple, but perform thepoorest. Adding hysteresis and filter states to the model aidsperformance, at some cost in complexity.

We have also seen how to identify the parameters of thecell models given cell-test data. Models that are linear in theparameters may have their parameters fit in a very straight-forward way using methods from least-squares-estimationtheory. Models with more dynamics than simply SOC re-quire more sophisticated techniques. One possibility is to

use an extended Kalman filter to identify the cell parametersin an on-line or off-line manner.


In the third paper [4], we will employ extended Kalmanfiltering from [3], using the cell models developed here, toimplement HEV BMS algorithms. We will see how to useEKF to estimate SOC and all other model states as the sys-tem operates. This model state will then allow us to accu-rately compute a dynamic estimate of available power. Wecan additionally employ a technique called dual extendedKalman filtering to simultaneously estimate cell state andparameters, allowing tracking of cell power fade and capac-ity fade, for example. Finally, the parameter data and SOCestimate may be combined to determine which cells in thepack must have their charge levels modified in order to bringthe pack into equalization.

Acknowledgements

This work was supported in part by Compact Power Inc.(CPI). The use of company facilities, and many enlighteningdiscussions with Drs. Mohamed Alamgir and Dan Riversand others are gratefully acknowledged.

References

[1] C. Barbier, H. Meyer, B. Nogarede, S. Bensaoud, A battery stateof charge indicator for electric vehicle, in: Proceedings of the In-ternational Conference of the Institution of Mechanical Engineers,Automotive Electronics, London, UK, 1719 May 1994, pp. 2934.

[2] P. Lrkens, W. Steffens, Ladezustandsschtzung von bleibatterienmit hilfe des Kalman-filters (State of charge estimation of lead-acidbatteries using a Kalman filtering technique), etzArchiv, vol. 8, No.7, July 1986, pp. 231236 (in German).

[3] G. Plett, Extended Kalman filtering for battery management systemsof LiPB-based HEV battery packs. Part 1. Background, J. PowerSour. 134 (2) (2004) 252261.

[4] G. Plett, Extended Kalman filtering for battery management systemsof LiPB-based HEV battery packs. Part 3. State and parameterestimation, J. Power Sour. 134 (2) (2004) 277292.

[5] G. Plett, LiPB dynamic cell models for Kalman-filter SOC estima-tion, in: Proceedings of the EVS-19, Busan, Korea, October 2002(CD-ROM).

[6] G. Plett, Advances in EKF SOC estimation for LiPB HEV batterypacks, in: Proceedings of the EVS-20, Long Beach, CA, November2003.

[7] S. Piller, M. Perrin, A. Jossen, Methods for state-of-charge deter-mination and their applications, J. Power Sour. 96 (2001) 113120.

[8] W.B. Gu, C.Y. Wang, Thermalelectrochemical modeling of batterysystems, J. Electrochem. Soc. 147 (8) (2000) 29102922.

[9] K. Takano, K. Nozaki, Y. Saito, A. Negishi, K. Kato, Y. Yamaguchi,Simulation study of electrical dynamic characteristics of lithium-ionbattery, J. Power Sour. 90 (2) (2000) 214223.

[10] E. Barsoukov, J. Kim, C. Yoon, H. Lee, Universal battery param-eterization to yield a non-linear equivalent circuit valid for batterysimulation at arbitrary load, J. Power Sour. 83 (1/2) (1999) 6170.

[11] S. Rodrigues, N. Munichandraiah, A. Shukla, A review of state-of-charge indication of batteries by means of ac impedance mea-surements, J. Power Sour. 87 (1/2) (2000) 1220.

[12] A. Salkind, C. Fennie, P. Singh, T. Atwater, D. Reisner, Determinationof state-of-charge and state-of-health of batteries by fuzzy logicmethodology, J. Power Sour. 80 (1/2) (1999) 293300.

[13] P. Singh, C. Fennie, D. Reisner, A. Salkind, Fuzzy logic-enhancedelectrochemical impedance spectroscopy (fleeis) to determine batterystate of charge, in: Proceedings of the 15th Annual Battery Confer-ence on Applications and Advances, Long Beach, CA, USA, 1114January 2000, pp. 199204.

[14] A. Kawamura, T. Yanagihara, State of charge estimation of sealedlead-acid batteries used for electric vehicles, in: Proceedings ofthe 29th Annual IEEE Power Electronics Specialists Conference,Fukuoka, Japan, 1722 May 1998, pp. 583587.

[15] S. Bhatikar, R. Mahajan, K. Kipke, V. Johnson, Neural network basedbattery modeling for hybrid electric vehicles, in: Proceedings of the2000 Future Car Congress, Paper No. 2000-01-1564, Arlington, VA,26 April 2000.

[16] H. Chan, D. Sutanto, A new battery model for use with battery en-ergy storage systems and electric vehicle power systems, in: Proceed-ings of the 2000 IEEE Power Engineering Society Winter Meeting,Singapore, 2327 January 2000, pp. 470475.

[17] V. Johnson, A. Pesaran, T. Sack, Temperature-dependent batterymodels for high-power lithium-ion batteries, in: Proceedings of the17th Electric Vehicle Symposium, Montreal, Canada, 1518 October2000.

[18] R. Giglioli, P. Pelacchi, M. Raugi, G. Zini, A state of charge observerfor lead-acid batteries, Energia Elettrica 65 (1) (1988) 2733.

[19] ThermoAnalytics Inc., Battery modeling for HEV simulationby ThermoAnalytics Inc. http://www.thermoanalytics.com/support/publications/batterymodelsdoc.html.

[20] V. Srinivasan, J. Weidner, J. Newman, Hysteresis during cycling ofnickel hydroxide active material, J. Electrochem. Soc. 148 (9) (2001)A969A980.

[21] E. Wan, A. Nelson, Dual extended Kalman filter methods, in:S. Haykin (Ed.), Kalman Filtering and Neural Networks, Wiley/Interscience, New York, 2001, pp. 123174.

Extended Kalman filtering for battery management systems of LiPB-based HEV battery packsPart 2. Modeling and identificationIntroductionStandard cell-modeling methods for SOC estimationLaboratory and chemistry-dependent methodsElectro-chemical modelingImpedance spectroscopyCircuit modelsCoulomb counting

An evolution of cell model structuresCell tests for model fittingSOC as a state-vector componentModels with only SOC as a stateThe combined modelThe simple modelThe zero-state hysteresis model

Models with SOC and additional statesThe one-state hysteresis modelThe enhanced self-correcting (ESC) model

Adding temperature dependence to the models

System identificationExtended Kalman-filter system identification for the one-state hysteresis modelExtended Kalman-filter system identification for the enhanced self-correcting model

ConclusionsAcknowledgementsReferences

baseplett2004jps2-

Documents

battery chemistries

battery pack age

battery state of charge

power fade

battery management system

cell modeling

instantaneous available

extended kalman filtering