Electrical Power and Energy Systems 42 (2012) 487–494

Contents lists available at SciVerse ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Behavior and state-of-health monitoring of Li-ion batteries using impedance

spectroscopy and recurrent neural networks
Akram Eddahech ⇑, Olivier Briat, Nicolas Bertrand, Jean-Yves Delétage, Jean-Michel Vinassa
UMR 5218 CNRS – IPB – Université Bordeaux 1, Laboratoire IMS, 351 Cours de la Libération, Bat A31, 33400 Talence, Bordeaux, France

a r t i c l e i n f o a b s t r a c t

Article history: Research into the monitoring of lithium-ion batteries has become increasingly important, due to their use
Received 19 October 2011 in a variety of complex, high-performance, energy-storage applications in hybrid and electric vehicles
Received in revised form 9 April 2012 (HEV and EV). This paper investigates the behavior and state-of-health monitoring of lithium-ion batter-
Accepted 20 April 2012
ies. The first part presents a model for a high-energy-density lithium-ion cell dedicated to EV applica-
Available online 4 June 2012
tions, based on Electrochemical Impedance Spectroscopy (EIS) measurements. The key characteristic of
this model, based on an equivalent-circuit approach, is not only its simplicity, but also the fact it takes
Keywords:
into account several important phenomena that occur inside lithium cells, such as the dependence of part
Lithium-ion battery
Modeling
of the internal resistance and the open-circuit voltage on the state of charge (SOC). The second part
Electrochemical Impedance Spectroscopy describes state-of-health (SOH) monitoring of a high-power-density lithium-ion cell, using recurrent
Aging neural networks (RNNs) to predict the deterioration in battery performance. This comprehensive
Recurrent neural network approach was used to monitor several batteries dedicated to HEV and EV applications, covering the entire
process, from behavior modeling to predicting performance degradation and use.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction concerning their aging modes and life-expectancy, which remains

a crucial problem.
Electric energy has taken on increasing importance with the There is obviously an acute need for tools capable of simulating
ongoing development of hybrid electric vehicles (HEVs) and, more the behavior of ESS, as well as monitoring deterioration in their
recently, electric vehicles (EVs), for a large number of applications: performance, based on current understanding of the phenomena
involved. These tools require accurate ESS models, tested on realis-
– increasing demand for both electric equipment (safety, comfort, tic operating profiles. One challenge for these ESS models is to
assistance, communication) and motors in light vehicles, implement aging laws identified from experimental campaigns.
– widespread electrification of public transport (trolleybuses) The first objective of this research was to model the behavior of
requires new solutions, based on partial autonomy or power- a high-energy-density lithium-ion cell, on the basis of Electro-
demand smoothing. chemical Impedance Spectroscopy (EIS) measurements. The first
part of this section presents lithium battery modeling research,
All these technological innovations rely on a common compo- based on equivalent circuits. The next part presents our approach,
nent, an Energy Storage System (ESS), designed to meet the using equivalent circuit modeling based on spectroscopy measure-
requirements of each application and enhance on-board energy ments to simulate cell behavior. This section concludes with simu-
management (auxiliaries, regenerative braking energy, etc.) [1]. lations and results validating the model.
However, current systems present several drawbacks. In view of the variety of high-performance energy storage appli-
A simulation approach was required to size and optimize their cations for electric and hybrid electric vehicles, the second sections
complex architectures [2]. This global model, covering the entire focuses on accelerated aging tests applied to high-power-density
system, from ESS electrochemistry to vehicle systems and use, Li-ion cells, particularly examining cell performance degradation
was unusually complex and extremely difficult to construct [3]. during power cycling. The results, based on periodic characteriza-
Changes in the demands placed on ESS and the use of innovative tion tests, reveal the changes in both cell capacity and equivalent
components and architectures raise the level of uncertainty series resistance (ESR) over a number of cycles. Capitalizing on
the power-cycling aging experiment results, the final section
describes a state-of-health indicator based on recurrent neural net-
⇑ Corresponding author. Tel.: +33 5 40 00 26 13; fax: +33 5 56 37 15 45.
works, designed for accurate monitoring of battery performance
E-mail address: akram.eddahech@ims-bordeaux.fr (A. Eddahech).

0142-0615/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijepes.2012.04.050
488 A. Eddahech et al. / Electrical Power and Energy Systems 42 (2012) 487–494

fading. The results of predictor simulations are compared with occur at high current levels, as in real use. However, the acquisition
experimental values to validate their consistency. frequency may be limited by the equipment, making it impossible
to measure impedance correctly at high frequencies. In contrast,
EIS provides accurate impedance measurements over a broad
2. Li-Po modeling from EIS
range of frequencies but only for low current values [8]. It is con-
sidered a rapid, reliable indicator of the battery’s SOC and state
2.1. Related work
of health (SOH) [9].
We monitored the cell model parameters using EIS measure-
Several equivalent-circuit models simulating lithium battery
ments. Fig. 1 shows the model used to simulate lithium cell
behavior have already been presented in the literature. Depending
behavior.
on the application and level of accuracy required, these models
The main elements in this model are a voltage source (E), which
present some simplifications and do not always consider specific
varies with SOC, an ohmic resistance (R1), a nonlinearly SOC-
phenomena that occur inside the cell, such as the nonlinearity of
dependent resistance (R2) connected in parallel with a constant
some parameters. In fact, for many applications, the saving in com-
phase element (CPE1) and a second constant phase element (CPE2).
puting time may be more important than the loss of accuracy [4].
The voltage source represents the open-circuit voltage (OCV),
Liaw et al. [5] presented an equivalent-circuit model simulating
corresponding to the voltage of the cell at its thermo-dynamic
the charge and discharge behavior of lithium-ion batteries. This
equilibrium, and the ohmic resistance (R1) represents the resis-
simple model separates the ohmic component represented by the
tance of the materials. The parallel connection of the second resis-
resistor (R1) and the faradic nonlinear part, represented by the
tance (R2) and (CPE1) describes the time-dependent overvoltage
(R2C) parallel branch.
occurring in the cell during charging or discharging, i.e.: the charge
The work presented in [6] describes a dynamic model of a Li-ion
transfer and double-layer processes. A second constant-phase ele-
cell that takes nonlinear phenomena into account. The model is
ment was used to model diffusion phenomena at low frequencies.
composed of an internal resistance (R1, R2), an effective capacitance
Actually, the resistance and capacitance values depend on the
(C) that characterizes the transient response of the double layer in
particular operating point of the cell, so they are time-variant
the porous electrodes, and, finally, a nonlinear equilibrium poten-
and change with temperature, SOC, power demand, and degrada-
tial (E), which depends on temperature, time, and current rate. The
tion [10]. This research focused mainly on SOC dependence.
authors confirmed that battery component losses were dependent
on chemistry, design parameters (geometry, pore structure, elec-
2.2.2. Experiments
trode materials, and electrolyte), temperature, and discharge rate.
The experiments were conducted using a cobalt manganese
Basically, equivalent-circuit models aim to represent all the sta-
nickel oxide lithium polymer cell (see Fig. 2). This cell had a carbon
tic (ohmic behavior) and dynamic (charge transfer and double-
negative electrode and a positive electrode consisting of lithium
layer processes) phenomena that occur inside lithium batteries.
cobalt manganese nickel oxide (LiMnNiCoO2). The lithium hexa-
They also introduce nonlinearity and the varying dependence of
fluorophosphate (LiPF6) electrolyte was a solution in a mixture of
the model parameters on state of charge (SOC), temperature, cur-
organic solvents: ethylene carbonate (EC) + ethymethyl carbonate
rent rate, and aging. Thus, several elements may be introduced to
(EMC).
monitor battery behavior more accurately, for example: simple
The overall chemical reaction inside the cell is given by:
resistance and capacitance or a constant-phase-element (CPE),
which represents a depression factor, valid between zero and Lix C þ Li1x MnNiCoO2 $ C þ LiMnNiCoO2 ð1Þ
one, responsible for the depression of the semi-circle in a Nyquist
Every battery has an internal resistance, Ri, due to the limited
plot of the EIS [4]. Diffusion at low frequencies is approximated by
conductance of the contacts, inter-cell connections, electrodes,
the Warburg element. Finally, an inductance L may be introduced
and electrolyte [12]. Analysis of variations in internal resistance
to model the metallic connection between battery terminals and
(Ri) with SOC at a precise age of the cell revealed a reverse corre-
electrodes.
lation: Ri decreased when SOC increased. In other words, it in-
Andre et al. [7] built and parameterized two different EC models
creased with the depth of discharge.
for a commercial 6.5 Ah high-power lithium-ion cell. The first was
Experimental Ri values are obtained by Hajizadeh and Golkar
based on three RC-elements and the second was a more complex
[2]:
model, based on CPE and Warburg elements. Both of the EC param-
eters were estimated by fitting a large number of EIS measure- DU
Ri ¼ ð2Þ
ments. The authors confirmed that the first EC was unable to DI
reproduce the measured data, even with optimized parameters.
In contrast, the more advanced EC reproduced the spectroscopy - DU represents the voltage drop after charging the cell for
measurements very accurately. Dt = 10 s,
This paper presents a simple model that takes into account non- - DI represents the equivalent charging current.
linear phenomena occurring inside the cell and uses complex ele-
ments, such as CPE. For simplification purposes, this model did not consider depen-
We then describe our equivalent-circuit modeling approach for dence on temperature or aging, but nonlinearity with SOC for both
simulating high-energy-density battery behavior, based on spec-
troscopy measurements.

2.2. Modeling

2.2.1. Equivalent-circuit model

The equivalent-circuit approach is based on experimental
results representing the behavior of a lithium cell in response to
various techniques, such as chrono-potentiometry and EIS. The
advantage of chrono-potentiometry is that charge and discharge Fig. 1. Proposed equivalent-circuit model of a lithium-ion cell.
 A. Eddahech et al. / Electrical Power and Energy Systems 42 (2012) 487–494 489

0.63

0.62

0.61

R2(mΩ)
0.6

0.59

0.58

0.57

20 30 40 50 60 70 80 90 100
Fig. 2. 100 Ah lithium-ion cell [11]. SOC (%)

Fig. 4. Changes in the estimated parameter R2 with SOC.

charge-transfer resistance and OCV were taken into account. The

inductive effect at high frequencies was also ignored.
cles to represent the charge-transfer and double-layer processes
The final model consisted of five elements: E (SOC), R1, R2 (SOC),
[5].
(CPE1), and (CPE2).
An automatic approach was used to identify the parameters on
The first parameter necessary for modeling cell behavior is the
the basis of an optimization algorithm, using EIS measurements.
OCV, approximated by a lookup table and taking SOC dependency
This technique had previously given accurate results in defining a
into account. Due to various dynamics characteristic of battery
supercapacitor model [15].
behavior, it is only possible to obtain an accurate OCV measure-
Static resistance, R1, remained constant, while Ri was SOC-
ment after a long rest period (often several hours) [13].
dependent. Therefore, R2 depended nonlinearly on SOC and was
In our case, EIS tests were conducted in galvanostatic mode at
obtained from a lookup table (Fig. 4). Finally, we assumed that
different SOCs and at 25 °C. Complex impedance spectra were col-
parameters of both of CPE elements were SOC-independent.
lected in the 10 kHz to 0.01 Hz frequency range. The impact of SOC
Fig. 5 presents a comparison between the Nyquist diagram sim-
on cell impedance is illustrated in Fig. 3.
ulated by the model and measurements on a real cell. This revealed
The impact of SOC on cell impedance is mainly observed in the
that our model reproduced cell behavior correctly. In addition, we
low- and mid-frequency range, corresponding to electrochemical
validated the model with all SOC presented in Fig. 3. Here, we pres-
kinetics such as mass transport processes, including the diffusion
ent validation for 55% of SOC.
and migration of electro-active species. We point out that a syn-
As inductive behavior was not taken into account, the model
thesis of the Nyquist plot is significant for validation, since exper-
plot was limited to the [10 mHz; 372 Hz] frequency window.
imental data may be obtained from tests [14].

2.3. Time-domain model validation

2.2.3. Parameters determination
A comparison between the lithium cell’s response to a current
R1 and Ri (equal to R1 + R2) may be extracted from the Nyquist
profile similar to that used in EV and that of the model confirmed
plot and R2 is calculated from the diagrams, after plotting semicir-
the accuracy of the model. Fig. 6 shows the current profile applied
to the cell.
-4
x 10

-4
x 10
3
2 Exp
Model

1
2
-Im[Z](Ω)

-Im[Z](Ω)

-1 1
25%
40%
55%
-2 70%
85%
100%
0
7 8 9 10 11
-4
6 7 8 9 10 11 12 Re[Z](Ω) x 10
-4
Re[Z](Ω) x 10
Fig. 5. Comparison between simulated model and real measurements on a Nyquist
Fig. 3. Nyquist diagram of cell impedance for several SOC at 25 °C. diagram at SOC 55%.
490 A. Eddahech et al. / Electrical Power and Energy Systems 42 (2012) 487–494

150 3.95
I(A) Model

100 3.9 Exp

3.85
50

3.8
Current (A)

Voltage (V)
0

3.75
-50

3.7
-100
3.65
-150
3.6

-200
0 50 100 150 200 3.55
Time (s) 0 50 100 150 200
Time (s)
Fig. 6. Current profile used for validation.
Fig. 8. Second validation of the model.

The current profile consisted of many charge/discharge pulses, under an accurate use conditions, namely battery temperature and
at different current levels. This profile was designed to simulate age. Moreover, for practical reasons, it is preferable to use a simple
a typical driving pattern, the ECE 15 urban driving cycle, used on model that facilitates on-board energy management [17].
electric vehicles. However, battery performances fade with time: battery delivers
The ECE 15 driving cycle represents city driving conditions in a less range after a large number of charge-discharge cycles. In addi-
typical European city (e.g. Paris or Rome) [16], simulating a tion, battery life is also affected by its use profile [18], whether it is
1.013 km urban trip at an average speed of 18.7 km/h, with a dura- charged in minutes or hours, and temperature.
tion of 195 sec. It is characterized by low vehicle speed (max. Therefore, for comprehensive analysis, it is obviously necessary
50 km/h), low engine load, and low exhaust gas temperature. to monitor battery SOH and take aging and use conditions, such as
The voltage curve, presented in Fig. 7, evolves in the opposite temperature, into consideration. In view of these parameters, a
way to the current, as the positive current in Fig. 6 is considered a neural network predictor was developed to model Li-ion battery
discharge current and vice versa. Fig. 7 below shows a test of the aging behavior and performance degradation under various oper-
model’s performance and ability to simulate lithium-cell behavior. ating conditions.
Fig. 8 shows the model’s response to another current profile The aim of this work was to develop a comprehensive approach
opposite to that of the first test. for describing the battery behavior and monitoring battery perfor-
This confirmed the model’s accuracy, as no modification of the mance during aging. This approach, which has many applications,
model parameters was required. from analyzing individual elements to battery use monitoring, is
very important for predicting failure and ensuring that lithium-
2.4. Discussion ion batteries function over a reliable range, as required in EV and
HEV applications.
Experimental validation demonstrated that the equivalent- We mention that the use of intelligent techniques, such as fuzzy
circuit model, developed using EIS measurements, provided an logic and artificial neural networks, is taking more importance on re-
accurate representation of lithium-ion polymer cell behavior. This search into energy storage systems, especially lithium-ion batteries
modeling approach, based on EIS, is useful for monitoring batteries [19,20].

3. Li-ion life monitoring using recurrent neural networks

3.85
3.1. Accelerated aging tests

The experiments were performed on high-power cells, designed

3.8
for hybrid electric vehicle (HEV) and high power applications (see
Fig. 9). These cells consisted of LiNixCoyAlzO2 cathodes paired with
Voltage (V)

3.75
low-cost graphite anodes and proprietary electrolyte mixtures
based on alkyl carbonates and LiPF6 salt with a vinylene carbonate
3.7 additive.
The tests were designed to accelerate and accentuate the aging
3.65 phenomena that occur in lithium-ion cells. In this study, cell
temperature (T), the amplitude of SOC variations (DSOC = SOC-
3.6 Model max  SOCmin), current pulse (I) magnitude, and the charge
Exp throughput of a micro-cycle (CT in As) are considered the main fac-
tors of aging in power cycling tests [21–24].
3.55
0 50 100 150 200
Time (s)
3.1.1. Power cycling tests
Fig. 7. Comparison between real voltage and model response to the current profile The power cycling tests continued for several months with
at 25 °C. characterization protocols applied periodically every 6 weeks to
 A. Eddahech et al. / Electrical Power and Energy Systems 42 (2012) 487–494 491

4.6
7C
4.4 1C

4.2

R (mOhm)
3.8

3.6

3.4

3.2

Fig. 9. Lithium-ion battery with three 7 Ah cells.

3
0 1 2 3 4
5
Cycle number x 10
check cell SOH. Basically, the power cycling profile was based on a
micro-cycle of discharge and charge current pulses separated by Fig. 11. Changes in ESR with aging at 45 °C and SOC 80% at different current rates.

rest periods. The micro-cycle charge balance was adjusted in order

to produce an overall discharge or charge (macro-cycle) within a
mainly due to decomposition of the electrolyte, leading to a de-
predefined SOC interval. In these tests, the temperature was con-
crease in conductivity and, consequently, an increase in the cell’s
trolled and aging was monitored at different temperatures (T):
internal resistance [24]. Physicochemical modifications (conduc-
40 °C, 45 °C, and 50 °C, current pulse magnitudes (I): 65 A, 98 A,
tivity and thickness) in the SEI layer on the positive electrode dur-
and 130 A and SOC variations (DSOC): 20%, 30%, and 40%.
ing cycling have also been considered predominant factors in the
increase in cell impedance [25].
3.1.2. Periodic time-domain characterization
These results have made it possible to quantify the impact of
The power cells were characterized before power cycling
the first-order factors of aging. This method is also useful for com-
started and then periodically, after a given number of cycles. The
paring the performance of different cell types and technologies.
deterioration in cell performance was quantified in terms of capac-
Other characterization methods were also investigated in order
ity reduction and impedance increase. The voltage response to cur-
to identify aging laws that could be implemented in battery simu-
rent pulses with varying current rates was monitored at different
lation models.
SOC. Capacity was measured at 1C rate and variations over a num-
In the next part of this paper, capacity and ESR values predicted
ber of cycles are illustrated in Fig. 10.
using recurrent neural networks are used as an indicator of battery
Capacity loss had reached 18.7% at the end of the power cycling
SOH and the corresponding results are presented.
test. This degradation was mainly due to deterioration of the cath-
ode, anode, and electrolyte. Electrochemical analyses confirmed
3.2. RNN-based SOH indicator for lithium-ion batteries
that the growth of the solid electrolyte interface (SEI) due to side
reactions, resulting in an irreversible consumption of lithium, rep-
3.2.1. Approach
resented the most important factor in capacity reduction at the
By monitoring performance trends during power cycling tests,
negative electrode [23].
recurrent neural networks (RNNs) are capable of predicting varia-
The change in ESR with the number of cycles is shown in Fig. 11,
tions in capacity and ESR during battery life on the basis of previ-
depicting the same cell at different current rates.
ous measurements.
Note that ESR was higher at 1C than 7C. However, it rose by
First of all, in order to evaluate the remaining useful life (RUL) of
21.4% at 7C and only 10.1% at 1C after the same number of cycles.
the battery and its SOH, it is necessary to determine its end of life
From an electrochemical point of view, the increase in ESR was
criteria. Thus, we assume that the loss of 20% of initial capacity or a
100% increase in ESR indicates these criteria. Battery SOH and RUL
7 were then quantified, on the basis of the capacity and ESR values
C (Ah) predicted.

3.2.2. Structure of the RNN-based predictor

6.5
It is extremely difficult to calculate all the dependencies in-
volved in aging phenomena and correlate the influence of all the
Capacity (Ah)

factors impacting battery behavior, mainly complexities in design

6 and manufacture, as well as environmental and operational condi-
tions (temperature, battery technology, etc.). Nevertheless, artifi-
cial neural networks (ANNs) present a useful alternative for
forecasting nonlinear dynamics and complex behaviors [19–26].
5.5
ANN techniques are capable of solving complex problems where
traditional methods have failed. This model focuses on RNN-based
battery SOH prediction.
5
0 1 2 3 4
The prediction results are based on knowledge of some aspects
Cycle number x 10
5 of previous system behavior. Moreover, temperature (T), current
pulse magnitude (I), and SOC variations (DSOC) are used as
Fig. 10. Capacity variation with aging at 45 °C, DSOC = 20% and I = 98 A. additional network input. Indeed, experiments conducted under
492 A. Eddahech et al. / Electrical Power and Energy Systems 42 (2012) 487–494

C
ΔSOC(k)
Inputs
RNN1
I(k) Recurrent
R/C(k+1)
Neural Present value of
T(k) R/C(k+2) Capacity
Network capacity
R/C(k-2) R/C(k+3) predictor
Predictor
R/C(k-1)
R/C(k)
ESR State of life
Inputs
RNN2 indicator
Fig. 12. RNN for predicting lithium-ion battery performance.
Present value of
ESR predictor
resistance
various conditions revealed that operational conditions signifi-
cantly affected the capacity life of a battery, so the predictor takes Fig. 14. SOH evaluation process.

these parameters into account. RNNs, a special class of neural net-

where n is the number of recurrent inputs in the first layer of the
works capable of generating feedback information from their out-
network and h is its set of weights. These weights were corrected
puts have been proven to be computationally rich [27]. This work
at the end of each prediction window.
is based on a nonlinear autoregressive model with exogenous input
We developed two predictors for battery performance evalua-
[28] for predicting battery performance trends.
tion, with network outputs representing capacity and estimated
Fig. 12 shows the architecture of the RNN predictor.
resistance. Fig. 14 illustrates the lithium-ion battery SOH monitor-
The main reasons for using this recurrent structure are its sim-
ing process.
ple approach and easily-implemented training algorithm. In fact,
this structure represents a powerful calculating model. The learn-
ing algorithm is formulated to predict N steps ahead. For the 3.2.3. Simulation results
RNN shown in Fig. 13, the objective function was optimized, as In order to improve the predictor’s training capacity, many sim-
follows: ulation tests were conducted to identify the structure that gave the
best prediction results.
1X N
1X N
J¼ ½Y e ðk þ iÞ  Y r ðk þ iÞ2 ¼ Ai ð3Þ The network transfer function is a sigmoid function, expressed
2 i¼1 2 i¼1
as:
where N is the prediction window, Ye (k + i) is the estimated output
1
by the RNN, and Yr (k + i) is the actual output. The network weights, f ðxÞ ¼ ð6Þ
1 þ ex
h, were updated using the following gradient descent rules:
The following are the appropriate parameters for the RNN pre-
@J
hnew ¼ hold  e dictor, including the number of hidden and input neurons and the
@h
prediction window:
@J 1 X @Ai @Y e ðk þ iÞ
¼ e ð4Þ - Number of input neurons: 7.
@h 2 i @Y ðk þ iÞ @h
- Number of hidden neurons: 6.
- Prediction window: N = 3.
@J X N
- Transfer function: the sigmoid function.
¼ ½Y e ðk þ iÞ  Y r ðk þ iÞ:
@h i¼1 - Training algorithm: back propagation through time [29].
" #
X n
@Y e ðk þ iÞ @Y e ðk þ i  jÞ @Y e ðk þ iÞ
þ ð5Þ Figs. 15 and 16 below present a comparison between the esti-
j¼1
@Y e ðk þ i  jÞ @h @h mated and the real values of capacity and ESR during power cycling
tests.

1
Predicted
Experimental

0.95
Normalized capacity

0.9

0.85

0.8

0.75
0 100 200 300 400 500
K cycles

Fig. 13. Recurrent neural network model. Fig. 15. Predicted and measured capacity variations during cycling.
 A. Eddahech et al. / Electrical Power and Energy Systems 42 (2012) 487–494 493

2.3 capacity loss and an increase in equivalent series resistance (ESR)

over a number of cycles. The battery SOH indicator developed using
2.2
RNN provided an accurate simulation of lithium-ion cell aging
2.1 behavior, taking specific operational conditions into account and
providing practical information on the expected life of the battery.
2
RNN-based predictors are applicable to a variety of high-perfor-
Normalized ESR

1.9 mance energy storage systems for hybrid and electric vehicles. These
encouraging simulation results represent an initial contribution to
1.8
the development of a real-time automated system for monitoring
1.7 battery behavior and SOH.

1.6 Predicted References

Experimental
1.5
[1] Du W, Wang HF, Cheng S, Wen JY, Dunn R. Robustness of damping control
implemented by energy storage systems installed in power systems. Int J Electr
1.4
Power Energy Syst 2011;33:35–42.
[2] Hajizadeh A, Golkar MA. Control of hybrid fuel cell/energy storage distributed
1.3 generation system against voltage sag. Int J Electr Power Energy Syst
0 100 200 300 400 500
2010;32:488–97.
K cycles [3] Roscher MA, Leidholdt W, Trepte J. High efficiency energy management in BEV
applications. Int J Electr Power Energy Syst 2012;37:126–30.
Fig. 16. Predicted and measured ESR variations during cycling. [4] Buller S, Thele M, Karden E, De Doncker RW. Impedance based nonlinear
dynamic battery modeling for automotive applications. J Power Sources
2003;113:422–30.
[5] Liaw BY, Nagasubramanian G, Jungst RG, Doughty DH. Modeling of lithium ion
Table 1 cells-a simple equivalent-circuit model approach. Solid State Ionics
2004;175:835–9.
MSE error for both capacity and resistance
[6] Gao L, Liu S. Dynamic lithium-ion battery model for system simulation. IEEE
prediction.
Trans Compon Pack Technol 2002;25(3):495–505.
Performance predictor MSE [7] Andre D, Meiler M, Steiner K, Walz H, Soczka-Guth T, Sauer DU.
Characterization of high power lithium ion batteries by electrochemical
Capacity 0.462 impedance spectroscopy. II: Modeling. J Power Sources 2011;196:5349–56.
Resistance 0.296 [8] Zhuang QC, Wei T, Du LL, Cui YL, Fang L, Sun SG. An electrochemical impedance
spectroscopic study of the electronic and ionic transport properties of spinel
LiMn2O4. J Phys Chem 2010;114:8614–21.
[9] Mandal LP, Cox RW. A transient-based approach to estimation of the electrical
These results revealed that the neural predictors were able to parameters of a lead-acid battery model. In: Proc IEEE energy conversion
congress and exposition conf; 2010. p. 4238–42.
accurately predict variations in performance with aging. This pre- [10] Remmlinger J, Buchholz M, Meiler M, Bernreuter P, Dietmayer K. State-of-
dictor acts as a life indicator, giving information about fading rate. health monitoring of lithium-ion batteries in electric vehicles by on-board
The mean square error (MSE) on predicting performance pre- internal resistance estimation. J Power Sources 2011;196:5357–63.
[11] Eddahech A, Briat O, Chaari R, Bertrand N, Henry H, Vinassa J-M. Lithium-ion
sents the learning and generalization error of the normalized val- cell modeling from impedance spectroscopy for EV applications. In: Proc IEEE
ues of capacity and resistance (see Table 1). It is computed using energy conversion congress and exposition conf, Phoenix, Arizona; 2011.
Eq. (7), where K represents the total number of samples (K = 500). [12] Moss PL, Au G, Plichta EJ, Zheng JP. An electrical circuit for modeling the
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi dynamic response of Li-ion polymer batteries. J Electrochem Soc
Xn e 2008;155(12):A986–94.
MSE ¼ 1
½Y ðkÞ  Y r ðkÞ2 ð7Þ [13] Hu Y, Yurkovich S, Guezennec Y, Yurkovich BJ. Electro-thermal battery model
identification for automotive applications. J Power Sources 2011;196:449–57.
where Yr (k) is the real value for sample k and Ye (k) the correspond- [14] Verbrugge MW, Conell RS. Electrochemical and thermal characterization of
ing predicted value. battery modules commensurate with electric vehicle integration. J
Electrochem Soc 2002;149:45–53.
The peaks shown in these figures represent the performance [15] El Brouji E-H, Briat O, Vinassa J-M, Bertrand N, Woirgard E. Impact of calendar
recovery phenomenon, which occurred whenever cycling stopped. life and cycling aging on supercapacitor performance. IEEE Trans Veh Technol
This may be related to charge redistribution when there is no 2009;8:3917–29.
[16] Tzirakis E, Pitsas K, Zannikos F, Stournas S. Vehicle emissions and driving
charge or discharge force acting on them. cycles: comparison of the Athens driving cycle (ADC) with ECE-15 and
Predicting the real battery capacity values makes it possible to European driving cycle (EDC). Global NEST J 2006;8(3):282–90.
avoid overcharging. In practical terms, installing a battery health [17] Zhang J, Xia C. State-of-charge estimation of valve regulated lead acid battery
based on multi-state unscented Kalman filter. Int J Electr Power Energy Syst
indicator in an HEV or EV energy management system helps to
2011;33:472–6.
avoid energy shortages, system malfunctions, and failures, by mon- [18] Dursun E, Kilic O. Comparative evaluation of different power management
itoring battery SOH, providing advance warning that repairs or strategies of a stand-alone PV/Wind/PEMFC hybrid power system. Int J Electr
maintenance are required, and ensuring that the system fulfills Power Energy Syst 2012;34:81–9.
[19] Weigert T, Tian Q, Lian K. State-of-charge prediction of batteries and battery–
its energy requirement during use. supercapacitor hybrids using artificial neural networks. J Power Sources
2011;196:4061–6.
[20] Singh P, Vinjamuri R, Wang X, Reisner D. Design and implementation of a fuzzy
4. Conclusions logic-based state-of-charge meter for Li-ion batteries used in portable
defibrillators. J Power Sources 2006;162:829–36.
This paper presents a lithium-ion polymer cell model based on [21] Smart MC, Ratnakumar BV, Whitcanack LD, Puglia FJ, Santee S, Gitzendanner R.
Life verification of large capacity Yardney Li-ion cells and batteries in support
the equivalent-circuit approach, using EIS measurements. This of NASA missions. Int J Energy Res 2010;34:116–32.
model provides an accurate description of cell behavior, taking into [22] Wohlfahrt-Mehrens M, Vogler C, Garche J. Aging mechanisms of lithium
account several key phenomena that occur inside lithium cells, cathode materials. J Power Sources 2004;127:58–64.
[23] Broussely M, Biensan Ph, Bonhomme F, Blanchard Ph, Herreyre S, Nechev K,
such as the dependence of internal resistance and open circuit volt-
et al. Main aging mechanisms in Li ion batteries. J Power Sources
age on SOC. Validation with a real profile used in electric vehicles 2005;146:90–6.
confirmed the accuracy of the model. [24] Troltzsch U, Kanoun O, Trankler H-R. Characterizing aging effects of lithium
The next stage focused on accelerated aging tests applied to high- ion batteries by impedance spectroscopy. Electrochim Acta 2006;51:1664–72.
[25] Eddahech A, Briat O, Henry H, Delétage J-Y, Woirgard E, Vinassa J-M. Aging
power-density lithium-ion cells. Periodic characterization tests re- monitoring of lithium-ion cell during power cycling tests. Microelectr Reliab J
vealed the impact of aging factors on power cycling, resulting in 2011;51(9–11):1968–71.
494 A. Eddahech et al. / Electrical Power and Energy Systems 42 (2012) 487–494

[26] Eddahech A, Briat O, Vinassa J-M. Neural networks based model and voltage [28] Basso M, Giarré L, Groppi S, Zappa G. NARX models of an industrial power
control for lithium polymer batteries. In: 8th IEEE international symposium on plant gas Turbine. IEEE Trans Control Syst Technol 2005;13(4):599–604.
diagnostics for electric machines. Power Electronics and Drives, Bologna; [29] Mazumdar J, Harley RG. Recurrent neural networks trained with
2011. backpropagation through time algorithm to estimate nonlinear load
[27] Bonanno F, Capizzi G, Tina G. Long-term energy performance forecasting of harmonic currents. IEEE Trans Ind Electr 2008;55(9):3484–91.
integrated generation systems by recurrent neural networks. In: Proc IEEE
clean elecrtical power conf, Italie; 2009. p. 673–8.