Journal of Power Sources 274 (2015) 355e366

Contents lists available at ScienceDirect

Journal of Power Sources

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

Fuzzy modelling for the state-of-charge estimation of lead-acid

batteries

ez*, Marcos E. Orchard, Roberto Ca
Claudio Burgos, Doris Sa
rdenas
Department of Electrical Engineering, Faculty of Mathematical and Physical Sciences, University of Chile (DIE), Av. Tupper 2007, 8370451 Santiago, Chile

h i g h l i g h t s

A novel fuzzy model is proposed to characterise lead-acid battery discharging.

An extended Kalman filter based on a fuzzy model is proposed for SoC estimation.

The polarisation resistance can be estimated using this novel algorithm.

Experimental validation of the proposed methodology is presented and discussed.

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

Article history: This paper introduces a novel fuzzy model based structure for the characterisation of discharge processes
Received 6 July 2014 in lead-acid batteries. This structure is based on a fuzzy model that characterises the relationship be-
Received in revised form tween the battery open-circuit voltage (Voc), the state of charge (SoC), and the discharge current. The
3 October 2014
model is identified and validated using experimental data that is obtained from an experimental system
Accepted 6 October 2014
Available online 15 October 2014
designed to test battery banks with several charge/discharge profiles. For model identification purposes,
two standard experimental tests are implemented; one of these tests is used to identify the VoceSoC
curve, while the other helps to identify additional parameters of the model. The estimation of SoC is
Keywords:
State-of-charge
performed using an Extended Kalman Filter (EKF) with a state transition equation that is based on the
Lead-acid batteries proposed fuzzy model. Performance of the proposed estimation framework is compared with other
Fuzzy modelling parametric approaches that are inspired on electrical equivalents; e.g., Thevenin, Plett, and Copetti.
Extended Kalman filter © 2014 Elsevier B.V. All rights reserved.

1. Introduction The SoC may be considered as the remnant energy in a battery

bank. The SoC is usually defined as a value in the interval [0,1],
The estimation of the remnant energy stored in Battery Energy where 0 indicates complete discharge of the battery and 1, full
Storage Systems (BESS) is important for the evaluation of its future charge. This work focuses on the problem of estimating the SoC in
performance. In this regard, many different battery-based appli- lead-acid Batteries; a variable that is affected by charge/discharge
cations have considered the problem of State-of-Charge (SoC) rates, temperature, usage time, hysteresis and self-discharge effects
estimation; e.g., microgrids [1], electric cars [2], wind-diesel sys- (due to the internal impedance of the cell). There are a number of
tems, distribution grids [3], and Flexible Alternate Current Trans- aspects that have to be considered when implementing SoC esti-
mission Systems (FACTS). Moreover, in microgrids, the SoC is mation procedures, and in this regard, we will solely focus on: (i)
usually an input of the energy management system [1] and it is the model of the battery that will be used, and (ii) the algorithm
required to optimise its performance. Additionally, the SoC is also that will be utilised to estimate parameters in non-linear/non-
an important parameter for an adequate operation of electric cars observable systems. Several research efforts aim at providing a
and drones, because it provides information that helps to estimate solution for these issues using empirical, physicochemical, or
their respective autonomy. electric models in conjunction with estimation techniques based on
fuzzy logic [4], neural networks [5], or Bayesian approaches such as
the Extended Kalman Filter (EKF) [6,7] and particle filtering [8].
One of the main difficulties for SoC estimation is that the SoC
* Corresponding author. cannot be directly measured in a battery-based storage system.
E-mail address: [email protected] (D. S

aez). Therefore, this value must be inferred from the observation of other

http://dx.doi.org/10.1016/j.jpowsour.2014.10.036
0378-7753/© 2014 Elsevier B.V. All rights reserved.
356 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366

variables such as the battery current, voltage, temperature, State- is augmented with a fuzzy inference rule which considers both the
of-Health (SoH) [8], and indicators associated with the self- temperature and the battery current level. However, the proposed
discharge phenomena [9e12]. The utilisation of more complex estimator is only validated by simulation experiments and it is not
electrochemical models for the estimation of the SoC is also tested with real data from an experimental system.
possible. However these techniques are only suitable for off-line Neural networks have also been used to represent a non-linear
studies, mainly because these models (i) require a large number relationship between battery measurements respect to the evo-
of variables to represent the battery internal processes, (ii) assume lution of SoC in time; however, they usually require huge data sets
extremely accurate measurements [5,9], and (iii) have an elevated for the identification process [9,5,11]. A neuro-fuzzy system is
computational cost [5,9]. Other options for SoC monitoring include proposed for estimation of the SoC of a Li-Ion battery in Ref. [21].
the open-circuit voltage (Voc) method [13]; where an equivalent The data is collected using Hybrid Pulse Power Characterisation
circuit model is developed using resistors, capacitors and voltage (HPPC) tests specified in the standard “Partnership for a new
sources to form a circuit network (such as polarisation-impedance- Generation of Vehicles” [22]. In Ref. [21] an adaptive neuro-fuzzy
based models or Thevenin-equivalent-based models). This inference system (ANFIS) models the relation between SoC and
approach has the advantage of providing a direct relationship be- the open circuit voltage. Given that the estimator input corre-
tween battery SoC and voltage measurements e the higher the Voc, sponds to the Voc, this method is not suitable for on-line appli-
the higher the SoC [14]. Unfortunately, the implementation of this cation. A slightly different approach was presented in Refs. [23],
test requires long battery resting periods, and limiting its use for where ANFIS models and Principal Components Analysis (PCA) are
online applications [5,9,12,14]. Similarly, Electrochemical Imped- used to estimate the SoC of Ni-MH batteries. In this method, the
ance Spectroscopy (EIS) [9,15] is outlined as a non-invasive method main inputs of the ANFIS estimator are the discharge current,
that intends to provide a complete characterisation of the battery battery terminal voltage and the battery energy released. Ac-
internal equivalent circuit. However, the implementation of an EIS cording with Ref. [23] the proposed neuro-fuzzy model has several
test requires the acquisition of costly equipment (usually available advantages when compared with conventional neural network
only at laboratory test sites), which severely limits its wide spread estimator. However, huge data sets are required for training the
use in practice [16]. Moreover, data obtained using EIS, usually is large number of parameters required by the proposed ANFIS sys-
very noisy [6,16]. It is for this reason that current research efforts tem. This may produce a substantial increase in the requirements
are mostly focused on the development of SoC estimation algo- associated with the implementation.
rithms based on empirical models that incorporate relevant A new learning structure called Merged Fuzzy Neural Network
phenomenological aspects of the process; i.e., the relationship be- (MFNN) is proposed in Ref. [24] to estimate the SoC of a lithium-
tween currents, voltages and temperatures of cells. Among these ion battery bank. The SoC of each battery cell is modelled by a
methods, it is worth mentioning those that are based on fuzzy logic, neural network using the battery terminal voltage, discharge
neural networks, and Bayesian approaches. current and battery surface temperature as inputs. The parameters
Fuzzy logic models have been used for SoC estimation mainly for of the MFNN are obtained using genetic algorithms. Although this
the identification of battery from EIS data [4,17,18] and to a lesser new approach allows the modelling of the whole battery bank, a
degree, directly from voltage or discharge current measurements. very large data set is required because measurements in each cell
Salkind et al. in Ref. [4] propose a SoC estimator based on fuzzy have to be realised. Therefore, the complexity of this model is high,
models for two battery storage systems: lithium/sulphur dioxide increasing the difficulties associated with a real time
and nickel/metal hydride. In this estimator, the relationships be- implementation.
tween the imaginary component of the impedance at 10.3 Hz, In recent years there has been a growing interest in the use of
41.01 Hz and 4101 Hz and the SoC of the cell are modelled using a stochastic filtering techniques (unscented Kalman Filter [25],
Sugeno fuzzy system. A similar approach for the SoC estimation of a extended Kalman Filter [6,7], and unscented particle filter [26]) to
Ni/MH battery is presented in Ref. [17]. This paper also considers estimate the SoC of battery cells. In fact, experience has demon-
the relationship between the imaginary component of the imped- strated that Bayesian estimators are well suited for real-time esti-
ance and the SoC using fuzzy clustering to obtain of the corre- mation problems that incorporate dynamic state transition models
sponding fuzzy sets. Zenati et al. in Ref. [18] propose to use EIS [6,8,16]. Particularly related to the proposed problem is a Stochastic
measurements for the SoC estimation of Li-Ion batteries. The Fuzzy Neural Network (SFNN) presented in Ref. [27] to model the
experimental work discussed in this paper considers temperature Ni-MH battery behaviour. The SoC is estimated with extended
variations and several battery charging levels. The influence of DC Kalman filters based on SFNN models. The load voltage is modelled
current over the AC impedance during EIS measurement is ana- by the SFNN considering the charge and discharge currents, tem-
lysed and interpreted by a fuzzy system. The methods proposed in perature and SoC as inputs. The SFNN has a filtering effect on the
Refs. [4,17,18] could be used for on-line estimation of the SoC. inputs and this improves the signal to noise relationship of the
However, it requires access to electrochemical impedance spec- output signals. In this case the validation discharge profile is very
troscopy (EIS) data. Moreover, because of the EIS disadvantages similar to the training set. Therefore, the methodology proposed in
previously discussed, methods to characterise the SoC using voltage Ref. [27] has not been fully validated for different battery operating
and current measurements have more advantages. In this regard, a conditions.
different approach that generates a SoC estimator from voltage Particularly in this work, a novel estimator of the SoC for a lead-
recovery profiles, without the need of impedance measurements, is acid battery bank is proposed. The SoC estimator is based on an
presented in Ref. [19]. Particularly, this method uses the minimum Extended Kalman Filter (EKF) that uses a fuzzy model as its
battery voltage, at a fixed load current, as one of the fuzzy system observation equation. The scheme of the proposed estimator is
inputs. The other input is the difference between the maximum and shown in Fig. 1. The EKF uses the measurements of both voltage,
minimum voltage of the battery pack. This methodology is pro- V(k) and current, I(k). The current measurements are used to esti-
posed in Ref. [19] to estimate the SoC in batteries that undergo mate the polarisation resistance Rint by using a fuzzy model and this
constant charge/discharge profiles as for instance in batteries used is one of the inputs of the state estimator for generating the esti-
to feed portable defibrillators. mated voltage. The voltage modelling error is used for updating the
Another fuzzy logic based SoC estimator for lead-acid batteries SoC estimator. The proposed EKF based on a fuzzy model will be
is presented in Ref. [20]. In this case, an improved Coulomb metric described on detailed in Sections 3 and 4.
 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366 357

In particular, the Voc-SoC curve in lead-acid batteries can be

modelled using a linear-in-the-parameters structure [30], as shown
in Eq. (1).

Vock ¼ an SoCnk þ an1 SoCn1

k þ … þ a1 SoCk þ a0 ; (1)

where ai(i ¼ 1, …, n) are model parameters. In this model, the SoC is

empirically calculated as the integral of the instantaneous current.
The discrete model is given by:

Ts
SoCkþ1 ¼ SoCk  I ; (2)
Cn k

where Cn is the cell nominal capacity (which can be experimentally

obtained [22]), Ts is the sampling time, and Ik is the instantaneous
discharge current.

2.2. Thevenin circuital equivalent model

Fig. 1. Scheme of the proposed EKF based on a fuzzy model for SoC estimations.

The Thevenin circuital equivalent model consists of a controlled

voltage source (whose value depends on the SoC), a series RC branch,
Two major contributions are highlighted:
and a series resistance [31e35]. Using this topology and assuming a
positive current for battery discharging, a discrete model for the

A new fuzzy model for the output voltage of a lead-acid battery
output voltage (as a function of the current) is given by:
bank, which is inspired on the process phenomenology. The
polarisation resistance is modelled by a non-linear interpolation Ts
(fuzzy based) of a set of available curves obtained experimen- Vk ¼ Vock þ ½Vk1  Vock1 $eR0 $C0

 (3)
tally at different operating points of the battery bank. The re- Ts

quirements for experimental data of this model are similar to  R0  ðR0 þ Rint Þ$eR0 C0 $Ik1  Rint $Ik ;
that required for the identification of classical circuital models of
battery banks. where Vock is obtained from Eq. (1), Rint is the polarisation resis-

The proposed EKF is based on a fuzzy model which requires only tance of the battery, C0 is the capacitance between the electrodes,
one internal state in its formulation. This configuration allows and R0 is a non-linear element that represents the contact resis-
an efficient microprocessor implementation, with a reduced tance between the electrodes and electrolyte. For parameter
number of numerical calculations. identification purposes, the voltage relaxation test is used for
determining the curve SoCeVoc and a test shown in Fig. 2 is used in
The structure of the paper is as follows. Section 2 presents the order to obtain the model resistances and capacitor values.
lead-acid battery models reported in the literature and their rela- The Thevenin model can be improved by modifying the polar-
tionship with the problem of lead-acid battery SoC estimation. In isation resistance, according to the battery operating point; e.g.,
Section 3, the fuzzy model proposed by the authors is extensively assuming different Rint for charging and discharging [33].
discussed. The use of an Extended Kalman Filter, based on a fuzzy
model, for SoC estimation is presented in Section 4. In Section 5, the
design of an experimental system used for system modelling and 2.3. Plett model
validation is discussed. In Section 6, the experimental results are
shown. Finally, analysis, comments, and further research lines are Plett in Ref. [36] presents a model for a battery bank that de-
presented in the conclusions. pends explicitly on the SoC. In this structure, the output voltage is
given by:
2. Classic models for batteries

This work presents a comparison between the performance of

the proposed fuzzy model with respect to other well-known bat-
tery models (Thevenin, Plett, and Copetti). Thus, we present in this
section a brief analysis of all these models, as well as their main
features.

2.1. Open circuit voltage curve

There is a well-known relationship between the open-circuit

voltage (Voc) and the state-of-charge (SoC) [22,28,29]. Most of the
times, this relationship is obtained using an experimental test known
as “voltage relaxation”: a procedure that basically applies a known
discharge profile to the battery for a given period, then forces null
discharge current (open circuit) for an appropriate “rest” time (usu-
ally an hour), and then measures the (Voc) output voltage [28]. Fig. 2. Training set used for parameter identification purposes.
358 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366

K1 3.1. Fuzzy model based on rules

Vk ¼ K0  Rint Ik   K2 SoCk þ K3 ln½SoCk 
SoCk (4)
Fuzzy logic defines a fuzzy set A, whose elements belong to a
þ K4 ln½1  SoCk ;
discourse universe X as the set of tuples A ¼ [(x, mA (x))/x2X] where
mA(x) is denominated the membership function associated with
where K0 represents the polarisation effects, Rint is the constant
fuzzy set A. The membership function assigns a membership degree
polarisation resistance of the battery storage system and Ki(i ¼ 1, …,
between 0 and 1 to each element of X. The most commonly used
4) are the model coefficients. All the terms associated with Ki
membership functions are triangular, trapezoidal and Gaussian
represent the relationship between VoceSoC, and thus the relaxing
shape. Based on this definition, a Takagi & Sugeno fuzzy system
test is not required as these terms are directly obtained from the
corresponds to a set of rules of the form [40]:
experimental tests.
An EKF implementation based on a Plett model has been used in Rule j : If x is Aj then yj ¼ fj ðx; zÞ (7)
Ref. [37] for the estimation of the SoC in ion-lithium batteries. In
this case, only one state variable was included, which is the SoC. Pr
This estimator was tested for values higher than 40%. Notice that j¼1wj ðxÞ$yj ðx; zÞ
yðx; zÞ ¼ Pr ; (8)
the Plett model (4) could have numerical errors for SoC near to zero. j¼1 wj ðxÞ

where Aj are the fuzzy sets, fj is a non-linear function for the rule j,
2.4. Copetti model wj is the activation degree of rule j and r are the number of rules.

The Copetti model [38,39] considers an equivalent circuit for the

3.2. Fuzzy model for a battery bank
battery that contains a constant voltage source and a variable
polarisation resistance Rint(SoC, I, T). The voltage source represents
The fuzzy model proposed in this work is based on the circuit
the relationship between Voc and SoC, while Rint represents the
shown in Fig. 3. It is assumed that the battery output voltage de-
battery electrochemical effects that oppose to the current flow. The
pends on both, the SoC, and the battery current. Temperature var-
Copetti model describes four operation modes for the battery sys-
iations are not considered in the model, because the battery bank is
tem; a charging zone, a discharging zone, an overcharging zone and
located in an environment with controlled temperature.
an intermediate zone representing a soft transition between
In Fig. 3, the polarisation resistance is modelled as a non-linear
charging/discharging modes. In this work only the discharging
function dependant on both the SoC and the discharge current. It is
zone is studied; therefore, the analytical expression for the output
important to note that this equivalent circuit does not consider
voltage is given by:
capacitances. In this regard, it is similar to the equivalent circuits
Vk ¼ ½V0d  K0d ð1  SoCk Þ obtained from Plett and Copetti models. Therefore they are used to
" # describe the steady state behaviour of the battery bank.
Ik P1d P3d (5) Based on Fig. 3, the fuzzy model discrete equation is given by:
 þ þ P5d ½1  qd DT;
C10 1 þ I P2d SoCP4d
k k
b ðSoC ; I Þ;
Vk ¼ Vock  Ik $ R (9)
int k k
where Pid(i ¼ 1, …, 5) are unknown coefficients associated to the
polarisation resistance. C10 corresponds to the battery capacity (in where the voltage source Vock is obtained from Eq. (1). Therefore,
AH) for 10 h [39] and qd relates the model with the temperature the values of the proposed polarisation resistance can be obtained
variation (DT).The term [V0d  K0d(1  SoCk)] models the rela- using Eq. (9), i.e.:
tionship VoceSoC. If temperature effects are neglected, the Copetti
model equation is: b ðSoC ; I Þ ¼ Vock  Vk :
R (10)
int k k
Ik
" #
Ik P1d P To obtain the polarisation resistance of the proposed model (see
Vk ¼ Vock  þ 3d þ P5d : (6)
C10 1 þ I P2d SoCP4d Fig. 3), experimental tests at fixed discharge current levels are
k k
performed (see Fig. 2). Fig. 4 shows the computed values of re-
Also, in this model, the curve VoceSoC is defined using Eq. (1). sistances in function of SoC, using Eq. (10) and current values of
Parameters associated with the battery polarisation resistance are 10 A, 15 A, 25 A, 32 A, together with the corresponding output
identified using data from the training set shown in Fig. 2. voltage measurements Vk. To calculate the equivalent polarisation
As well as the Plett model, the Copetti model (6) also presents resistance for any value of discharge current within the operating
discontinuities when the SoC is near to zero. range, the proposed fuzzy model interpolates between the resis-
tance curves depicted in Fig. 4. Thus, the structure for the proposed
fuzzy model and its rules for the calculation of R b ðSoC ; I Þ are:
int k k
3. Fuzzy modelling of lead-acid batteries

The classical models discussed above, have disadvantages (e.g.,

polarisation resistance is assumed to be constant, numerical errors,
discontinuities in the vicinity of SoC near to zero), which could limit
their performance in a SoC estimation procedure.
In this section a novel battery model is proposed. It is based on
fuzzy system and has the following advantages: i) it offers a non-
linear characterization of the battery polarisation resistance, (ii) it
requires a limited number of experimental tests for model identi-
fication purposes, and (iii) it does not present discontinuities for
SoC values in the vicinity of zero. Fig. 3. Circuital representation of the fuzzy model.
 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366 359

formulation of the EKF considers a non-linear stochastic system

that can be represented by the following discrete time state space
equations [7]:

xk ¼ f ðxk1 ; uk1 Þ þ uk (13)

yk ¼ hðxk ; uk Þ þ vk ; (14)

where xkεRn is the unknown state vector, ukεRr are known process
inputs, and ykεRm are observation outputs. Process and measure-
ment noises are denoted by uk and vk (Gaussian), with variances Q
and R respectively; while f($) and h($) are non-linear functions that
characterise the state transition and measurement equations, in
that order.
Fig. 4. Polarisation resistance values at different current levels. The EKF has the same structure of the well-known linear Kal-
man filter [41]. Eqs. (13) and (14) are used to characterise the non-
linear state transition equation. These equations are linearised us-
Rule j : If Ik is AIj ;j b ing the Taylor expansion, and only selecting its first term. Thus, the
then R intj ¼ fj ðSoCk Þ; (11)
expectation of the estimate state probability density function (PDF)
Pr is given by:
j¼1 wj ðIk Þ$fj ðSoCk Þ
b ðSoC ; I Þ ¼
R Pr ; (12)
int k k
j¼1 wj ðIk Þ x k=k1 ¼ f ðb
b x k ; uk Þ þ uk (15)
where wj is the activation degree of rule j (in this case, equal to the
The Pk/k1 covariance matrix is obtained by calculating a deriv-
membership degree to the fuzzy set AIj ;j ) and fj is a non-linear
b ative respect to b
x k=k1 as:
function adjusted for the resistance values R intj in function of
SoCk at a fixed level of the current Ij (see curves from Fig. 4). The 
vf 
fuzzy sets are obtained from the experimental data (Fig. 4) and they Ak1 ¼ (16)
vxk1 bx k=k1 ;uk1
are shown in Fig. 5.
Gaussian membership functions for the fuzzy sets AIj ;j are pro-
posed in this work. Gaussian membership functions have been Pk=k1 ¼ Ak1 Pk1=k1 ATk1 þ Q : (17)
selected because they only require two parameters (mean value
and standard deviation) to be completely defined. In this particular Once the prediction step is completed, the following equations
case, the mean value of these functions is defined as Ij, and the are computed:
standard deviation is selected to minimise the model mean square  
error. ~k ¼ yk  h b
y x k=k1 ; uk (18)
Finally, it is important to mention that the fuzzy model imple-
mentation requires similar experimental data size than that
Sk=k1 ¼ Ck Pk=k1 CkT þ R; (19)
required for the implementation of Thevenin and Copetti models.
where Ck is given by:
4. Extended Kalman filter based on fuzzy models for SoC
estimation 
vh 
Ck ¼ (20)
vxk bx k=k1
The proposed fuzzy model is important for the estimation of the
SoC using the Extended Kalman Filter algorithm. The classic If (20) is evaluated at the prior state estimate b x k=k1 , then the
Kalman gain Kk is computed as in Eq. (21). This Kalman gain is af-
terwards used in Eq. (22) to generate the posterior estimate b x k=k .
Finally, the error covariance matrix Pk/k is calculated using Eq. (23).
 1
Kk ¼ Pk=k1 CkT Ck Pk=k1 CkT þ R (21)

  
x k=k ¼ b
b x k=k1 þ Kk yk  h bx k=k1 ; uk (22)

Pk=k ¼ ðI  Kk Ck ÞPk=k1 (23)

4.1. EKF based on a fuzzy model

The state space model that is proposed in this implementation

uses Eq. (2), which relates the SoC with the integral of the instan-
taneous current, as the state equation and the output voltage of the
Fig. 5. Fuzzy sets for current values. proposed model (9) as the observation equation, respectively:
360 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366

Table 1
Parameters of the experimental system.

Nominal power of the 4 kW

experimental system
Inductance for battery charging 30 mH
Battery bank 3 Trojan T-105. 18 V in total (nominal)
Nominal capacity 185 Ah
Switching frequency 4 KHz
Inductance and resistance 15 mH, 1 U
for battery discharging
Control platform DSK 6713, FPGA ACTEL A3P400
DC-link capacitance 2 Parallel-connected capacitors of
75 V, 33,000 mF each.

In summary, the EKF described is based on a new fuzzy model

which it is used as the observation equation with just one state
defined by the SoC. Thus, the resulting algorithm is efficient and
Fig. 6. Setup of the experimental system.
entails low computational burden when programmed on a
microprocessor.

d d Ts Ik1
SoC k=k1 ¼ SoC k1=k1  (24) 5. Experimental system
Cn
    In this work an experimental system has been designed and
b d b d
V k=k1 ¼ Voc SoC k=k1  Ik R int SoC k=k1 ; Ik : (25) implemented in order to

Therefore, the estimation of the SoC is obtained using the EKF

Validate the proposed methodology using experimental data,
(see Eqs. (15)e(23)), which is based on the model described by Eqs.
obtained by discharging a lead-acid battery bank. The same
(9)e(12). Thus the SoC is estimated using:
system is used to charge the batteries according to the charging
h i profile recommended by the manufacturers.
d
d ¼ SoC b
SoC k k=k1 þ Kk $ Vk  V k=k1 (26)
Experimental data has been used to comparate the performance
of the proposed model respect to the performance of other
h i1 models reported in the literature.
Kk ¼ Pk=k1 $CkT $ Ck $Pk=k1 $CkT þ R (27)
Experimental data has been used for comparate the perfor-
mance of the SoC estimator based on the fuzzy model and a SoC
Pk ¼ ½I  Kk $Ck $Pk=k1 ; (28) estimator based on a conventional model.

where Ck is the first term in a Taylor expansion of the fuzzy model The experimental prototype is shown in Fig. 6 and its topology
(see Eqs. (9)e(12)): in Fig. 7. This converter topology is relatively inexpensive and


vVk 
Ck ¼
vSoC cSoC k=k1
0 1
 r 
vVOC  Ik X vfj ðSoCk Þ
¼  Pr @ wj ðIk Þ$  A
vSoC SoC
c k=k1 j¼1 wj ðIk Þ j¼1
vSoCk  SoCc k=k1

(29)

and:

Pk=k1 ¼ Pk1=k1 þ Q : (30)

Fig. 7. Experimental system topology. Fig. 8. Control system and associate hardware for battery charging.
 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366 361

Fig. 9. Control system for battery bank discharging.

simple to implement. Information about the parameters of the

experimental system is provided in Table 1.
The input to the battery charging/discharging system (see
Fig. 7), is a variable output transformer designed to supply a voltage
of 40 V, 100 A. A bridge rectifier is required to feed a DC-link
capacitor bank of 66,000 mF. This DC-link is connected to two
blocks labelled “Battery charge” and “Battery discharge”. A Digital Fig. 11. VoCeSoC relation.
Signal Processor (DSP), Texas Instrument DSK 6713 is used to
control the whole system. As shown in Fig. 8 this DSP platform is
connected to a host computer. An FPGA based board is utilised to
generate the IGBT switching pulses and for protection purposes. An 2
IB ¼ ðv  vB Þ: (31)
additional board with 10 ADs channels of 14 bits, 1 ms of conversion sL2 2
time is provided. Hall-effect transducers are used to measure the
Using Eq. (31) and root-locus, it is relatively simple to design a
DC-link voltages, the voltage of the battery bank, the currents (IB1,
current control system. For this application a natural frequency of
IB2) and the current in the resistance where the energy is dissipated
un z 20 Hz and damping factor z ¼ 0.8 are used. This natural fre-
when the batteries are being discharged. The control systems
quency is considered sufficient for this application.
required to charge and discharge the battery bank are discussed
below.

5.1. Control system for charging the battery bank 5.2. Control system for discharging the battery bank

To charge the battery bank, the control system and associate To discharge the battery bank, the reference current regulated
hardware shown in Fig. 8 have been implemented in this work. To by the control system of Fig. 8 is negative (i.e. IB* < 0). Therefore in
operate with currents close to 100 A with minimum ripple in the steady state v2 < vB and the actuator is operated as a boost con-
battery bank, two inductors are used with interleaved switching verter, increasing the DC-link voltage Edc.
signals. The switching frequency used to control the system is 4 kHz When the DC-link voltage is above a given value, the control
for a sampling time (in the DSP implementation) of 250 ms. The host system shown in Fig. 9 is activated. A PI controller is used to
computer is also used to acquire and store intermediate data, with a regulate the DC-voltage v1 which is applied to the load composed of
sampling time of 1 s. This time is considered appropriate consid- an inductor L and a resistance R (see “battery discharge” block in
ering the relatively slow variation of the current profile. A MATLAB® Fig. 7). Neglecting the losses, in steady state the PI controller drives
interface has been implemented to acquire and process the data the current IL to the value:
sent from the DSP.
The reference charging current (see bottom left side of Fig. 8) vB IB
IL ¼ IT ¼ ; (32)
corresponds to a profile which is stored in the host computer. A PI E*
controller regulates the DC component of the voltage (see v2 in
Figs. 7 and 8) at the actuator output. Assuming that the two in- where IT is the total current provided by the battery bank to the DC-
ductors/IGBT legs are identical, the battery current IB is obtained as: link.
Considering the large DC-link capacitance, the natural frequency
of this control loop is un z 5 Hz, with a damping coefficient of
z z 0.8.

Fig. 10. Interleaved operation of the two legs used for battery charging. Fig. 12. Validation set 1.
362 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366

5.3. Interleaved operation

In order to charge and discharge the battery bank, with a rela-

tively high current and low ripple, the two inductors available for
this work were connected for interleaved operation. This allows a
considerable reduction in the current ripple, maintaining the
inductor currents below the rating values. For interleaved opera-
tion the FPGA was programmed to switch each IGBT with identical
duty cycle but with a phase shift of 180 .
The waveforms are shown in Fig. 10 where the switching pulses
and its associated inductor currents are displayed in Fig. 10a and b.
Notice that the DC current components are identical in both in-
ductors, however part of the harmonic distortion is eliminated
when the currents IB1 and IB2 are added. More information about
interleaved operation of current control system is elsewhere [42]. Fig. 14. Validation set 3. The subset 3 is defined from t ¼ 0 until t z 15 (see dotted
line).

6. Experimental evaluation of the model performance

Open circuit voltage curve
In this section data from experimental system discussed in
Thevenin model
Section 5 is used to identify the parameter of all the models pre-
Plett model
sented in Section 2. These are:
Copetti model

Proposed fuzzy model

For this evaluation the parameters of Eq. (1) have been identi-
fied using the relaxation tests described in Section 2.1 and genetic
algorithms [43]. This VoceSoC curve was obtained experimentally
and it is shown in Fig. 11. This relationship can be mathematically
written as (see Eq. (1)):

Vock ¼ 3:755SoC3k  5:059SoC2k þ 3:959SoCk þ 17:064 (33)

As described in Eq. (2), the SoC is calculated as the integral of the
instantaneous current with a sampling time of Ts ¼ 1 s and nominal
capacity (Cn) equal to 165 Ah, which is obtained from a capacity
experimental test at 10 A.
The comparative analysis of model performances is generated in
terms of the Root Mean Square Error (RMSE), which is given by:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uP  2
u N b
t i¼1 Vi  V i
RMSE ¼ ; (34)
N

where N is the number of samples (from experimental data), Vi is

the actual output voltage (obtained using the experimental system
b is the estimated output voltage obtained from
of Section 5) and V i
the battery bank model considered.

6.1. Parameter identification

The data used in the training and validation sets are obtained
from battery discharging profiles programmed in the DSP con-
trolling the experimental platform. Notice that temperature effects
in this work were not considered since the experimental setup was
allocated in a room with constant temperature. The parameters

Table 2
RMSE performance models on validation sets.

Models Set #1 Set #2 Set #3 Set #3*

RMSE [V] RMSE [V] RMSE [V] RMSE [V]

Thevenin 0.1548 0.1259 0.2457 0.1584

Plett 0.1174 0.1087 2.8840 0.2096
Copetti 0.0717 0.0978 1.7005 0.0644
Fig. 13. a) Validation set 2, b) model comparison in validation set 2, c) model com-
Fuzzy rules 0.0819 0.0866 0.0753 0.0589
parison in validation set 2, zoom for a time period.
 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366 363

Fig. 15. EKF estimators based on Copetti and fuzzy models in validation set 1. Fig. 17. EKF estimators based on Copetti and fuzzy models in validation set 3.

required in the Thevenin, Copetti, Plett, and fuzzy models are Rule 1 : If Ik is A10;1 then b
R int1
identified using the training sets shown in Fig. 2.
Using the discrete expression defined in Eq. (3) the parameters ¼ 0:070SoC4k 0:382SoC3k þ 0:619SoC2k
of the Thevenin model are estimated as: Rint ¼ 0.019 U;  0:383SoCk þ 0:118
C0 ¼ 28,747.99 F and R0 ¼ 0.013 U.
Rule 2 : If Ik is A15;2 then b
R int2

1
¼ 0:067SoC4k  0:338SoC3k þ 0:529SoC2k
Vk ¼ Vock þ ½Vk1  Vock1 $e0:013$28747:99  0:316SoCk þ 0:095
h 1
i
 0:013  ð0:013 þ 0:019Þ$e0:013$28747:99 $Ik1 (35)
Rule 3 : If Ik is A25;3 then b
R
 0:019$Ik ; int3

¼ 0:031SoC  0:219SoC3k þ 0:391SoC2k

4
The parameters of the Plett model (see (4)) are estimated as:
K0 ¼ 15.33, K1 ¼ 0.471, K2 ¼ 4.408, K3 ¼ 2.249, K4 ¼ 0.085 and  0:253SoCk þ 0:079
Rint ¼ 0.026 U.
Rule 4 : If Ik is A32;4 then b
R int4

¼ 0:083SoC  0:284SoC3k þ 0:374SoC2k

4

0:471  0:208SoCk þ 0:063

Vk ¼ 15:33  0:026Ik  þ 4:408SoCk
SoCk (36)
 2:249 ln½SoCk   0:085 ln½1  SoCk ; Fig. 5 shows the membership functions of the fuzzy sets AIj ;j .
b ðSoC ; I Þ is ob-
Therefore using AIj ;j (see Fig. 5) and Eq. (12), R int k k
The parameters of the Copetti model described in Eq. (6) are tained as model output.
estimated as: C10 ¼ 138.003 Ah, P1d ¼ 49.246, P2d ¼ 1.089,
P3d ¼ 0.063, P4d ¼ 2.082 and P5d ¼ 1.986.
6.2. Evaluating of the model performance using the RMSE criterion

In this work, three separated sets are used for validating pur-
" #
Ik 49:246 0:063 poses. These validation sets are not correlated with the training sets
Vk ¼ Vock  þ þ 1:986 ; (37) (see Fig. 2). A short description of each validation sets is shown
138:003 1 þ I 1:089 SoC2:082
k k below:
The fuzzy model rules obtained (based on the methodology Validation set #1 (Fig. 12). It includes battery discharge profiles
discussed in Section 3.2) are: obtained in different operating points. This validation set has been
designed to use instantaneous current levels different to those of
the training set.
Validation set #2 (Fig. 13a). It includes data from a battery dis-
charging current profile with three different levels of constant
power. This validation set has been designed to emulate the con-
sumption of electric vehicles operating at cruise speed [22].

Table 3
Performance indices on validation set 1.

Points EKF-Copetti [%] EKF-fuzzy rules [%]

C
I1k C
I2k F
I1k F
I2k

A1 0.40 1.19 0.17 0.47

B1 2.76 4.31 1.26 1.62
C1 4.93 4.68 0.17 0.18
D1 5.76 6.11 9.41 10.86
Fig. 16. EKF estimators based on Copetti and fuzzy models in validation set 2.
364 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366

Table 4 Table 6
Performance indices on validation set 2. Convergence times in validation set 1 using EKF estimators.

Points EFK-Copetti [%] EFK-fuzzy rules [%] Initial SoC Error EKF-Copetti [s] EKF-fuzzy rules [s]
C
I1k C
I2k F
I1k F
I2k 0.1 90% 4.6 2.9
0.3 70% 1.8 2.9
A2 2.91 8.68 0.56 1.53
0.5 50% 2.7 2.5
B2 0.77 1.52 3.51 5.68
0.7 30% 2.0 2.6
C2 1.77 2.55 0.63 0.79
D2 1.63 1.67 5.49 6.62
E2 13.0 15.6 10.1 13.3
F2 15.2 27.2 17.5 19.5 Table 7
Convergence times in validation set 2 using EKF estimators.

Initial SoC Error EFK-Copetti [s] EFK-fuzzy rules [s]

Validation set #3 (Fig. 14). It is generated by using a current 0.1 90% 4.6 3.1
profile that consists of periods where the battery discharges 30 Ah 0.3 70% 1.8 2.9
0.5 50% 2.7 2.5
(using different discharge currents each time) with interspersed 0.7 30% 2.0 2.5
periods of relaxation. This validation set has been designed to test
the performance of the model with different discharging current.
Notice that at the beginning of each experimental test, the value
of the SoC is 100%. Also, it must be noted that the charging profile  
d 
current recommended by the manufacturer, includes two stages. In  SoC k  SoCk 
I2k ¼ q ffiffiffiffiffiffiffiffi
ffi $100%; (39)
the first stage, the battery is charged at a constant current of 25 A; 2 Pk=k
while in the second stage constant voltage is applied to the battery
until the injected current is less than or equal to 5 A.
where Pk/k is the EKF covariance matrix (evaluated at the sampling
Table 2 shows the performance indices (RMSE) for Thevenin,
instant k). The index of Eq. (38) describes the error between the
Copetti, Plett, and fuzzy model, using the three aforementioned d and the real value SoCk.
estimated value SoC k
validation sets. Also, a subset of the set #3 (set #3*) includes only
The index of Eq. (39) is based on the error of the estimated SoC
SoC values between 0.1 and 1 to avoid the discontinuities that affect
value respect to the length of the confidence interval at 95%. If both
both Plett and Copetti models. Fig. 13b and c shows the results of
indices are below a given threshold, the SoC estimation would be
the models for the validation set #2.
close to its actual value.
From Table 2, it is concluded that the best results in terms of the
The R (covariance of measurement noise) and Q (covariance of
RMSE index are obtained with the Copetti and fuzzy models. Except
process noise) matrices required in the Kalman filter (see Eqs.
in set #1, the proposed fuzzy model behaves consistently better
(17)e(19)) were obtained using the data sheets of the corre-
than the other three circuital models. The proposed fuzzy model
sponding sensors, where information about the typical noise
has a slightly worse performance with validation set #1, because
introduced to the measurements is stated. For the voltage trans-
this model has not been trained to operate with low discharging
ducer, there is an error z0.9% with a noise variance of s2 ¼ (0.9/
currents (see Fig. 4). In all the cases where the proposed fuzzy
2)2 ¼ 0.2025, then R ¼ 0.2 is used.
model has been adequately trained, the fuzzy model outperformed
For determining the lower bound of R, the model error has to be
the others models including the Copetti model.
higher that the input variable error. For this application, the
Considering the results obtained using the validation tests, the
instantaneous current is the input, and the associated sensor error
two models with the best performance are used in the next section
is 1%. Then the variance is: s2 ¼ (1/2)2 ¼ 0.25. As the matrix R
to investigate the performance of a SoC estimation algorithm based
should be higher than 0.25 then Q ¼ 0.5 is selected. Thus, the initial
on an EKF.
covariance matrix is Po ¼ 0.5.

7. EKF estimators
7.1. Performance of the EKF estimators based on Copetti and fuzzy
To evaluate the performance of the EKF SoC estimators, the models
following indices are defined
SoC EKF estimators based on both Copetti and fuzzy models
  have been evaluated using the aforementioned validation sets.
d 
 SoC k  SoCk  Figs. 15e17 show the SoC ground truth (values that are considered
I1k ¼ $100% (38) as actual SoC) and the results of EKF estimators, assuming the
SoCk
battery is fully charged at the beginning of the experiment. SoC
ground truth was computed off-line by considering both the inte-
gral of the instantaneous current and battery open circuit voltage
Table 5 after each relaxing period.
Performance indices on validation set 3.

Points EFK-Copetti [%] EFK-fuzzy rules [%]

Table 8
C
I1k C
I2k F
I1k F
I2k Convergence times in validation set 3 using EKF estimators.
A3 2.38 8.32 0.46 1.53
Initial SoC Error EKF-Copetti [s] EKF-fuzzy rules [s]
B3 1.95 4.13 1.89 3.32
C3 2.70 3.54 4.65 5.61 0.1 90% 4.7 3.4
D3 2.63 2.47 2.48 2.82 0.3 70% 1.8 2.9
E3 6.47 6.08 0.45 0.48 0.5 50% 2.8 2.7
F3 1.47 1.80 2.47 1.87 0.7 30% 2.3 2.7
 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366 365

Fig. 18. Performance SoC estimator in validation set 1 (estimator is turn on in t ¼ 6 h). Fig. 20. Performance SoC estimator in validation set 3 (estimator is turn on in t ¼ 6 h).

Tables 3e5 show the values obtained by evaluating the perfor- Tables 6e8 show the convergence times for both estimators. From
mance indices of Eqs. (38) and (39) using the experimental data the information displayed in the aforementioned tables, we
depicted in Figs. 15e17. From Figs. 16 and 17, it can be noted that the conclude that a fast convergence for both EKF estimators is ach-
proposed estimators provide outputs that are consistent with the ieved. Also, note that convergence times are influenced by the
ground truth that was acquired in an off-line manner (validation initial value assumed for the SoC.
sets #2 and #3). Finally, Figs. 18e20 show the estimator performance consid-
According to what it was presented in Section 6.2, the estimator ering all three validation sets. In these experiments, the proposed
accuracy is directly affected by the fact that the fuzzy model was SoC estimator is turned on after the battery bank has been dis-
trained using data with battery discharge currents higher than 10 A. charged for at least six consecutive hours approx. As shown in
This issue becomes evident when analysing the estimator perfor- Figs. 18e20, the estimator converges to the SoC ground truth values
mance using set #1 (see Fig. 15). Also, the EKF estimators are turned after few iterations.
off during the relaxation periods; i.e., when the discharge current is
null.
8. Conclusions
Additionally, the estimator performance is evaluated using three
validation sets for different values of the initial conditions of SoC
In this paper, a new fuzzy model for the estimation of lead-acid
(30%, 50%, 70% and 90% of the initial SoC errors). Validation tests are
battery bank SoCs has been presented. The fuzzy rules are based on
used to obtain the convergence times of both EKF estimators
the battery polarisation resistance behaviour at different current
(Copetti and fuzzy). The convergence time is defined as the time tc
levels.
when the estimator output gets into a predefined zone, based on an
The fuzzy model has been compared with several conventional
upper and a lower band without leaving this zone again for t > tc.
models using the RMS error as goodness factor. In all the cases its
performance was better than that obtained by the conventional
models discussed in this work. Moreover, the proposed SoC esti-
mator does not have any numerical issues or mathematical dis-
continuities when the battery SoC is near to zero. This is certainly
an advantage of the proposed algorithm when compared to the
Plett and Copetti models.
In this work it has also been observed that the performance of
the models is improved by modelling the battery polarisation
resistance as a non-linear resistor whose value is dependant on the
SoC value and the discharging current. This is an additional
advantage of the proposed model when compared to other algo-
rithms where a fixed polarisation resistance value is assumed.
The fast converge of the proposed EKF estimator based on a
fuzzy model, allows on-line implementation. Moreover the fuzzy
model discussed in this work has converged to the correct SoC
value even when the initial SoC value has been very badly
estimated.

Acknowledgements

This work has been partially supported by FONDECYT project

Fig. 19. Performance SoC estimator in validation set 2 (estimator is turn on in t ¼ 6 h). 1140775 and the Complex Engineering Systems Institute (ICM: P-
366 C. Burgos et al. / Journal of Power Sources 274 (2015) 355e366

05-004-F, CONICYT: FBO16). The support of the Conicyt scholarship [19] P. Singh, R. Vinjamuri, X. Wang, D. Reisner, J. Power Sources 162 (2) (2006)
829e836.
programme for postgraduated studies, is also acknowledged.
[20] S. Malkhandi, Eng. Appl. Artif. Intell. 19 (5) (2006) 479e485.
[21] P. Singh, A. Nallanchakravarthula, in: Proceedings of the 13th International
References Conference on Intelligent Systems Application to Power Systems, Arlington,
VA, 2005.
[1] R. Palma-Behnke, C. Benavides, F. Lanas, B. Severino, L. Reyes, J. Llanos, D. S
aez, [22] U.S. Department of Energy, PNGV Battery Test Manual, Revision 3, DOE/ID-
IEEE Trans. Smart Grid 4 (2) (2013) 996e1006. 10597, 2001.
[2] V. Agarwal, K. Uthaichana, R.A. DeCarlo, L.H. Tsoukalas, IEEE Trans. Energy [23] C. Cai, D. Du, Z. Liu, in: The 12th IEEE International Conference on FUZZ '03,
Convers. 25 (3) (2010) 821e835. 2003.
[3] A. Purvins, M. Sumner, J. Power Sources (242) (2013) 742e755. [24] I. Li, W. Wang, S. Su, Y. Lee, IEEE Trans. Energy Convers. 22 (3) (2007)
[4] A. Salkind, C. Fennie, P. Singh, T. Atwater, D. Reisner, J. Power Sources 80 (1e2) 697e708.
(1999) 293e300. [25] S. Santhanagopalan, R. White, Int. J. Energy Res. 34 (2010) 152e163.
[5] M. Charkhgard, M. Farrokhi, IEEE Trans. Ind. Electron. 57 (12) (2010) [26] Y. He, X. Liu, C. Zhang, Z. Chen, Appl. Energy 101 (2013) 808e814.
4178e4187. [27] L. Xu, J. Wang, Q. Chen, Energy Convers. Manag. 53 (2012) 33e39.
[6] B. Saha, K. Goebel, in: Annual Conference of the Prognostics and Health [28] H. Zhang, M.-Y. Chow, in: IECON 2010 e 36th Annual Conference on IEEE
Management Society, San Diego, CA, 2009. Industrial Electronics Society, 2010, pp. 1844e1849.
[7] D. Vinh Do, C. Forgez, K. El-Kadri-Benkara, G. Friedrich, IEEE Trans. Veh. [29] S. Abu-Sharkh, D. Doerffel, J. Power Sources 130 (2004) 266e274.
Technol. 58 (8) (2009) 3930e3937. [30] S. Piller, M. Perrin, A. Jossen, J. Power Sources (2001) 113e120.
[8] B. Olivares, M. Cerda, M. Orchard, J.F. Silva, IEEE Trans. Instrum. Meas. 62 (2) [31] H. Dai, Z. Sun, X. Wei, in: ICVES, IEEE International Conference on Vehicular
(2013) 364e376. Electronics and Safety, 2006, pp. 342e347.
[9] B. Pattipati, C. Sankavaram, K. Pattipati, IEEE Trans. Syst. Man Cybern. Part C [32] T. Hu, B. Zanchi, J. Zhao, IEEE Trans. Energy Convers. 26 (3) (2011) 787e798.
Appl. Rev. 41 (6) (2011) 869e884. [33] F. Sun, R. Xiong, H. He, W. Li, J.E.E. Aussems, Appl. Energy (96) (2012)
[10] D. Cadar, D. Petreus, C. Orian, in: 15th International Symposium for Design 378e386.
and Technology of Electronics Packages (SIITME) 2009, 2009, pp. 257e260. [34] J. Kim, B.H. Cho, IEEE Trans. Veh. Technol. 60 (9) (2011) 4249e4260.
[11] S. Qingsheng, Z. Chenghui, C. Naxin, Z. Xiaoping, in: 29th Chinese Control [35] A. Vasebi, M. Partovibakhsh, S.M. Taghi Bathaee, J. Power Sources (174) (2007)
Conference (CCC) 2010, 2010, pp. 5999e6003. 30e40.
[12] Z. Di, M. Yan, B. Qing-Wen, in: 30th Chinese Control Conference (CCC) 2011, [36] G.L. Plett, J. Power Sources 134 (2004) 262e276.
2011, pp. 6256e6260. [37] X. Liu, Y. He, Z. Chen, in: 2nd International Conference on Software Engi-
[13] S. Lee, J. Kim, J. Lee, B. Cho, J. Power Sources 185 (2) (2008) 1367e1373. neering and Data Mining (SEDM), 2010, pp. 27e31.
[14] X. Tang, X. Mao, J. Lin, B. Koch, in: American Control Conference (ACC) 2011, [38] J. Copetti, F. Chenlo, J. Power Sources 47 (1994) 109e118.
2011. [39] J. Copetti, E. Lorenzo, F. Chenlo, Prog. Photovoltaics 1 (1993) 283e292.
[15] L. Ran, W. Haiying, L. Gechen, in: 5th IEEE Conference on Industrial Electronics [40] R. Babuska, Fuzzy Modeling for Control, KAP, United States, 1998.
and Applications (ICIEA), 2010. [41] K.J. Astrom, B. Wittenmark, Computer-controlled Systems: Theory and Design,
[16] M. Dalal, J. Ma, D. He, Proc. Inst. Mech. Eng. Part O J. Risk Reliab. (2011) 81e90. Prentice Hall, 1996.
[17] P. Singh, C. Fennie, D. Reisner, J. Power Sources 139 (2) (2004) 322e333. [42] D. Newlin, R. Ramalakshmi, S. Rajasekaran, in: IEEE International Conference
[18] A. Zenati, P. Desprez, H. Razik, in: 36th Annual Conference on IEEE Industrial on Green High Performance Computing (ICGHPC), 2013, pp. 1e6.
Electronics Society EICON 2010, Glendale, AZ, 2010. [43] K. Man, K. Tang, S. Kwong, Genetic Algorithms, Springer-Verlag, 1999.