This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
1

Battery Model Parameterization using

Manufacturer Datasheet and Field Measurement
for Real-time HIL Applications
Fuhong Xie, Student Member, IEEE, Hui Yu, Student Member, IEEE, Qian Long, Student Member,
IEEE, Wente Zeng, Member, IEEE, and Ning Lu, Senior Member, IEEE

 Indices
Abstract — This paper presents a novel battery model
parameterization method using actual field measurement and r Index of battery model outputs.
manufacturer datasheet for real-time hardware-in-the-loop (HIL) k Index of iteration conducting local exploitation.
applications. It is critical that real-time HIL models can accurately n Index of the battery model parameters.
reproduce field test results so that tests can be conducted on HIL i, j Index of the battery model parameter sets within
testbeds instead of in the field. In the past, numerical heuristic the solution pool calculated by the GA.
optimization algorithms were often used to derive parameters for s Index of iteration conducting global search.
battery models. However, the deterministic algorithms often reach
a locally optimal solution and stochastic heuristic searching Variables
strategies suffer from low searching efficiency. Therefore, in this Em(t) Battery internal voltage source at time t.
paper, we propose a global-local searching enhanced genetic R0(t) Battery terminal resistance at time t.
algorithm (GL-SEGA). By applying the generalized opposition-
R1(t) Battery dynamic branch resistance at time t.
based learning mechanism, GL-SEGA can efficiently explore the
global solution space. By using the trust-region-reflective method C1(t) Battery dynamic branch capacitance at time t.
to perform the local search, the GL-SEGA can improve the R2(t) Battery thermodynamic resistance at time t.
accuracy and convergence in its local exploitations. Field T(t) Battery electrolyte temperature at time t.
measurements and manufactory datasheets are used to test and Ta(t) Ambient temperature at time t.
validate the accuracy and robustness of the GL-SEGA algorithm. SOC(t) Battery state-of-charge at time t.
PR (t) Battery internal heat loss at time t.
Index Terms — battery model, genetic algorithm, opposition
V (t) Battery terminal voltage at time t.
based learning, parameterization, trust-region-reflective.
Im(t) Battery load current at time t.
Zmain(t) Total battery branch impedance at time t.
I. NOMENCLATURE x (θ, t) Simulated model output with θ at time t.
x (t ) Field measured output at time t.
Acronyms
Set
BESS Battery energy storage systems.
CFDM Computational fluid dynamics model. θ Set of N parameters for battery model.
ECM Equivalent circuit model. l Set of lower limit for N model parameters.
EM Empirical model. u Set of upper limit for N model parameters.
GA Genetic algorithm. θ0 Set of initial N parameter for local exploitation.
GN Gauss-Newton. δ Set of trail step for N model parameters.
GOL Generalized opposition-based learning. Δ Set of scaled radius of the trust region.
GL- Global-local searching enhanced genetic Δmax Set of max radius of the trust region.
SEGA algorithm. θ* Output of Algorithm 1.
HIL Hardware-in-the-loop. θ op Output of Algorithm 2.
KOP Key operating points. θ rec Output of recombination operation.
NM Nelder-Mead. θ GA Output of the Improved Genetic Algorithm.
OL Opposition-based learning. Parameters
PSO Particle swam optimization. Em0 Battery open-circuit voltage.
SOC State of charge. KE Positive constant for Em(t).
TRR Trust region reflective. R00 Battery internal resistance when fully charged.
TSD Time series data. K0 Positive constant for R0(t).

The authors are with the Electrical & Computer Engineering Department (FREEDM) Systems Center at North Carolina State University, Raleigh, NC
and the Future Renewable Electric Energy Delivery and Management 27606 USA. (Emails: [email protected]; [email protected]).

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
2

R10 Positive constant for R1(t). parameterization using both factory datasheets and time-series
τ1 Time constant for battery dynamic branch. field measurements for HIL applications.
K21, K22 Negative constants for R2(t). There are two critical steps for developing a high-fidelity
Im0 Battery nominal C10 current. BESS model: the first step is to select a BESS model that can
Tini Battery electrolyte initial temperature. faithfully reproduce both steady-state and dynamic operating
RT Environment-to-battery thermal resistance. characteristics and is computationally efficient to meet the real-
CT Battery thermal capacitance. time simulation requirements. The second step is to select a
Qini Battery initial extracted charge. model parameterization method that uses time-series data as
C0 Battery nominal capacity at 0°C. inputs to derive an optimal set of model parameters that can
KT Temperature-dependent look-up table. produce accurate simulation results under a wide range of
KC Positive constant for battery SOC estimation. operating conditions.
N Number of model parameter. Three different approaches for the battery modeling were
NGA Number of parameter set for genetic algorithm. proposed in the literature [1]-[6]: computational fluid dynamics
Nsample Total number of the field measurement in a day. model (CFDM), equivalent circuit model (ECM), and the
w Weighting factor for different model outputs. empirical model (EM).
J Model Jacobian matrix at parameters set θ. The EM approach is computationally efficient but its
H Model Hessian matrix at parameters set θ. accuracy is insufficient for conducting precise simulations. The
C Compensation matric for H. CFDM method is a physics-based modeling approach that
D Scaling matrix for trail step. allows electrochemical characteristics of a BESS to be modeled
Jv Sign of Jacobian matrix. in details to produce accurate results. There are three defects of
𝜙 Performance metrics for local exploitation. using CFDM for HIL applications. First, CFDM requires too
μ1, μ2 Updating criterion for the trust region radius. many model parameters. Second, different kinds of CFDMs are
α1, α2 Updating scale for the trust region radius. required to model different types of batteries. Third, because of
ε Max allowed tolerance. the model complexity, CFDMs developed for BESSs can only
λ Radom real number within [0, 1]. be used for millisecond level simulations, making them unfit
lnd Minimum of nth parameter within solution pool. for the microsecond level HIL simulations [7]-[8].
und Maximum of nth parameter within solution pool. The ECM approach is measurement-based and it requires
p Probability for selecting the parameter set. only a small number of model parameters. Different battery
technologies can be modeled by using the same ECM structure,
making the parameterization process of the ECM easier to be
II. INTRODUCTION
standardized than that of the CFDM. Moreover, ECM is

I N recent years, battery energy storage systems (BESSs) have

been rapidly integrated into power grid operation. Before
large-scale field deployment of BESSs, it is necessary to
computationally efficient so it is widely used in the
microsecond level real-time HIL simulations [9]. Therefore, in
this paper, we select ECM as the battery modeling approach.
conduct field experiments to benchmark both the steady-state In general, the KOP-based method uses nominal data given
operation and the dynamic response characteristics of a BESS by manufacturers or collected in simple field tests to identify
under a wide variety of system operating conditions. model parameters [10] and it is a straightforward approach to
However, conducting test experiments in the field is not only develop a generic battery model. However, because of the non-
prohibitively expensive but also risky because of the lack of linearity inherent in battery operations, the KOP-derived model
operational experience and possible design flaws. Reduction of produces results that are accurate only within a small region
field test cost and safety considerations make the development around KOPs and the model performance is susceptible to noise
of high-fidelity BESS models an attractive solution. and error [11]. Thus, KOP-method tuned model often result in
A high fidelity, hardware-in-the-loop (HIL) test system poor performance when the model run under different operating
allows controllers and equipment to be tested and validated in conditions, making it unfit for high-fidelity HIL applications.
laboratory environments under different operating conditions. The TSD-based method optimizes overall time-series errors
To make a HIL testbed high fidelity, several field tests are between simulation results and field measurements [12]-[15],
often needed to tune the model parameters. This is to verify how [25]; therefore, TSD-tuned models can produce more accurate
well the simulated results produced by the HIL test system can simulation results for different operating conditions than KOP-
match the actual field measurements. tuned models. Rahman et al. [13] proposed a gradient-free
For example, to guarantee the HIL battery model can technique where the particle swarm optimization (PSO) method
accurately represent different microgrid operating scenarios in is used to parameterize the lithium-ion battery model. However,
both steady-state and dynamic response performance, field tests the heuristic optimization-based method could have numerical
normally last for hours or days to capture different operating difficulty because of the ill conditionality inherent in the battery
conditions. The obtained field measurements are therefore parameterization problem. Barreras et al. [14] presented a
time-series data (TSD) in contrast to isolated data sets collected nonlinear dynamic electro-thermal ECM for lithium-ion
from key operating point (KOP) based tests. batteries to test battery management software on a HIL testbed,
Thus, in this paper, we focus on discussing battery model where Nelder-Mead (NM) method is used to solve the model

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
3

parameterization problem. However, the NM searching method with the battery state of discharge (SOC) at time t, we have
often converges to a non-stationary point. Dvorak et al. [15]
R0 (t )  R00  1  K0 1  SOC(t )   (2)
combined the PSO with the Gauss-Newton (GN) method to
achieve a balance between searching efficiency and accuracy. We also assume the battery branch resistance R1 varies with
Because of no box-constraint is considered in the GN method, respect to SOC instead of depth of discharge, it is calculated as
the obtained parameters may violate the allowed limits. R1 ( t )   R10 ln  SOC ( t )  (3)
To improve the computational speed, achieve optimality and
guarantee parameter bounds, we developed a global-local Assuming that the branch time constant τ1 is a fixed positive
searching enhanced genetic algorithm (GL-SEGA) for battery constant, C1 is calculated as
model parameterization using manufacturer datasheet and field C1 (t )   1 R1 (t ) (4)
measurements for high-fidelity HIL applications. Our main Thermodynamic resistance R2 is related with the battery
contributions are summarized as follows: current Im following through the main branch and increases
1) GL-SEGA can efficiently explore the entire solution space exponentially when the battery SOC increases, so we have
for an optimal parameter set by applying the generalized
opposition-based learning (GOL) mechanism. R2 (t )  R20  e

K 21 1 SOC ( t ) 
1  e K 22  I m ( t ) I m 0 
 (5)
2) The estimation accuracy is significantly improved by using
Then, the heat loss of battery, PR (t), can be represented as
the trust-region algorithm and considering the bound of
PR (t )   R0 (t )  R2 (t )    I m (t ) 
2
model parameters. Compared with existing methods, GL- (6)
SEGA-tuned models converge faster with higher accuracy.
The battery terminal voltage V (t) can be calculated as
3) GL-SEGA can be used as a generic method for battery
model parameterization. This is because many types of V (t )  Em (t )  I m (t )  Z main (t ) (7)
BESSs can be represented by the same ECM structure. The where Zmain is the total impedance of the main circuit branch
proposed GL-SEGA method represents an initial step represented in the dotted box in Fig. 1(a).
towards online BESS model parameterization, the goal of Using the thermodynamic branch model shown in Fig. 1(b),
which is to derive battery system control and operation we can calculate the battery electrolyte temperature T (t) as
parameters in real-time for different operation conditions.
1 t
The rest of the paper is organized as follows. Section III T (t )  Tini    PR ( )  T ( )  Ta ( )  RT   d (8)
describes the modeling method. Section IV introduced the CT 0

formulation of the proposed algorithm. Section V presents the In this paper, we use temperature data measured from the
simulation results and performance evaluation. Conclusions battery surface to represent the battery electrolyte temperature.
and future work are provided in Section VI. This is because temperature sensors often measure cell surface
temperature instead of the internal electrolyte temperature.
III. MODEL METHODOLOGY The battery SOC varies concerning battery activities and it is
This section presents the modeling for battery ECM and the correlated to the battery temperature T at time t, so we have

 
objective function for the optimal parameterization problem.  1 t 
SOC (t )  1   Qini   I m ( )d   100% (9)
 K C KT (T )C0
0
A. Battery Equivalent Circuit Model 
In [16], Jackey proposed a simple but effective ECM for where KT is a temperature-dependent look-up table that can be
battery cell modeling. As shown in Fig. 1(a), we simplified found in the manufacture datasheet [17].
Jackey’s model by removing the parasitic branch assuming that Although the ECM only represents one single battery cell, by
the battery parasitic current loss is negligible. assuming all battery cells are identical and multiplying with the
Zmain C1 number of cells, the ECM can be used to mimic the dynamic
POut

electric behaviors of a multi-cell battery pack.
R2 R0 T
R1 Ta Moreover, minute-by-minute test measurements collected in
PR RT
Em Terminal the field have shown that simulation results produced by the one
Voltage CT
RC branch ECM match the field measurements with reasonable
Im  accuracy. A multi-RC-branches ECM can be used if field
(a) (b) measurements at higher resolutions are available or modeling
Fig. 1. Battery equivalent circuit models: (a) battery electrical main branch
circuit; (b) battery thermodynamic branch circuit.
greater details of electric dynamic responses are requested [18].

When the battery is fully charged, the internal voltage source B. Objective Function
Em equals to the open-circuit voltage, Em0. At time t, Em(t) can To benchmark the battery ECM model for HIL applications,
be calculated as Total S.A. conducted a series of benchmark field tests at their
E m (t )  E m 0  K E T ( t )  1  SOC ( t )  (1) testbed in Lyon, France. In each test, five variables are
measured: SOC, Im, V, Ta and T. Note that temperature sensors
Assuming the battery internal resistance, R0, is independent are mounted on the battery surface because the battery internal
of the battery temperature but it varies negative-proportionally electrolyte temperature cannot be directly measured. For the Li-

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
4

ion battery, we use the SOC measurement provided by the As introduced in [20]-[21], the general constrained TRR sub-
battery manufacturer. For the lead-acid battery, the battery SOC problem is used to compute the trail step δ that minimizes ψ (δ)
is measured by an embedded SOC estimator inside the inverter. over G. The optimization problem is formulated as
As shown in Fig. 2, the measured Im and Ta are used as inputs  1 
of the ECM and the simulated V, SOC, and T are model outputs. minN   k   J k T  k   k T ( H k  Ck ) k  (15a)
 k   2 
The summation of squared errors (SSE), f (θ), is calculated as
2 s.t.
f ( )   t sample  x( , t )  x (t )  ,    N
N
(10)
1 Dk  k   k (15b)
The goal of the parameter estimation problem is to find an
optimal set of model parameters so that the mismatch between Dk  diag 1  v( k )  (15c)
simulation results and field measurements is minimized. The
objective function for minimizing errors is formulated as Ck  Dk  diag  J k   J kv  Dk (15d)

  N

min F ( )   r 1 wr  f r ( ) : l    u , l , u   N
3
 (11) where Ck is formulated to ensure the reversibility of Hk.
For the nth parameter θk,n of a model parameters set θk, at
iteration k,1 < n < N, a vector function vn (θk,n) is defined as

3
r 1
wr  1 (12)
 n  u n , J k ( k , n )  0 and un  
Note that wr is the weighting factor corresponding to the three 
  l , J k ( k , n )  0 and ln  
model outputs [V, SOC, T], as shown in Fig. 2. Within this vn  k , n    n n (16)
paper, SOC is within [0, 100] (Unit: %), V is within [40, 70]  1, J k ( k , n )  0 and un  
(Unit: Volt); T is within [0, 70] (Unit: °C).  1, J k ( k , n )  0 and ln  

v
Model Im
V
Model Note that J is a diagonal matrix representing the sign of
k
Battery Model SOC
Inputs Ta
T
Outputs Jacobian matrix Jk at iteration k and it is expressed as
Fig. 2. Inputs and outputs for lead-acid battery model.  0, J k ( k , n )  0 or u n ln  
J kv, n   (17)
 sign  J k ( k , n )  , otherwise
IV. SEARCHING ENHANCED GENETIC ALGORITHM
Trust region method guarantees a sufficient decrease of the
This section introduces the model parameterization method. objective function by adjusting the radius Δk of G while using
First, the trust-region-reflective (TRR) method for improving the trial step δk to indicate the direction reducing the objective
the accuracy and convergence of the local minimum search is
function. The performance metrics, 𝜙k, is calculated as
presented. Then, the GOL algorithm for efficiently exploring
the global solution space of the genetic algorithm (GA) is 1  1 T 
k   F ( k   k )  F ( k )   k Ck  k  (18)
presented. In the end, the GL-SEGA approach is introduced.  ( k )  2 
A. Trust-Region-Reflective Algorithm where 0.5×δkTCkδk is the compensation term by introducing Ck.
Because the battery model is a complicated nonlinear system, If the trial step can sufficiently reduce the objective function
and the updating index is larger than the upper threshold μ2, the
the associated model parameterization is an ill-conditioned
trust region is expanded by letting α2 > 1 to allow a larger step
problem. Therefore, it is difficult to use traditional least-squares
methods to find an accurate and robust solution. In this paper, in the next iteration. Otherwise, if the approximation is poor and
we use the trust-region method to parameterize battery models. the updating index is less than the lower threshold μ1, the trust
region G is contracted by letting α1 < 1. Thus, we have
The trust region method searches for the optimal solution
within a small subset. We denote the subset as the trust region  1   k , k  1

G, which can be approximated around the initial position using  k 1   min  2   k ,  max  , k   2
a model function [19]. Because battery model parameters are  (19)
 k , otherwise
usually limited within a range, the TRR method can be used to
bound the model parameters by applying a single reflection 0  1   2  1
transformation [20]-[21]. By nature, TRR is a local exploitation The model parameters are updated based on 𝜙k, such that
method, making its performance heavily affected by initial
   k , k   2
points. To avoid potential initialization issues, we propose a  k 1   k (20)
global searching method in the later section to provide good  k , otherwise
initial points for the TRR algorithm. When using TRR to solve a bounded problem, given a trial
To implement the TRR method, we first simplify (11) to a step δk,m for the mth parameter θk,m, it can be re-calculated as
quadratic function ψ (θ) using the first two terms of the Taylor   k ,n , nm
expansion at an initial point θ0 within G. Then, we have  kR, n   (21)
  k , n  2 J
v 2
k ,n D , nm
k ,n
1
 ( )  F ( 0 )  F ( 0 )T    T H  (13)
 k ,n ,
2  nm
 kB, n   2 (22)
  0   (14)  v
 k ,n k ,n n  m
J D ,

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
5

where (21) and (22) mean if θk,m+δk,m falls out of the parameter Algorithm 1: TRR for Model Parameterization
range, δk,m will be reflected by the boundary it violates (δk,nR ) or Initialization: Select the initial values for θ0, Δ0, Δmax, kmax, and ε.
Output: optimal parameters set θ* = θk.
capped by the scaled steepest descent direction –D-2k Jkv (δk,nB ). 1. for k = 0, 1, …, kmax
Then, we calculate the trial step δk again 2. Compute Jk, Dk, Hk, Ck and J vk .

 
3. Compute the trial step δk by solving (15).
 k  arg min   kR  ,   kB  (23) 4. if θk + δk cross any bound constraint then
 k  N
Recalculate δk using (21)-(23).
The TRR algorithm is summarized as Algorithm 1. end if
5. Compute the updating index 𝜙k using (18).
6. Compute Δk+1 using (19) and θk+1 using (20).
B. Generalized Opposition-based Learning Mechanism
7. if ||Jk||2 ≤ ε then Stop.
The opposition-based learning (OL) method is an effective else
way to improve the performance of heuristic optimization based Go to step 2 until the maximum iteration number is reached.
end if
algorithms. OL method evaluates the corresponding opposite 8. end for
solution of the original candidate solution to increase the chance
of finding a better solution [22]. As shown in Fig. 3(a), let θ be Algorithm 2: GOL-based Transformation
the solution of the parameter estimation problem and θ op is the Input: Get the values of θ, and the upper and lower bounds of l, u.
opposite solution. For the nth parameter θn, we have Output: transformed opposite solution θop.
1. for n = 1, 2,…, N
 nop    (ln  u n )   n ,   1 (24) 2. Compute θop using (25)
3. if θnop < ln or θnop > un then Correct θnop using (26).
If λ is a random number within [0, 1], the OL method can be end if
4. end for
extended to the GOL method that can have better performance
[23] by using the knowledge of the current search space. Then, Define the fitness function, Γ(F), as
the opposite solution can be calculated using the dynamically
  F ( )    F ( ) 
1
updated boundaries [ld, ud], as shown in (25) and Fig. 3(b). (27)
 nop    (lnd  u nd )   n ,   [0,1] (25a) The parent individuals are selected from NGA individuals
based on individual fitness Γ(F). Higher fitness leads to a
l  min(i , n ) , u  max(i , n )
d
n
d
n (25b) greater chance for a set of θ to be selected. Thus, the possibility
i i

Comparing (24) with (25), we can see that the basic opposite to be selected can be defined as


NGA
solution is the symmetry point of the original number centered pi  (i ) i 1
(i ),    N (28)
on the point (ln + un)/2 and it is still located in [ln, un], as shown
in Fig. 3(a). However, when using the GOL method and θ op is The recombination operation is used to exchange the genetic
located in [λ(lnd + und ) - und , λ(lnd + und ) - lnd ], it is possible that the information of two parent individuals to generate two new child
GOL calculated θ op will jump out of [ln, un], making the solution individuals whose goal is to randomly generate new solutions
infeasible, as shown in Fig. 3(b). from an existing individual pool. For the nth parameter of the ith
If the nth opposite parameter θnop is outside [ln, un] and violates and jth set of θ, the recombination method can be expressed as
the given ranges, the GOL will reassign θnop to a random value  i , n  1      i , n     j , n
within [lnd , und ] by (29)
 j , n  1      j , n     i , n
 nop  rand (lnd , u nd ), if  nop  [ln , u n ] (26)
The mutation operation is used to maintain genetic diversity.
The GOL algorithm is summarized as Algorithm 2. In this paper, we introduce the GOL scheme into the mutation
 nop
operation to formulate an improved GA for a better searching
θn
performance. Half of NGA individuals with relatively worse
ln (ln+un)/2 un fitness will be transformed into the opposite individuals using
(a) the GOL-based Transformation (i.e., Algorithm 2).
 nop θn The best set θiGA from GA will be used as the initial point θ0
for TRR algorithm (i.e., Algorithm 1) to find the local optimal
λ(lnd+und)-und λ(lnd+und)-lnd ln un solution θi* , which will be fed back to GA again as a new child
(b) individual. Fig. 4 illustrates an example is used to the global and
Fig. 3. Conceptual descriptions of the learning process: (a) opposition-based
learning; (b) generalized opposition-based learning. local search process of the GL-SEGA.
The diagram for implementing GL-SEGA is shown in Fig. 5.
C. Improved Genetic Algorithm Within each iteration s, Half of NGA individuals with relatively
In section III.A, we introduced the TRR method to search the poorer fitness will be mutated using Algorithm 2. Because it
local optimum within the trust region G around θ0. Because the may take some time before Algorithm 1 reaches stop criteria, to
optimality of the TRR solution can be significantly affected by improve searching efficiency and lower computation time, GL-
θ0, a heuristic improved GA is used to search globally for sets SEGA will request the Algorithm 1 every sL iteration.
of θ values within the given ranges of each parameter. We will When calling Algorithm 2, the TRR algorithm will be used
use the best set of θi GA values among NGA sets of θ as θ0 for to further optimize the solution with the best fitness function so
Algorithm 1 to find the local optimum θi* around θ0. that a local optimum can be found (see Set 3 in Fig. 4). Then,

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
6

Step 1: Select 6 parent sets of θ values for GA algorithm the optimal individual θj* obtained from TRR, the mutated sets
based on the fitness Γ.
from GOL, and the rest of the sets will be recombined as the
1 2 3 4 5 6
new child population for the next round of GA optimization.
θ1 θ2 θ3 θ4 θ5 θ6
The best individual θjbest, which has the highest fitness, will be
Step 2: If Set 1, 4, and 5 have the lowest fitness, they will
be selected to mutate. updated and recorded at the end of each iteration.
1 2 3 4 5 6 If the iteration index s reaches its maximum number smax, the
θ1op θ2 θ3 θ4op θ5op θ6 estimation algorithm will be terminated and the latest best
Step 3: Randomly select two sets of θ to recombine and if solution θjbest will be used as the battery model parameters.
Set 2 and 6 are selected.
1 2 3 4 5 6 V. CASE STUDIES AND ANALYSIS
θ1op θ2rec θ3 θ4op θ5op θ6rec
Step 4: Calculate the fitness Γ of the 6 sets of parameters.
A. Algorithm Setting
If Set 3 has the highest fitness, Set 3 (θ3GA) is selected as The GL-SEGA is implemented in MATLAB with Parallel
the one to proceed with the local exploitation for the θ3*.
Computation Toolbox. The value of all the constants used in the
1 2 3 4 5 6
GL-SEGA are listed in Table I. The battery model is created in
θ1op θ2rec θ3* θ4op θ5op θ6rec
Simulink and run on an OPAL-RT based HIL testbed.
Step 5: Repeat step 1.
Simulation results are compared with field measurements
Fig. 4. An illustration of the global and local search process (NGA = 6).
using the performance metrics listed in Table II. SSE and the
Start mean square error represent the square matching error thus they
vary from zero (perfect fit) to positive infinity (worst fit). The
Iteration index s = 0. normalized root mean square error and the normalized mean
square error facilitate the comparison of results with different
Randomly generate the initial solution population and
calculate the fitness for each individual.
scales and they vary from negative infinity (worst fit) to one
(perfect fit). The battery specifications are listed in Table. III.
s=s+1 TABLE I
SETUPS OF THE GL-SEGA ALGORITHM
Parent individuals selection Symbol Description Quantity
operation for GA. w1 , w2 , w3 Weighting factors in the objective function. 1/3, 1/3, 1/3
kmax , smax Max iteration number. 100
Mutation operation for the half individual population Δmax Max scaled radius of the trust region. 3
who has relatively poor fitness using GOL. Δ0 Initial scaled radius of the trust region. 1.5
μ1 , μ2 Constants used to update estimations. 0.25, 0.75
Recombination operation for α1 , α2 Constant used to update the trust region. 0.5, 2
parent individuals for GA. ε Max tolerance for TRR algorithm. 10-6
NGA Number of the individuals in the pool. 20
sL Interval for calling TRR algorithm. 10
Calculate the fitness for each individual.
pre Possibility of the recombination operation. 0.6
pmu Possibility of the mutation operation. 0.9
No
Mod(s, sL) = 0
TABLE II
Yes DEFINITION OF THE PERFORMANCE METRIC
Metrics Equation
Select the best fitness
individual as θjGA . Summation of squared errors
Remove the best (10)
(SSE).
fitness individual (θjGA)
1 2
Local exploration
 sample  x( , t )  x (t ) 
N
from the solution pool. Mean square error (MSE).
using TRRA. N sample t 1
2
  x( , t )  x(t ) 
N sample
Output the optimal individual θj* Normalized root mean square
and calculate the fitness. error (NRMSE). 1 t 1

N sample max  x (t )   min  x (t )  

2

2
  x( , t )  x (t ) 
N sample
Reform the population.
Normalized mean square error
1 t 1
(NMSE). N sample  max  x (t )   min  x (t )  
2

Record and update the best

fitness individual θjbest.
TABLE III
BATTERY SPECIFICATIONS
Yes
s < smax Case 1 Case 2 Case 3/4/5
No Battery Type Li-ion Li-ion Lead-acid
Model Number \ Phi 3.4 SBS 900
Output the latest
Cell Nominal Voltage (V) 7 51.2 2
best individual θjbest
Battery Cell Nominal Capacity (Ah) 5.4 67 900
Number of Series 1 1 1
End
Number of Cells in each Series 1 1 24
Fig. 5. Flowchart of the proposed GL-SEGA.

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
7

B. Case 1: Benchmark Test

In the benchmark test, the convergence speed and optimality
of GL-SEGA is compared with three existing parameterization
methods: the PSO-GN algorithm [15], the hybrid multi-swarm
PSO algorithm [24], and the PSO-NM algorithm [14], [25].
(a) (b)
Identical or similar algorithm settings (See Table I) are used to
achieve impartial performance comparison.
The benchmark test model is shown in Fig. 6. The battery
model given in MATLAB is configured as a Li-ion battery cell
using settings given in Table III. Assuming that the ambient
temperature is held at 20°C, a current profile (see Fig. 7(a)) is
injected into the model and the responses (see Figs. 7(b)-(d)) (c) (d)
Fig. 7. Benchmark test profiles: (a) model current; (b) battery terminal voltage;
are recorded as target profiles for model parameterization. (c) battery SOC and (d) battery cell temperature.

Voltage
Test Current
Profile
SOC
Ambient Battery
20
Temperature Temperature

Fig. 6. Model setups for the benchmark test.

Then, parameterization algorithms are used to parameterize

the battery model by comparing model outputs with target
profiles. Model parameters and their true values are listed in
Table IV. Parameterization algorithms search for the optimal
Cmax (Ah)

parameter set within the searching range that minimizes the

mismatches between simulated outputs and target profiles.
As shown in Fig. 8 and Table. V, compared with GL-SEGA,
other methods have relatively slower converge speed. It takes
40 iterations for GL-SEGA to converge, while PSO-GN and
HMPSO require 83 iterations and 92 iterations, respectively.
PSO-NM cannot reach a stable estimation for Rin within 100
iterations. Regarding accuracy, both GL-SEGA and PSO-GN
converge to the correct value, while hybrid multi-swarm PSO
Fig. 8. Evolution curve for model parameters using different algorithms.
and PSO-NM have about 1% estimation error.
TABLE IV
C. Case 2: Li-ion Battery Cell MODEL PARAMETERS TO BE ESTIMATED
Because GL-SEGA and PSO-GN achieve better matching Definition True Value Unit Search Range
Internal resistance Rin 0.0133 Ω 0.01 ~ 0.02
results among four algorithms in Case 1, they are adopted to Thermal resistance Rth 0.6 °C/W 0~1
derive ECM parameters for a single Li-ion battery cell using Maximum capacity Cmax 5.6 Ah 0 ~ 10
characteristic data sets (i.e., terminal voltage, battery SOC, and Nominal Voltage En 7 V 0 ~ 7.9
temperature) provided by the battery manufacturer [26]. The TABLE V
characteristic data was collected from a constant current ALGORITHM EVALUATION FOR BENCHMARK
discharging test and the discharging current is 32A. The battery Metrics Parameter GL-SEGA PSO-GN HMPSO PSO-NM
specifications are listed in Table III. Rin 35 83 92 /
Converge Rth 24 75 89 88
As shown in Fig. 9, when implementing the ECM parameters Iteration Cmax 40 75 88 81
derived by GL-SEGA, the battery model can produce battery En 10 78 91 86
voltage, SOC, and temperature curves that match very well with Rin 0.0133 0.0133 0.0133 0.01245
the curves provided in the manufacturer datasheets. Final Rth 0.6 0.6 0.5986 0.5947
Value Cmax 5.6 5.6 5.606 5.582
If started with the same initial conditions, as shown in Fig. En 7 7 7.003 7
10, SSE curves of both PSO-GN and the proposed GL-SEGA Rin 0 0 0 0.0009
drop quickly right after the first iteration and GL-SEGA has a Error
Rth 0 0 0.0014 0.0053
lower SSE performance compared with PSO-GN. This is Cmax 0 0 -0.006 0.018
En 0 0 -0.003 0
because the GOL-based transformation can search the solution
space more efficiently, leading to a better model matching model matching SSE is reduced. At approximately the 40th
performance. At the 10th and 20th iteration, with the help of iteration, both algorithms found the optimal solution. Compared
local search, both algorithms can reach a local optimum and the with PSO-GN, the proposed GL-SEGA obtains a lower SSE

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
8

(a)

(b) (c) Fig. 11. Lead-acid battery modeled voltage curve versus the field measurement.
Fig. 9. Li-ion battery model simulated results versus datasheet curves: (a)
voltage; (b) SOC; and (c) temperature.

SOC (%)
Fig. 12. Lead-acid battery modeled SOC curve versus the field measurement.

consisting of 24 serially connected lead-acid battery cells [27].

We assume that all battery cells are identical such that the
battery bank model can be constructed by multiplying the cell
voltage with the total cell number per string. These assumptions
allow the system thermodynamics to be captured by using a
lumped battery cell model. The battery specifications are listed
Fig. 10. SSE comparison between GL-SEGA and PSO-GN for Li-ion battery. in Table III.
TABLE VI A 24-hour field test is conducted to validate the battery
ALGORITHM EVALUATION FOR LI-ION BATTERY model performance. As shown in Fig. 11-13, the simulated
Metrics Output GL-SEGA PSO-GN voltage and SOC match the field measurements closely. The
SOC 176.61 195.02
SSE V 456.80 604.67 modeled battery bank temperature deviates the field
T 22.40 98.06 measurements slightly. That is because the accuracy of the
SOC 0.0273 0.0386 temperature measurement is not very high. Besides, the
MSE V 0.0721 0.0881
T 0.0031 0.0136
temperature sensor is placed on the surface of the battery bank,
SOC 0.9783 0.9708 making the measurement susceptible to ambient temperature
NRMSE V 0.7857 0.7656 variations. Thus, the measurement data consists of some
T 0.7656 0.5337
unpredictable uncertainties.
SOC 0.9995 0.9991
NMSE V 0.9541 0.9451 The GL-SEGA parameterization method is compared with
T 0.9451 0.7825 the PSO-GN method and algorithm results are summarized in
Fig. 14 and Table. VII. As shown in Fig. 14, in the first 10
value, showing that the proposed algorithm can escape from the
iterations, compared with PSO-GN, the SSE curve of GL-
local minimum and reach the global optimum.
SEGA drops quickly because the improved GA can efficiently
Because the exact battery parameters are unknown, the set of
traverse the solution space searching for a good parameter set
parameters attain a lower SSE/MSE or a larger NMSE/NRMSE
that can reduce the matching error. However, at 10th iteration,
is considered as a better set of parameters. From the results
PSO-GN achieves a lower SSE than GL-SEGA because PSO-
shown in Table VI, GL-SEGA achieves lower SSE/MSE for all
GN reaches a better local optimum than GL-SEGA. In the next
measurements. Thus, GL-SEGA outperforms PSO-GN in every
20 iterations, GL-SEGA keeps traversing the solution space and
aspect. Moreover, compared with the battery voltage and cell
reducing the matching SSE, and it reaches a minimum point at
temperature simulated results, the battery SOC data that has a
30th iteration. The SSE curve of PSO-GN drops slowly after 10th
higher NMSE/NRMSE indicates that the proposed battery
iteration and it reaches a minimum point at 60th iteration. As
model and parameterization method have a better performance
shown in Table. VII, compared with the PSO-GN method, the
in Li-ion battery SOC matching.
GL-SEGA achieves a lower SSE for all outputs. Moreover, the
D. Case 3: Lead Acid Battery Bank voltage data has a higher NMSE/ NRMSE indicating that the
The GL-SEGA parameterization method is tested using field proposed battery model and parameterization method have a
test data collected from an actual 48V/900Ah battery bank better performance in lead-acid battery voltage matching.

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
9

the main grid. The microgrid controller first adjusts the

35 Model
Field
microgrid frequency to match the grid frequency. Then, at
30 t = 50s, when connected to the grid, power drawn from the
25 main grid will quickly increase to pick up the loads in the
microgrid. Because the microgrid load is a constant
20
impedance load, it increases when the voltage increases
0 6 12 18 24
Time (hour) after the microgrid is reconnected to the main grid. Since
Fig. 13. Lead-acid battery bank modeled temperature curve versus field the battery SOC is greater than 45%, a default SOC setting
measurement. for allowing the battery to supply loads, the grid power will
decrease slowly to let the battery system pick up the loads.
3) After t = 230s, the PV is back online and its active power
starts to increase. The PV will supply all the loads and also
charge the battery after t = 310s.

Utility grid Composite

load

Transformer AC Lead acid

DC batteries back

Circuit
breaker AC Photovoltaic
Fig. 14. SSE comparison for Lead-acid battery. DC panels
TABLE VII
ALGORITHM EVALUATION FOR LEAD-ACID BATTERY Multi-cluster box
Metrics Output GL-SEGA PSO-GN Fig. 15. The configuration of the AC hybrid microgrid model.
SOC 43795.28 70009.08
SSE V 8260.80 10827.23
T 109763.32 112651.08
SOC 0.5069 0.8103
MSE V 0.0956 0.1253
T 1.2704 1.3038
SOC 0.8688 0.8342
NRMSE V 0.9176 0.9057
T 0.7099 0.7061 (a) (b)
SOC 0.9828 0.9725
NMSE V 0.9932 0.9911
T 0.9156 0.9136

E. Case 4: HIL Microgrid Energy Storage

After the ECM parameters are derived for the battery banks,
we test the battery model in a microgrid real-time HIL test (c) (d)
system. The HIL test system is developed based on the Fig. 16. Microgrid switching transients from islanding modes to grid-connected
microgrid prototype in Lyon which consists of an AC hybrid modes: (a) battery power comparison; (b) field measured grid power; (c) field
measured PV power; and (d) field measured load consumption.
microgrid connected to the power distribution system via a
2.4kV/230V step-up transformer. The microgrid is powered by
a lead-acid battery system and a rooftop photovoltaic (PV)
system, as shown in Fig. 15.
A field test is conducted to obtain the transient response of
the battery operation when the microgrid is switching from
islanding to grid-connected. Because the PV and battery control
schemes are designed by the inverter manufacturer and are not
Fig. 17. Comparison of the lead-acid battery voltage.
in the scope of this paper, we focused our effort on duplicating
the field test and validating whether or not the modeled battery
performance will match the actual field measurements.
The testing procedures are summarized as follows.
1) In the first 50 seconds, the microgrid is operated under the
islanding mode so the grid power is zero. Since the PV is
switched OFF, the system load is mainly supplied by the
battery bank, as shown in Fig. 16(a).
Fig. 18. Comparison of the lead-acid battery SOC.
2) Between t = 50s to t = 230s, the microgrid reconnects to

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
10

[7] Boovaragavan, Vijayasekaran, S. Harinipriya, and Venkat R.

Subramanian. "Towards real-time (milliseconds) parameter estimation of
lithium-ion batteries using reformulated physics-based models." Journal
of Power Sources 183.1 (2008): 361-365.
[8] Subramanian, Venkat R., Vijayasekaran Boovaragavan, and Vinten D.
Diwakar. "Toward real-time simulation of physics based lithium-ion
battery models." Electrochemical and Solid-State Letters 10.11 (2007):
A255-A260.
[9] Barreras, Jorge Varela, et al. "An advanced HIL simulation battery model
Fig. 19. Comparison of the lead-acid battery bank temperature. for battery management system testing." IEEE Transactions on Industry
Applications 52.6 (2016): 5086-5099.
Simulation results in Fig. 16 (a) and Fig. 17-19 show that the [10] Motapon, Souleman Njoya, et al. "A Generic electrothermal li-ion battery
model for rapid evaluation of cell temperature temporal evolution." IEEE
GL-SEGA tuned battery model can track the field battery Transactions on Industrial Electronics 64.2 (2017): 998-1008.
measurements accurately. From simulated results we obtained, [11] Fleischer, Christian, et al. "On-line adaptive battery impedance parameter
the matching errors of the modeled battery voltage, SOC, and and state estimation considering physical principles in reduced order
equivalent circuit battery models: Part 1. Requirements, critical review of
temperature are normally within 0.2V, 1% and 0.25C, methods and modeling." Journal of Power Sources 260 (2014): 276-291.
respectively. Thus, the modeled SOC is very accurate and the [12] Qiu, Xin, et al. "A Field Validated Model of a Vanadium Redox Flow
modeled battery voltage and temperature variations can track Battery for Microgrids." IEEE Trans. Smart Grid 5.4 (2014): 1592-1601.
[13] Rahman, Md Ashiqur, Sohel Anwar, and Afshin Izadian.
the battery electrical and thermal changes very well. "Electrochemical model parameter identification of a lithium-ion battery
using particle swarm optimization method." Journal of Power
VI. CONCLUSION Sources 307 (2016): 86-97.
[14] Barreras, Jorge Varela, et al. "An advanced HIL simulation battery model
In this paper, we presented the GL-SEGA, a novel hybrid for battery management system testing." IEEE Transactions on Industry
optimization-based parameterization algorithm. This approach Applications 52.6 (2016): 5086-5099.
allows modelers to use field measurements or manufactory [15] Dvorak, Dominik, et al. "A comprehensive algorithm for estimating
lithium-ion battery parameters from measurements." IEEE Transactions
datasheets to derive battery parameters that can be used to on Sustainable Energy 9.2 (2017): 771-779.
model a wide variety of battery operating conditions. The GL- [16] Jackey, Robyn A. "A simple, effective lead-acid battery modeling process
SEGA method combines the strong global searching capability for electrical system component selection." SAE Transactions (2007):
219-227.
of the heuristic improved GA with the powerful local [17] Sunny Island 6.0H / 8.0H Datasheet. [Online]. Available from: https:
exploitation capability of the deterministic TRR algorithm to //www.sma.de/en/products/battery-inverters/sunny-island-60h-80h.html
achieve a balance among robustness, accuracy and convergence [18] Zhang, Lei, et al. "A comparative study of equivalent circuit models of
ultracapacitors for electric vehicles." Journal of Power Sources 274
requirements. By applying the GOL-based transformation, the (2015): 899-906.
GL-SEGA can escape the local optimum and keep on searching [19] Conn, Andrew R., Nicholas IM Gould, and Ph L. Toint. Trust region
for the global optimum, making it more efficient compared with methods. Vol. 1. Siam, 2000.
[20] Branch, Mary Ann, Thomas F. Coleman, and Yuying Li. "A subspace,
the PSO method. In the case studies section, we compared GL- interior, and conjugate gradient method for large-scale bound-constrained
SEGA with other state-of-the-art algorithms and demonstrated minimization problems." SIAM Journal on Scientific Computing 21.1
that when applying to Li-ion and lead-acid battery model (1999): 1-23.
[21] Coleman, Thomas F., and Yuying Li. "An interior trust region approach
parameterization, GL-SEGA outperforms PSO-GN, HMPSO, for nonlinear minimization subject to bounds." SIAM Journal on
and PSO-NM in the speed of convergence and accuracy. optimization 6.2 (1996): 418-445.
[22] Zhou, Yangming, Jin-Kao Hao, and Béatrice Duval. "Opposition-based
ACKNOWLEDGMENT memetic search for the maximum diversity problem." IEEE Transactions
on Evolutionary Computation21.5 (2017): 731-745.
The authors would like to thank Mr. Stanislas DE Crevoisier [23] Wang, Hui, et al. "Space transformation search: a new evolutionary
D' Hurbache with Total S.A. for his support of this work. technique." Proceedings of the first ACM/SIGEVO Summit on Genetic
and Evolutionary Computation. ACM, 2009.
[24] Wang B, Li S E, Peng H, et al. Fractional-order modeling and parameter
REFERENCES identification for lithium-ion batteries[J]. Journal of Power Sources, 2015,
293: 151-161.
[1] Saw, Lip Huat, et al. "Computational fluid dynamic and thermal analysis
[25] Hamid, Nurul Farhana Abdul, Nasrudin Abd Rahim, and Jeyraj Selvaraj.
of Lithium-ion battery pack with air cooling." Applied Energy 177
"Solar cell parameters identification using hybrid Nelder-Mead and
(2016): 783-792.
modified particle swarm optimization." Journal of Renewable and
[2] Fotouhi, Abbas, et al. "A review on electric vehicle battery modelling:
Sustainable Energy 8.1 (2016): 015502.
From Lithium-ion toward Lithium–Sulphur." Renewable and Sustainable
[26] SimpliPhi 3.4kWh/48VDC Lithium Battery. [Online]. Available from:
Energy Reviews 56 (2016): 1008-1021.
https://simpliphipower.com/product/phi-3-5-battery/.
[3] Jackey, Robyn A. "A simple, effective lead-acid battery modeling process
[27] EnerSys 1.8kWh/2VDC Lead-acid Battery. [Online]. Available from:
for electrical system component selection." SAE Transactions (2007):
https://www.enersys.com/Components/Templates/General.aspx?id=848.
219-227.
[4] Wang, Qian, et al. "A critical review of thermal management models and
solutions of lithium-ion batteries for the development of pure electric
vehicles." Renewable and Sustainable Energy Reviews 64 (2016): 106-
Fuhong Xie (S’16) received the B.S. and M.S. degree in
128.
electrical engineering from Harbin Institute of Technology,
[5] Franke, Marcel, and Julia Kowal. "Empirical sulfation model for valve-
Harbin, China, in 2014 and 2016, respectively. Now he is
regulated lead-acid batteries under cycling operation." Journal of Power
pursuing the Ph.D. degree at the North Carolina State
Sources 380 (2018): 76-82.
University, NC, USA. His research interests are in the
[6] Li, Bingcong, et al. "Real-time energy management in microgrids with
theoretical and algorithmic studies in microgrid modeling,
reduced battery capacity requirements." IEEE Transactions on Smart
control, and real-time hardware-in-the-loop simulation.
Grid 10.2 (2017): 1928-1938.

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2019.2953718, IEEE
Transactions on Smart Grid
11

Hui Yu (S'17) received the Bachelor and Master degrees in Wente Zeng (S’09–M’15) received the B.S. and M.S. degree
electrical engineering from Huazhong University of Science in Automation from Shanghai Jiao Tong University, Shanghai,
and Technology, Wuhan, China, in 2013 and 2016, China, in 2006 and 2009, respectively and Ph.D. degree in
respectively. He is currently working toward the Ph.D. degree Electrical Engineering from North Carolina State University,
at the Future Renewable Electric Energy Delivery and Raleigh, NC, in 2013. He is currently a Research Scientist with
Management (FREEDM) Systems Center, North Carolina Total New Energies Ventures, USA. His research interests
State University, Raleigh, NC, USA. His research interests include smart grid, microgrids, energy management systems,
include power electronics converter modeling, design, and control in energy storage and electric vehicles.
microgrids and energy storage systems. He was a recipient of the Outstanding
Presentation Award of the 2018 IEEE Applied Power Electronics Conference.
Ning Lu (M’98–SM’05) received the B.S. degree in electrical
engineering from the Harbin Institute of Technology, Harbin,
Qian Long (S'17) received the B.S. degree in electrical China, in 1993, and the M.S. and Ph.D. degrees in electric
engineering from China Agricultural University, Beijing, power engineering from Rensselaer Polytechnic Institute,
China, in 2014, and the M.S. degree in electrical engineering Troy, NY, USA, in 1999 and 2002, respectively. Dr. Lu is
from North Carolina State University, Raleigh, NC, USA, in currently a professor with the Electrical and Computer
2016, where he is currently pursuing the Ph.D. degree with the Engineering Department, North Carolina State University. She
Department of Electrical and Computer Engineering. His was with Pacific Northwest National Laboratory. Her current research interests
research interests are in the areas of modeling, operation, include microgrid modeling and control, power distribution systems operation
control and optimization of hybrid energy systems. and planning, and power system data analysis.

1949-3053 (c) 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.