Anda di halaman 1dari 13

IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO.

7, JULY 2015 4181

A Linear Identi cation of Diode Models from


Single I V Characteristics of PV Panels
Li Hong Idris Lim, Member, IEEE, Zhen Ye, Jiaying Ye, Dazhi Yang, and Hui Du

Abstract This paper presents a novel approach on


diode model parameter identi cation from the I V charac-
teristics of photovoltaic panels. Other than the prevailing
methodology of solving a group of nonlinear equations
from a few points on the I V curve, the proposed method-
ology views the diode model as the equivalent output of
a dynamic system. From this new viewpoint, diode model
parameters are linked to the transfer function (after Laplace
transform) of the same dynamic system whose parameters Fig. 1. Equivalent circuit of diode models.
are then identi ed by a simple integral-based linear square.
The indoor ash test shows the accuracy and effective-
ness of the proposed method, and outdoor module testing The general mathematical description of the diode model in
shows its ability of online monitoring and diagnostics. Fig. 1 is given by
Comparisons to the methods of Lambert W function and
evolution algorithms are also included. m

Index TermsBinary search algorithm, diode model, I = IL IDi Ish


I V characteristics, linear least squares. i=1

m
V + Rs I V + Rs I
I. I NTRODUCTION = IL Ioi e ai
1 (1)
i=1
Rsh
HE currentvoltage (IV ) characteristics of photovoltaic
T (PV) cell/modules play an important role in solar industry
because it exactly re ects the cell/module performance [1].
where IL is the photocurrent proportional to the irradiance;
Ioi is the reverse saturation, ai = Ns ni kTc =q is the modi ed
Lumpedcircuit models with multiple diodes (as shown in ideality factor [6] for the ith diode (Ns is the number of cells
Fig. 1) have been broadly accepted to accurately describe the connected in series, ni is the ideality factor, k is Boltzmanns
IV characteristics [2], where diode D1 accounts for carriers constant, Tc is the cell temperature, and q is the electronic
diffusing across the PN junction and recombining in the bulk charge); Rs and Rsh are resistances in series and parallel,
or at surfaces. Diode D2 is sometimes attributed to carrier respectively. Only I and V are known variables from the
recombination by traps within the depletion region [3], or datasheet or real measurements.
recombination at an unpassivated cell edge [4]. Theoretically, Due to the inherent nonlinearity, it is not straightforward
more diodes (m > 2) can be added to the circuit in Fig. 1 to determine the model parameters (IL , Io , a, Rs , and Rsh )
to better account for distributed and localized effects in solar from the inputoutput data (IV characteristics), even for the
cells such as Auger recombination; however, their contribu simplest case of m = 1. The current methods in the literature
tions are too small as compared with D1 and D2 and can be can be divided into two categories.
negligible [5]. One category is the deterministic solution, which solves
the ve model parameters (m = 1) from the ve independent
equations. Usually, the four independent equations are chosen
from the opencircuit (OC), shortcircuit (SC), and maximum
power points (MPPs) at STC (1000 W/m2 , Tc = 25 C, AM =
Manuscript received August 3, 2014; revised October 29, 2014 and 1:5) as follows.
November 28, 2014; accepted December 16, 2014. Date of publication At SC (V = 0),
January 12, 2015; date of current version May 15, 2015.
L. H. I. Lim is with the Department of Electronic Systems, University
of Glasgow, Glasgow, G12 8QQ, U.K. (e-mail: LiHonIdris.Lim@glasgow. Rs Isc Rs Isc
ac.uk). Isc = IL Io e a 1 : (2)
Rsh
Z. Ye is with REC Cells Pte Ltd., Singapore 637312.
J. Ye, D. Yang, and H. Du are with the Solar Energy Research
Institute of Singapore (SERIS), National University of Singapore (NUS), At OC (I = 0),
Singapore 117574.
Color versions of one or more of the gures in this paper are available
online at http://ieeexplore.ieee.org.
Voc Voc
IL Io e a 1 = 0: (3)
Digital Object Identi er 10.1109/TIE.2015.2390193 Rsh
0278-0046 2015 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.
4182 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 7, JULY 2015

At MPP, The current trend is to combine the above two categories, i.e.,
Vmpp + Rs Impp employing both the nonlinear tting procedure and algebraic
Impp = IL Io e a 1 equations solving [31][33]. With a single parameter tting
procedure, numerical solutions to (2)(5) will be obtained
Vmpp + Rs Impp by empirical formulas or iterative algorithms. The drawbacks
: (4)
Rsh of the above two categories are mitigated this way. Recent
dIV Vmpp progress is reported by Laudani et al. [15], [34]. By applying
= 1 the Lambert W function [35], [36] to (1) (m = 1), the implicit
dV mpp Rs + Vmpp + Impp Rs
Io
a e a + R1
sh
format of I is converted to its equivalent explicit format as
+ Impp = 0: (5)
Rsh (IL + Io ) V
I=
There are many options for the fth independent equation: Rs + Rsh
One way is to estimate one of the ve parameters independently. a Io Rs Rsh Rsh (V + Rs ( IL + Io ) )

For example, IL can be estimated from the in uence of the W e a( Rs + Rsh )


: (6)
Rs a(Rs + Rsh )
structure parameters of a silicon solar cell on photocurrent
[7]. Io is material independent and can be explicitly related
to a solidstate parameter, i.e., the 0 K Debye temperature of The bene t of (6) over (1) is that the former is not transcen
the semiconductor [8]. a can be determined from the use of dental anymore, which makes it possible to nd solutions to
properties of the special trans function theory [9]. Rsh can (2)(5) by iterative algorithms. Laudani et al. further reduce
be approximated by the inverse of the slope at SC [10], i.e., the dimension of searching space from 5 to 2 by splitting the
dI=dV |sc 1=Rsh . Rs estimation is well summarized in model parameters into two independent unknowns (a and Rs )
[11][14]. The other way is to apply one of (2)(5) to non and three dependent unknowns (IL , Io , and Rsh ). This way, the
STC. For example, applying (3) to Tc = Tc + T (T = 0) burden of iterative searching is greatly relieved, and it becomes
gives [15] easy to get a and Rs numerically or graphically. To the best
of our knowledge, this represents the best achievement ever
3
Tc Eg

Eg ( 1 0 :0 0 0 2 6 7 7 T ) reported in the literature.
0 = IL + T T Io e kTc kTc
This paper opens a new angle to view the diode model from
Tc
the systems perspective. In fact, one of the biggest applications
Voc + T T Tc
Voc + T T of the Lambert W function is to solve differential equations,
e a Tc
1
Rsh which is directly linked to the representation of a linear system
in the time domain. For example, the rstorder linear system
where Eg = 1:17 4:73 104 Tc2 =(Tc + 636) is the band can be described as T dy(t)=dt + y(t) = u(t), whose unit ramp
gap energy; T and T are the temperature coef cients of SC response, i.e., y(t) = t + T (et=T 1), has the same format
current and OC voltage, respectively. as (1). This motivates us that the IV curve governed by (1)
No matter what the fth equation is, if the approximated pa can be viewed as the output of some linear system, and the
rameter is signi cantly different from the real value, it may lead model parameters can be linked to the parameters of a linear
to a singular solution to the rest of four parameters [16]. Even differential equation, which is then determined by system iden
if there is no approximation in the fth equation, there are no ti cation methods available in the literature [37], where linear
analytical solutions available due to the inherent nonlinearity. least squares is good enough to yield satisfactory solutions.
Usually, partial linearization has to be made to yield empirical This paper is organized as follows. Section II describes the
formulas [17][21], which is a tradeoff between simplicity and way to transform the static IV curve to the dynamic linear
accuracy. system output. Integralbased system identi cation methods
The other category is the optimal solution, which employs and the linear least squares algorithm are then proposed in
nonlinear tting procedures based on the minimization of Section III. Examples of the indoor ash test and applications
deviations between modeled and measured IV curves, in of outdoor module testing (OMT) are given in Section IV
accordance with some metric function (usually least squares) to illustrate the accuracy and effectiveness of the proposed
[22][25]. Iterative searching algorithms are usually used [26], method. Comparison with the existing methods is demonstrated
[27], but their convergence and accuracy heavily depend on in Section V. Section VI draws the conclusion.
the initial values and are easily trapped in the local optima.
For different initial value guess, such approaches can result in
widely different parameter sets, all leading to satisfactory curve II. DYNAMIC S YSTEM F ORMULATION
tting [28]. Although a good match between estimation and
A. One-Diode Model
measured data can be obtained, there is no guarantee that the
estimated IV curve would pass the SC, OC, and MPP points. Recall the IV curve described by (1) with m = 1. Let y = I
To achieve the global optimum, evolution algorithms such as and x = V + Rs I, (1) then becomes
differential evolution (DE) and genetic algorithm (GA) have to
be used [29], [30]. However, they are too complicated to be x x
y = IL + Io Io e a : (7)
implemented as online calculation. Rsh
LIM et al.: IDENTIFICATION OF DIODE MODELS FROM SINGLE IV CHRACTERISTICS OF PV PANELS 4183

Taking differential once on both sides of (7) gives B. Multidiode Model

dy Io x 1 Similarly, by letting y = I and x = V + Rs I in (1), it yields


= ea : (8)
dx a Rsh m m
x x
y = IL + I oi Ioi e ai : (15)
Differentiating once more for (8) gives i=1 i=1
Rsh

d2 y Io x Taking differential once on both sides of (15) gives


2
= 2 ea : (9)
dx a m
dy Ioi ax 1
= e i : (16)
Eliminating ex=a from (8) and (9) gives dx i=1
a i R sh

d2 y dy 1 Differentiating (16) for k times, k = 1; 2; : : : ; m, yields


a = : (10)
dx2 dx Rsh m
Ioi ax
y(k+1) (x) = e i (17)
Let t = x and u(t) 1, (10) is equivalent to i=1
ak+1
i

d2 y(t) dy(t) u(t) where y (k) (x) = dk y=dxk . Rewrite (17) in matrix format as
a 2
= (11)
dt dt Rsh x
I
y (2) (x) a1
1 a1
2 a1
m
ao11 e a1
which is a standard differential equation representation of a y (3) (x) a2 a2 a2 I x
1 2 m ao22 e a2
secondorder linear system. t is the time, and u(t) and .. = .. .. .. .. :
. . . . . ..
y(t) are the system input and output, respectively. Since .
u(t) 1, y(t) is the unit step response of the system in the y (m+1) (x) am
1 am
2 am
m
I m ax
aom e m
time domain. Take Laplace transform, i.e., F (s) = L[f (t)] = B A
st
0
e f (t)dt, on both sides of (11), then
Since ak = 0, A is a Vandermonde matrix with det(A) = 0;
2 U (s) hence, A1 exists, and
a s Y (s)sy(0)y (0) [sY (s)y(0)] = : (12)
Rsh T
Io1 ax Io x Io x
e 1 ; 2 e a2 ; : : : m e am = A1 B (18)
Utilize sU (s) = 1, and (12) is equivalent to a1 a2 am

a s2 Y (s) s2 U (s)y(0) sU (s)y (0) where A1 = [i;j ] Rmm with

1 (1)j1 a1 1
k1 akn j
[sY (s) sU (s)y(0)] = U (s): 1 k1 < <kn j n
Rsh k1 ; ;kn j = i
i;j = : (19)
a1
i a1 1
k ai

It follows from (7) that y(0) = IL , y (0) = Io =a 1=Rsh ; 1 k n
k= i
hence, the transfer function from Y (s) to U (s) is
Substituting (18) into (16) yields
2 1
Y (s) ay(0)s +[ay (0)y(0)] s+ Rsh m m
G(s) := = 1
U (s) as2 s y (1) (x) i;j y (j+1) (x) = : (20)
j=1 i=1
Rsh
aIL s2 Io + Rash +IL s+ R1sh
= : (13) Let t = x and u(t) 1, (20) becomes the differential equation
as2 s
representation of an mthorder dynamic system, i.e.,
The corresponding timedomain differential equation is m m
u(t)
y (1) (t) i;j y (j+1) (t) = : (21)
2
d y(t) dy(t) d u(t) 2
j=1 i=1
Rsh
a = aIL
dt2 dt dt2
Taking Laplace transform for both sides of (21) yields
a du(t) u(t)
IL + Io + + : (14)
Rsh dt Rsh sY (s) y(0)
m m j+1
It should be noted that (11) is different from (14) because j+1
i;j s Y (s) sk1 y (j+1k) (0)
of the nonzero initial conditions. In other words, (14) is the
j=1 i=1 k=1
description of the same system of (11) but with zero initial
conditions. This will facilitate the calculation of the integral U (s)
= : (22)
based identi cation proposed in Section III. Rsh
4184 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 7, JULY 2015

It follows from (15)(17) that y(0) = IL , y (1) (0) = m i=1 Ioi =


A. One-Diode Model
ai 1=Rsh , y (k+1) (0) = m I
i=1 io =a k+1
i for k = 1; 2; : : : ; m. Applying (27) to (14) for T1 = 0, T2 = t, and n = 2 gives
Since sU (s) = 1, (22) becomes
m m a
ay(t) aIL u(t) + IL + Io +
sY (s) IL sU (s) i;j Rsh
j=1 i=1
(1) (2) (1)
1
u( ) u( ) = y( ): (28)
sj+1 Y (s) U (s) Rsh
[0;t] [0;t] [0;t]

j m (1)
Ioi sj U (s) Let (t) = y( ), = [a; aIL ; (IL +Io +(a=Rsh)); (1=Rsh)]T ,
sk j+1k
+IL sj+1 = : [0;t]

i=1 ai Rsh Rsh (1) (2)


(t) = [y(t); u(t); [0;t] u( ); [0;t] u( )]T , and (28) can be
k=1
rewritten as the matrix format of T (t) = (t). Note that the
The transfer function is G(s) = Y (s)=U (s) = N=D, where
matrix format holds for any ti [0; t], i = 1; 2; : : : ; N , where
m m
N is the number of data samples on the IV curve. This
D= i;j sj+1 s; actually casts an equation group of = with = [(t1 );
j=1 i=1
(t2 ); : : : ; (tN )]T and = [(t1 ); (t2 ); : : : ; (tN )]T . If
1
m m T is nonsingular, the linear least squares solution for is
N= IL s + i;j given by
Rsh j=1 i=1

j m = (T )1 T (29)
j+1 sj k Ioi
IL s s :
Rsh
k=1 i=1 aj+1k
i
which will minimize the square error of ( )T ( ).
Once is determined from (29), the parameters of the one
The corresponding timedomain differential equation with zero diode model can be obtained by a = 1 , IL = 2 =1 , Io =
initial condition is 3 2 =1 1 4 , Rsh = 1=4 .
m+1 y (m+1) (t) + + 2 y(2) (t) y (1) (t)
B. Multidiode Model
(m+1) (1) u(t)
= m+1 u (t) + + 1 u (t) + (23) Apply (27) to (23) for T1 = 0, T2 = t, and n = m + 1, then
Rsh
(m1) (m)
where for j = 1; 2; : : : ; m, we have
m
m+1 y(t) + + 2 y( ) y( )
j+1 = i;j (24) [0;t] [0;t]
i=1
(m) (m+1)
m+1 = m+1 IL (25) 1
m m
= m+1 u(t) + + 1 u( ) + u( ):
j+1 k+1 Ioi Rsh
j = j IL ; [0;t] [0;t]
Rsh
k=j i=1
ak+1j
i
(m)
Let (t) = y(), = [m+1 ; : : : ; 2 ; m+1 ; : : : ; 1 ; 1=Rsh]T ,
(1 = 1): (26) [0;t]
(m1) (m+1)
(t) = [y(t); : : : ; [o;t] y( ); u(t); : : : ; [0;t] u( )]T ,
In general, by introducing a virtual time of t = x, the static and (t) R(2m+2)1 , we have T (t) = (t). For ti [0; t],
relationship between two variables y and x can be regarded as i = 1; 2; : : : ; N , the equation group can be described by
dynamics from the linear system governed by (23). Once i = with = [(t1 ); (t2 ); : : : ; (tN )]T and = [(t1 );
and i are determined from system identi cation, diode model (t2 ); : : : ; (tN )]T . If T is nonsingular, the least squares
parameters IL , Ioi , ai , and Rsh can be linearly solved from solution for will be
(24)(26).
= (T )1 T : (30)
III. I NTEGRAL -B ASED L INEAR I DENTIFICATION
Once is determined from (30), Rsh = 1=2m+2 is imme
For an integer n 1, de ne the multiple integral as [37] diately derived. It follows from (25) that IL = m+1 =m+1 =
(n) T2 n 2
m+1 =1 . ai ; i = 1; 2; : : : ; m, will be derived in the following
way. Rewriting (24) in matrix format gives
f ( ) = f (1 )d1 d2 : : : dn : (27)
[T1 ;T2 ] T1 T1 T1 [2 ; : : : ; m+1 ] = [1; : : : ; 1] A1 :
n m
LIM et al.: IDENTIFICATION OF DIODE MODELS FROM SINGLE IV CHRACTERISTICS OF PV PANELS 4185

Rightmultiplying A for both sides gives C. Nonsingularity of T

a1 a1 The existence of the linear least squares solution by (29) and


1 m
.. .. .. (30) depends on the nonsingularity of T , which is shown by
[2 ; : : : ; m+1 ] . . . = [1; : : : ; 1] the following lemma.
am
1 am
m m 1) Lemma 1: T is nonsingular if ai = aj for i = j,
i; j = 1; 2; : : : ; m, and the sampling number N 2m + 2.
which implies that 1=ai ; i = 1; 2; : : : ; m, are the roots of the 2) Proof: See the Appendix.
following characteristic equation:
D. Calculation of Multiple Integrals
m+1 m + m m1 + + 2 1 = 0: (31)
In practice, the integral shown as (27) is numerically esti
Solve (31) for i , and ai = 1=i ; i = 1; 2; : : : ; m. Ioi , i = mated by rectangular or trapezoidal integration. For example,
1; 2; : : : ; m, will be derived as follows. Equation (26) can be suppose there are N samples at t1 ; t2 ; : : : ; tN , the rectangular
rewritten as integration gives
(1) ti
m m i1
j+1 k+1 f ( ) = f (1 )d1 f (k)(tk+1 tk ) := f1 (i);
j = j IL Ioi k+1j
:
Rsh i=1 ai k=1
k=j [t1 ;ti ] t1

(2)
Rewrite further as matrix format i1
f ( ) f1 (k)(tk+1 tk ) := f2 (i);
m m m
k+ 1 k+ 1 k+ 1 k=1
ak ak
ak
[t1 ;ti ]
1 2 m
k=1 k=1 k=1
m m m ..
k+ 1 k+ 1
k+ 1 .
ak
1
1
ak
2
1
ak
m
1
k=2 k=2 k=2
.. .. .. .. (n)
i1
. . . .
m
k+ 1
m
k+ 1
m
k+ 1
f ( ) fn1 (k)(tk+1 tk ) := fn (i)
ak+ 1 m
ak+ 1 m ak+ 1 m k=1
k=m 1 k=m 2 k=m m [t1 ;ti ]


for i = 1; 2; : : : ; N . The more number of samples, i.e., fi , the
more accurate the estimation to the multiple integrals will be.
I o1 1 + IL + Rsh
2

I o2 2 2 IL + Rsh
3

.. = .. : E. Determination of Rs
. .
Iom m m IL + m+ 1
Rsh
To calculate from (29) or (30), and must be known.
As both of them are integrals to t, t must be known as well.

Since t = V + Rs I, Rs must be determined before applying
m integrals. It is clear that if Rs is bigger than its real value, t will
Note from (31) that k+1 =aki = 1 for i = 1; 2; : : : ; m,
k=1 increase so that the whole IV curve will move to the right,
can be simpli ed as and the error between the real and estimated IV curves will be
positive; if Rs decreases, the whole IV curve will move to the
1 1 1
left, and the error between the real and estimated IV curves
a1 a2 am
= .. .. .. .. will be negative. Thus, Rs can be used as a tuning parameter
. . . . such that the rootmeansquare error (RMSE) is minimized.
am1
1 am1
2 am1
m It derives from (1) that
1 1
dI
= Rs + m Voc
> Rs
Ioi 1
0 0 dV oc
ai
e ai
+ Rsh
i=1
2 2
.. .. ..
. . . : which implies the upper bound of Rs , i.e., Rsupp =
m1 m1 1=(dI=dV )|oc . The lower bound of Rs can be zero at rst,
k+1 am1k k+1 am1k
k=1
1
k=1
m i.e., Rslow = 0. With such a band of Rs [Rslow ; Rsupp ], the
binary search algorithm is applied to determine Rs in the
This implies that after elementary row operations, is following way.
similar to , which is a Vandermonde matrix with Step 1: Arbitrarily choose Rs from [Rslow ; Rupp s ] and calcu
det( ) = 0. Therefore, 1 exists ( is full rank), and late ai , IL , Ioi , and Rsh from the proposed linear
[Io1 ; Io2 ; : : : ; Iom ]T = 1 . least squares (29) or (30).
4186 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 7, JULY 2015

Step 2: Calculate from (1) that The remaining procedures are the same as aforementioned.
(1) (2)
Let (t) = [y(t); u(t); [0;t] u( ); [0;t] u( )]T , (t) =
m (1)
V + Rs I V + Rs I y( ), and
y(t) = IL Ioi e ai
1 [0;t]
Rsh
i=1 a
V oc
aVoc
aIL aIo (e a 1) Rsh
= oc a
and RM SE = N IL + Io VR
i=1 [y(ti ) y(ti )] =N .
2 sh
1
Step 3: Calculate ERR = N i=1 [y(ti ) y(ti )]. If ERR >
Rsh

0, adjust Rs = (Rs + Rlows )=2. Otherwise, adjust the linear least squares solution is = (T )1 T
Rs = (Rs + Rsupp )=2. with = [(t1 ); (t2 ); : : : ; (tN )]T and =
Step 4: Update Rsupp and Rslow according to the sign of [(t1 ); (t2 ); : : : ; (tN )]T . Once is determined, the
ERR. If ERR > 0, Rupps = Rs ; otherwise, Rslow = parameters of the onediode model are obtained by a = 1 ,
Rs . IL = 2 =1 + (3 2 =1 1 4 )(1 eVoc =1 ) + Voc 4 ,
Step 5: If RM SE is less than some tolerance or the iterative Io = (3 2 =1 1 4 )=eVoc =1 , and Rsh = 1=4 .
cycle reaches some preset number, stop the search. In case of the multidiode model, with the same transform of
Otherwise, update Rupp
s and Rlow
s according to the x = Voc x, (15) becomes
sign of ERR and go back to Step 2.
m m
Voc
ax Voc x
y = IL + Ioi Ioi e ai
e i + : (32)
i=1 i=1
Rsh Rsh
F. Robustness Enhancement
From the viewpoint of control theory, the transfer function Let ai = ai , IL = IL + mi=1 Ioi (1 e
Voc =ai
) Voc =Rsh ,
(13) has a pole of s = 1=a > 0, which implies that system Io = Io e oc i , Rsh = Rsh , and (32) is equivalent to
i
V =a
i

(14) is unstable. This is also true for the general case of the m m
x x
multidiode model. Identi cation for an unstable system is not y = IL + Ioi Ioi e ai
preferred because the convergence of the proposed algorithm i=1 i=1
Rsh
might be sensitive to the accuracy of the integral calculation
in such a case. To improve the robustness of the proposed which has the same format as (15). This means that all
algorithm, V is introduced to yield a stable system. the derivations aforementioned are applicable to the param
In case of the onediode model, let V = Voc V , 0 V eter set {ai ; IL ; Ioi ; Rsh }. Once they are determined, the
Voc , and x = V Rs I , thus x = V + Rs I = Voc (V parameter set {ai ; IL ; Ioi ; Rsh } is immediately derived by
Rs I) = Voc x. It follows from (7)(9) that ai = ai , Rsh = Rsh , Ioi = Ioi eVoc =ai , and IL = IL
m
i=1 Ioi (1 e
Voc =ai
) + Voc =Rsh .
Voc Voc x x
y = IL + Io Io e a e a + ;
Rsh Rsh IV. VALIDATION

dy Io Voc x 1 A. Indoor Flash Test


= e a e a + ;
dx a Rsh The IV characteristics of fullsized commercial modules
were measured indoor by a pulsed solar simulator (PASAN
d2 y Io Voc x IIIB) with a constant illumination intensity plateau of about
= 2 e a e a :
dx2 a 12 ms used. The data acquisition, which requires about 10 ms,
occurs during the plateau period, whereby the light intensity
Let t = x and u(t) 1, by eliminating ex=a , it gives varies by less than 1%. The intensity of the solar simulator
is calibrated with a cSi reference cell certi ed by Fraunhofer
ISE. The overall uncertainty of module power measurement is
d2 y(t) dy(t) u(t)
a + = : within 2%.
dt2 dt Rsh Example 1 (c-Si Modules): The IV characteristic of a
crystalline PV module from the indoor ash test under STC
The corresponding transfer function is (1000 W/m2 , 25 C) is shown in Fig. 2. Both onediode and
twodiode models are considered for this case study.
Y (s) ay(0)s2 + [ay (0) + y(0)] s + 1 1) One-Diode Model: First, use the last ten points at OC
Rsh
G(s) = = to derive a linear tting: I = kV + p, where k = 0:9131.
U (s) as2 + s
Rupp
s 1=k = 1:0952. Rslow = 0. Arbitrarily choose Rs
[Rs ; Rsupp ], e.g., Rs = 1:0952, and follow the proposed
low

where y(0) = IL Io (eVoc =a 1) Voc =Rsh , and y (0) = integralbased linear identi cation presented in Section IIIA,
Io eVoc =a =a + 1=Rsh . This way, the unstable pole s = 1=a > 0 Rs converges to Rs = 0:655 after about 30 steps with the
becomes stable as s = 1=a < 0. proposed binary search, as shown in Fig. 3. Multiple integrals
LIM et al.: IDENTIFICATION OF DIODE MODELS FROM SINGLE IV CHRACTERISTICS OF PV PANELS 4187

Equations (34) and (35) in matrix format are


I I
d2 y ao21 ao22 Voc
e Voca1
x
dx2 = 1 2 x :
d3 y Io1 Io2
dx3 a31 a32
e a2

Thus,

I I 1
Voc
e Voca1
x
ao21 ao22 d2 y
x = 1 2 dx2
Io1 Io2 d3 y
e a2 a31 a32 dx3

a31 a31 a2
d2 y
Io1 (a2 a1 ) Io1 (a2 a1 ) dx2
= a3 a a32 d3 y
:
Fig. 2. Accuracy of the proposed method for the c-Si module. Io (a22a1 ) Io (a1 2 a
2 2 1) dx3

Substituting it into (33) yields

d3 y(t) d2 y(t) dy(t) u(t)


a1 a2 + (a 1 + a2 ) + = (36)
dt3 dt2 dt Rsh

where t = x and u(t) 1. After Laplace transform, (36) be


comes

a1 a2 s3 Y (s)y (0)sy (0)s2 y(0) +(a1 + a2 )


U (s)
s2 Y (s)y (0)sy(0) +[sY (s)y(0)] = (37)
Rsh
Fig. 3. Convergence of Rs and RM SE for the c-Si module.
where
from (27) are estimated by the numerical integration presented Voc Voc Voc
in Section IIID. It follows from (29) that 1 = 1:9891, 2 = y(0) = IL +Io1 1e a1 +Io2 1e a2 (38)
9:8295, 3 = 4:9434, 4 = 8:9631 104 . Thus, a = 1 = Rsh
1:9891 V, IL = 2 =1 = 4:9416 A, Io = 3 2 =1 1 4 =
Io1 Vaoc Io Voc 1
4:1785 109 A, and Rsh = 1=4 = 1:1157 103 . y (0) = e 1 + 2 e a2 + (39)
a1 a2 Rsh
Fig. 2 also shows the comparison between the IV curves
from the real measurement and the onecode model, where the
Io1 Vaoc Io2 Vaoc
average absolute error E = 1=N N y (0) =
2
i=1 |ERR| = 0:0085. The (40)
e 1 e 2 :
a21 a2
RM SE is shown in Fig. 3, which nally converges to 1.67%
after 35 steps with Tol = 2%.
Utilize sU (s) = 1, and (37) is equivalent to
2) Two-Diode Model: It is clear in Fig. 2 that the one
diode model is good enough to represent the whole IV curve a1 a2 s3 Y (s) + (a1 + a2 )s2 Y (s) a1 a2 y(0)s3 U (s)
accurately. This implies that if the twodiode model is applied,
Io2 0, which will cause a singular matrix in the identi cation U (s)
in Section IIIB. To avoid such a potential problem, robustness [a1 a2 y (0) + (a1 + a2 )y(0)] s2 U (s)
Rsh
enhancement discussed in Section IIIF will be applied. With
m = 2, (32) becomes
[a1 a2 y (0)+(a1 +a2 )y (0)+y(0)] sU (s) = sY (s):
Voc x Voc x Voc x
y = IL + Io1 1 e a1 + Io2 1 e a2
Rsh Therefore, the differential equation representation with zero
initial conditions is
where x = V Rs I, and V = Voc V . Furthermore, its mul
tiple differentials are d3 y(t) d2 y(t) d3 u(t)
a1 a2 +(a 1 +a 2 ) a 1 a 2 y(0)
dy Io Voc x Io Voc x 1 dt3 dt2 dt3
= 1 e a1 + 2 e a2 + (33)
dx a1 a2 Rsh d2 u(t) u(t)
d2 y Io Voc x Io Voc x [a1 a2 y (0)+(a1 +a2 )y(0)]
= 21 e a1 22 e a2 (34) dt2 Rsh
dx 2 a1 a2
d3 y Io1 Voca x Io2 Voca x du(t) dy(t)
= 3e 1 + 3e 2 : (35) [a1 a2 y (0)+(a1 +a2 )y (0)+y(0)] = : (41)
dx3 a1 a2 dt dt
4188 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 7, JULY 2015

Apply triple integral (27) (with n = 3) to (41), we have


(1)

a1 a2 y(t) + (a1 + a2 ) y( ) a1 a2 y(0)u(t)


[0;t]
(1)

[a1 a2 y (0) + (a1 + a2 )y(0)] u( )
[0;t]
(2)

[a1 a2 y (0) + (a1 + a2 )y (0) + y(0)] u( )
[0;t]
(3) (2)
1
u( ) = y( ): (42)
Rsh Fig. 4. Environmental factors of a typical day in SERIS OMT testbed.
[0;t] [0;t]

(2) (1) (1) The average absolute error E = 0:0080 and RM SE =


Let (t) = [0;t] y(), (t) = [y(t); [0;t] y(); u(t); [0;t] u();
(2) (3)
1:35%, both of which are slightly reduced as compared with the
[0;t]
u( ); [0;t]
u( )]T , and onediode model result. As expected, Io2 is indeed extremely
1 a1 a2 close to zero, whereas other parameters are comparable to their
2 a1 + a2 counterparts in the onediode model result.
3 1 y(0) It should be highlighted that the diode model parameters
:= = (43) derived from the indoor ash test are not constant. In fact,
4 1 y (0) + 2 y(0)
5 1 y (0) + 2 y (0) + y(0) they vary with temperature and solar radiation. Therefore, it
6 1
Rsh
is necessary to check the online computability of the proposed
method for PV modules under nonconstant environment, which
then (42) can be rewritten in matrix format of (t)T = (t). is demonstrated by the OMT as follows.
The linear least squares solution to
! is given by (30). Immedi
ately, Rsh = 1=6 , a1;2 = (2 22 41 )=2, and
B. OMT
3 1 0 0 y(0)
4 = 2 1 0 y (0) : OMT is usually carried out by many PV panel manufactur
5 1 2 1 y (0) ers and solar research institutes for the module performance
evaluation under the real operating environments. DC param
Therefore, eters including full IV curves, Voc , Isc , Vmpp , Impp , and
1 Pmpp , together with module temperature, are measured and
y(0) 1 0 0 3 logged every minute. Environmental parameters including in
y (0) = 2 1 0 4 : plane solar irradiance Gsi , ambient temperature Tamb , mod
y (0) 1 2 1 5 ule temperature Tmod , wind speed, and wind direction are
It follows from (38)(40) that simultaneously logged with the dc parameters. Between IV
Voc Voc
measurements, electrical energy is maintained at the module
Voc 1 1e a1
1e a2 MPP. The uncertainty of all electrical measured parameters is
y(0) + R Voc Voc IL
sh
e a1 e a2 within 0:1% for full scale. With these IV data in time series,
y (0) R1sh
= 0 I o1 :
a1
Voc
a2
Voc
the diode model parameters can be identi ed online by the
y (0) e a1 e a2 I o2 proposed method and correlated to the environmental factors
0 a21
a22
such as irradiance,temperature, etc.
Thus, Fig. 4 shows the time series of Gsi , Tamb , and Tmod on a
Voc Voc 1
typical day from the OMT testbed of the Solar Energy Research
1 1e a1
1 e a2 y(0) + RVoc Institute of Singapore (SERIS). The plot is centered around
IL Voc Voc

0 e a1 e a2 sh solar noon, which was at 13:10 on August 5, 2010.


Io1 = y (0) R1sh :
a1
Voc
a2
Voc By applying the proposed method in Section III, the time
Io2 e a1 e a2 y (0) varying onediode model parameters IL , Io , a, Rs , and Rsh for
0 a21
a22
the same day are identi ed, as shown in Fig. 5. The variation
This way, with the same IV characteristics data as of the identi ed parameters re ects the dynamics of the PV
shown in Fig. 2, we got 1 = 0:6849, 2 = 2:2356, 3 = module under different environmental conditions, which cannot
0:0247, 4 = 3:3348, 5 = 4:9034, 6 = 0:0010. The two be seen from the static IV curves.
diode model parameters are identi ed as a1 = 1:8691 V, a2 = The relationships between the identi ed parameters and
0:3664 V, Io1 = 1:5168 1010 A, Io2 = 7:9060 1054 A, the environmental operating conditions are further illustrated
IL = 4:9480 A, Rsh = 955:1229 , and Rs = 0:6845 . in Figs. 69. A proportional relationship between IL and
LIM et al.: IDENTIFICATION OF DIODE MODELS FROM SINGLE IV CHRACTERISTICS OF PV PANELS 4189

Fig. 8. Relationship between a and Gsi .

Fig. 5. Identi ed one-diode model parameters.

Fig. 9. Relationship between Rs , Rs h, and Gsi .

Fig. 6. Proportional relationship between IL and Gsi .

Fig. 10. RMSE and burden of online calculation.

where Eg is the band gap of silicon, and B is a temperature


independent constant [13]. Fig. 8 illustrates that a generally
decreases with increasing irradiance for Gsi < 300 W/m2 and
increases beyond that, which is as reported in [38]. When ir
radiance decreases in Fig. 9, the series resistance Rs decreases,
and the shunt resistance Rsh increases, which is consistent with
previous reported results [39]. The decrease in Rs is due to the
decreased thermal loss (I 2 Rs ) with decreasing irradiance.
The RMSE of the proposed algorithm in the OMT case
Fig. 7. Relationship between Io and Tmod .
is shown in Fig. 10, where the burden of online calculation
irradiance intensity is observed in Fig. 6. It is also apparent in for convergence (iterative steps for Rs until Tol or maximum
Fig. 7 that Io generally shows an increasing trend with rising cycle is achieved) is presented as well. Among 600 plus IV
module temperature. This also agrees with the theoretical tem scans during the day, there are only three cases with the RMSE
perature dependence of Io , as given by Io = BT 3 eEg =(kT ) , exceeding the preset 1% Tol when the maximum number (100)
4190 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 7, JULY 2015

TABLE I
S OLUTION C OMPARISON FOR S OLAR M ODULE

of steps is reached. Even for these three cases, the RMSE TABLE II
RMSE W ITH D IFFERENT DATA S AMPLES (M ODULE )
is still below 1.5%. The iterative steps are very stable, and
they are usually less than 30. This indicates that the online
calculation burden of the proposed algorithm is low, and the
identi cation can be done by an industrial PC locally between
two consecutive IV scans (1 min in our case).

V. C OMPARISON W ITH OTHER M ETHODS


As mentioned in Section I, the recent progress in the parame
ter identi cation of the diode model is mainly from the work of
Laudani et al. [15], [34] and the evolutionary algorithms [29],
[30]. Here, the comparison of the proposed method with these
two approaches is discussed. smaller the RMSE. When data samples increased to 100, the
RMSE for the experimental data is already better than the 1A/B
solutions of Laudani et al. and all the other results compared
A. Method of Laudani et al. in [34].
In [34], two data sets of IV curves (26 points) are presented, The result comparison for the solar cell IV data in [34]
which are initially proposed in [40] and are commonly used to is shown in Table III. The RMSE of the proposed method is
test the effectiveness of the extraction algorithms. One refers to smaller than the results of 2A/C of Laudani et al., and only
a solar module (PhotowattPWP 201) at 45 C, and the other slightly bigger than 2B/D of Laudani et al.. When data samples
refers to a solar cell (cSi) at 33 C. increased to 100, the proposed method already outperformed
The onediode model parameters IL , Io , and Rsh are proved 2B of Laudani et al. as shown in Table II.
to be functions of Rs and a. Hence, searching in the 2D In general, the method of Laudani et al. has many bene ts in
parameter space of Rs and a with the constrained conditions two aspects: 1) it utilizes the Lambert W function to convert a
of (2), (3), and (5) yields Solution A, whereas that with the nonconcave optimal problem into a concave optimal problem;
constrained conditions of (2)(4) yields Solution B. These two 2) it utilizes reduced forms to decrease the dimension of the
solutions are then ne tuned as the initial values of some parameter space from ve to two. It can deal with the IV data
nonlinear least squares for the experimental data, which yields from the datasheet (points at SC, OC, and MPP) or experiment
Solutions C and D, respectively. (full IV curve), and in most of cases, it yields the best results
With the solar module IV data in [34], the comparison of in terms of RMSE and/or MAE. The de ciencies of the method
the solutions of the onediode model by the proposed method of Laudani et al. may be that it 1) has no unique solutions,
and the method of Laudani et al. is shown in Table I, where 2) is inapplicable to the multidiode model (m > 1) parameter
MAE is the mean absolute error, and Step is the number of identi cation due to the limitations of the Lambert W function,
iterative searching cycles before convergence. It is clear that the and 3) is not easy to be implemented and unsuitable for online
proposed method is very close to the model parameter values parameter identi cation.
in the results by Laudani et al.. Although the error is slightly The proposed method further reduces the dimension of the
bigger, the number of iteration steps is less. parameter space to one. It uses linear squares other than non
The error mainly arises from the numerical integrations pre linear optimal algorithms to derive diode model parameters;
sented in Section IIID and the few IV data samples available hence, the drawbacks of nonlinear algorithms are avoided. It
(26 points only). If more data samples on the IV curve are can also be used for the multidiode model and simple enough
known, the error of the proposed method will be reduced. To to be implemented as online calculation. The de ciency is that
illustrate this point, model parameters from the 1D solution of it requires knowledge of the full IV curve data.
Laudani et al. was used to reproduce the whole IV curve with
the help of (6). The number of samples is selected to be 50, 100,
B. Evolution Algorithms
and 200. Based on such samples on the IV curve derived from
the 1D solution of Laudani et al., the RMSE of the proposed As mentioned in Section I, evolution algorithms are very suit
method to the whole IV curve and the experimental data are able for the search of a global optimal solution. Recently, two
shown in Table II. As expected, the more the data samples, the types of evolution algorithms using DE [29] and GA [30] have
LIM et al.: IDENTIFICATION OF DIODE MODELS FROM SINGLE IV CHRACTERISTICS OF PV PANELS 4191

TABLE III
S OLUTION C OMPARISON FOR S OLAR C ELL

TABLE IV VI. C ONCLUSION


S OLUTION C OMPARISON W ITH E VOLUTION A LGORITHMS
A novel method has been proposed in this paper to identify
all onediode model parameters of PV panels from a single
IV curve. By utilizing the mapping of the transfer function,
the nonlinear tting problem is converted equivalently into
linear system identi cation. Correspondingly, the dimension
of the parameter space is reduced from ve to one. Indoor
and outdoor module testing shows its effectiveness and online
computability, and its accuracy is also comparable to or better
than the best results from the literature.

A PPENDIX
TABLE V
C OMPARISON OF T WO D IODE M ODELS
P ROOF OF L EMMA 1
Consider the general case of the multidiode model with
= [(t1 ); (t2 ); : : : ; (tN )]T := [1 ; 2 ];
(1) (m1)
y(t1 ) [0;t1 ] y( ) [0;t1 ] y( )
(1) (m1)
y(t2 ) [0;t2 ] y( ) [0;t2 ] y( )
1 = := [i;j ];
.. .. .. ..
. . . .
(1) (m1)
y(tN ) [0;tN ] y( ) [0;tN ] y( )
(1) (m+1)
u(t1 ) [0;t1 ] u( ) [0;t1 ] u( )
(1) (m+1)
u(t2 ) [0;t2 ] u( ) [0;t2 ] u( )
yielded good results for diode model parameter identi cation. 2 =
.. .. ..
=: [i;l ]:
..
Hence, it is worthy to compare the proposed method with them. . . . .
(1) (m+1)
Since no full IV curve data are provided in [29] and [30], u(tN ) [0;tN ] u( ) [0;tN ] u( )
we do the comparison in an indirect way as follows. First, use
the identi ed parameters (IL , Io , a, Rs , and Rsh ) to reconstruct Recall from (15) that
m m
the IV curve by (6); second, use that IV curve data to t t
identify diode model parameters with the proposed method. y(t) = IL + Ioi Ioi e ai
Rsh
Since DE and GA are applied to derive a, Rs , and Rsh only i=1 i=1

(IL and Io are derived by formulas in [6] and [32]), we only and u(t) 1 by the de nition. For i = 1; 2; : : : ; N ,
compare the results of a, Rs , and Rsh . Table IV shows the (j1)
results of a, Rs , and Rsh from the proposed method and IL + m i=1 Ioi j1 tji
DE/GA. It is clear that the differences in between are very i;j = y( ) = ti
(j 1)! j!Rsh
minor. [0;ti ]
The result of the twodiode model for the aforementioned j2 m
tki
j
ti

Kyocera module (KyoceraKC200GT) was also reported in + Iol ajk1


l Iok aj1
k e
ak
;
k!
[30]. It is interesting to compare this result with ours. Carefully k=0 l=1 k=1
(l1)
looking at the comparison shown in Table V, the GA algorithm 1
gives comparable Io1 and Io2 (both in 109 A). a1 and a2 are i;l = u( ) = tli
j!
also near each other. Ignoring the differences between them, [0;ti ]
the twodiode model can be combined as one. This implies
that the GA algorithm actually gives a result of the onediode where j = 1; 2; : : : ; m, and l = 1; 2; : : : ; m + 2. After elemen
model but mathematically divides it into a twodiode format tary column operations for , 1 1 := [i;j ] with
with no physical meaning. That is a common issue for the j
ti
global optimization algorithm such as DE and GA, whereas the i;j = Iok aj1
k e
ak
:
proposed method has no such problems. k=1
4192 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 62, NO. 7, JULY 2015

In matrix format, R EFERENCES


t1 t1 t1 [1] J. S. C. M. Raj and A. E. Jeyakumar, A novel maximum power point
e a1
t2
e a2
t2
e am
t2 tracking technique for photovoltaic module based on power plane anal
e a1
e a2
e am ysis of I V characteristics, IEEE Trans. Ind. Electron., vol. 61, no. 9,
1 = .. .. .. .. pp. 47344745, Sep. 2014.
. . . . [2] J. J. Soon, K.S. Low, and S. T. Goh, Multidimension diode photovoltaic
tN tN tN
e a1
e a2
e am (PV) model for different PV cell technologies, in Proc. 23rd IEEE ISIE,
2014, pp. 24962501.
E [3] B. G. Streetman and S. Banerjee, Solid State Electronic Devices, 5th ed.
Io1 1 a1 am1
1
Upper Saddle River, NJ, USA: PrenticeHall, 2000.
[4] K. R. McIntosh, Lumps, humps and bumps: Three detrimental effects
Io2 1 a2 am1
2 in the currentvoltage curve of silicon solar cells, Ph.D. dissertation,
.. .. .. .. .. : Centre Photovoltaic Eng., Univ. New South Wales, Sydney, Australia,
. . . . . 2001.
Iom 1 am am1
m [5] D. Macdonald and A. Cuevas, Reduced ll factors in multicrystalline
silicon solar cells due to injectionlevel dependent bulk recombina
V tion lifetime, Prog. Photovolt, Res. Appl., vol. 8, no. 4, pp. 363375,
Jul./Aug. 2000.
Since is diagonal and V is a standard Vandermonde ma [6] W. De Soto, S. Klein, and W. Beckman, Improvement and validation of
trix, rank() = rank(V ) = m. If t2 t1 = t3 t2 = = a model for photovoltaic array performance, Sol. Energy, vol. 80, no. 1,
pp. 7888, Jan. 2006.
tm tm1 = Ts > 0, as N 2m + 2, the rst m row of E [7] F. Pelanchon, P. Mialhe, and J. Charles, The photocurrent and the open
1Ts 1Ts 1T circuit voltage of a silicon solar cell, Sol. Cells, vol. 28, no. 1, pp. 4155,
s Jan. 1990.
e a1 e a2 e am [8] N. Ravindra and B. Prasad, Saturation current in solar cells: An analy
Em = .. .. .. .. sis, Sol. Cells, vol. 2, no. 2, pp. 109113, Oct. 1980.
. . . .
Ts n1 Ts n1 Ts m1 [9] N. S. Singh, A. Jain, and A. Kapoor, Determination of the solar cell
e a1 e a2 e am junction ideality factor using Special Trans Function Theory (STFT), Sol.
t1 Energy Mater. Sol. C., vol. 93, no. 8, pp. 14231426, Aug. 2009.
e a1 [10] A. Chatterjee, A. Keyhani, and D. Kapoor, Identi cation of photovoltaic
t1
source models, IEEE Trans. Energy Convers., vol. 26, no. 3, pp. 883
e a2 889, Sep. 2011.
..
. [11] D. Pysch, A. Mette, and S. Glunz, A review and comparison of different
t1
methods to determine the series resistance of solar cells, Sol. Energy
e am Mater. Sol. C., vol. 91, no. 18, pp. 16981706, Nov. 2007.
[12] D. Sera and R. Teodorescu, Robust series resistance estimation for di
so rank(E) = rank(Em ) = m. Otherwise, it is always possi agnostics of photovoltaic modules, in Proc. 35th IEEE IECON, 2009,
ble to nd some T such that ti = ni T , ni N for i = pp. 800805.
[13] G. H. Yordanov, O.M. Midtgard, and T. O. Saetre, Series resistance
1; 2; : : : ; m. Construct matrix determination and further characterization of csi PV modules, Renew.
1 1 1 Energy, vol. 46, pp. 7280, Oct. 2012.
T T T [14] Y. S. Kim, S.M. Kang, B. Johnston, and R. Winston, A novel method to
e a1 e a2 e am extract the series resistances of individual cells in a photovoltaic module,
E = .. .. .. .. Rnm n Sol. Energy Mater. Sol. C., vol. 115, pp. 2128, Aug. 2013.
.
nm T
.
nm T
. . [15] A. Laudani, F. R. Fulginei, and A. Salvini, Identi cation of the onediode
nm T
e a1 e a2 e am
model for photovoltaic modules from datasheet values, Sol. Energy,
vol. 108, pp. 432446, Oct. 2014.
and Em is the submatrix of E . Since E is a Vander [16] S. Lineykin and A. Kuperman, Issues in modeling amorphous silicon
monde matrix with full column rank, rank(E) = rank(Em ) = photovoltaic modules by singlediode equivalent circuit, IEEE Trans.
Ind. Electron., vol. 61, no. 12, pp. 67856793, Dec. 2014.
rank(E ) = m. Hence, 1 is full column rank, i.e., [17] Y. A. Mahmoud, W. Xiao, and H. H. Zeineldin, A parameterization
rank(1 ) = m. Thus, approach for enhancing PV model accuracy, IEEE Trans. Ind. Electron.,
vol. 60, no. 12, pp. 57085716, Dec. 2013.
t1 t21 tm+2
1 [18] F. Adamo, F. Attivissimo, A. D. Nisio, and M. Spadavecchia, Characteri
t2 t22 tm+2 zation and testing of a tool for photovoltaic panel modeling, IEEE Trans.
2 Instrum. Meas., vol. 60, no. 5, pp. 16131622, May 2011.
2 = .. .. .. .. [19] Y. Mahmoud, W. Xiao, and H. H. Zeineldin, A simple approach to
. . . .
modeling and simulation of photovoltaic modules, IEEE Trans. Sustain.
tN t2N tm+2
N Energy, vol. 3, no. 1, pp. 185186, Jan. 2012.
[20] S. Lineykin, M. Averbukh, and A. Kuperman, Fiveparameter model of
V2
photovoltaic cell based on STC data, in Proc. 27th IEEE Convers. Elect.
1 Electron. Eng., Eilat, Israel, Nov. 2012, pp. 15.
.. [21] A. Izadian, A. Pourtaherial, and S. Motahari, Basic model and governing
. equation of solar cells used in power and control applications, in Proc.
1 :
(m+1)! IEEE Energy Convers. Congr. Expo., Raleigh, NC, USA, Sep. 2012,
1 pp. 14831488.
(m+2)! [22] M. AlRashidi, M. AlHajri, K. ElNaggar, and A. AlOthman, A new
estimation approach for determining the I V characteristics of solar
As N 2m + 2, the rst m + 2 row of V2 is a Vandermonde cells, Sol. Energy, vol. 85, no. 7, pp. 15431550, Jul. 2011.
matrix, so rank(2 ) = rank(V2 ) = m + 2, i.e., 2 is full col [23] H. Qin and J. W. Kimball, Parameter determination of photovoltaic cells
from eld testing data using particle swarm optimization, in Proc. IEEE
umn rank. Since = [1 ; 2 ] with the full column rank of PECI, 2011, pp. 14.
both 1 and 2 , is also full column rank. N 2m + 2 [24] J. J. Soon and K.S. Low, Optimizing photovoltaic model parameters for
implies that the row number of is no less than the column simulation, in Proc. IEEE 21st ISIE, 2012, pp. 18131818.
[25] Y. Li et al., Evaluation of methods to extract parameters from
number. Hence, rank() = 2m + 2, and T is full rank, i.e., currentvoltage characteristics of solar cells, Sol. Energy, vol. 90, pp. 51
(T )1 exists. 57, Apr. 2013.
LIM et al.: IDENTIFICATION OF DIODE MODELS FROM SINGLE IV CHRACTERISTICS OF PV PANELS 4193

[26] K. F. Teng and P. Wu, PV module characterization using QR decom Zhen Ye received the B.Eng. and M.Eng. de-
position based on the least square method, IEEE Trans. Ind. Electron., grees from Wuhan University, Wuhan, China, in
vol. 36, no. 1, pp. 7175, Feb. 1989. 2000 and 2003, respectively, and the Ph.D. de-
[27] H. Park and H. Kim, PV cell modeling on singlediode equivalent cir gree from the National University of Singapore,
cuit, in Proc. 39th IEEE IECON, 2013, pp. 18451849. Singapore, in 2008, all in electrical engineering.
[28] J. Appelbaum, A. Chait, and D. Thompson, Parameter estimation and After graduation, he worked at the Solar En-
screening of solar cells, Prog. Photovolt, Res. Appl., vol. 1, no. 2, pp. 93 ergy Research Institute of Singapore as a Re-
106, Feb. 1993. search Scientist. Since 2011, he has been a
[29] K. Ishaque and Z. Salam, An improved modeling method to determine Principal Engineer for photovoltaic (PV) module
the model parameters of photovoltaic (PV) modules using Differential technology with REC Cells Pte Ltd., Singapore.
Evolution (DE), Sol. Energy, vol. 85, no. 9, pp. 23492359, Sep. 2011. His research interests include proportional
[30] M. S. Ismail, M. Moghavvemi, and T. M. I. Mahlia, Characterization of integralderivative control, system identi cation, PV system monitoring,
pv panel and global optimization of its model parameters using genetic solar power integrity, and smart grids.
algorithm, Energy Convers. Manage., vol. 73, pp. 1025, Sep. 2013.
[31] W. Xiao, P. R. Dunford, W. G. Palmer, and A. Capel, Regulation of pho
tovoltaic voltage, IEEE Trans. Ind. Electron., vol. 54, no. 3, pp. 1365
1374, Jun. 2007. Jiaying Ye received the Bachelors degree in
[32] M. G. Villalva, J. R. Gazoli, and E. R. Filho, Comprehensive approach microelectronics from Sun Yat-sen University,
to modeling and simulation of photovoltaic arrays, IEEE Trans. Power Guangzhou, China. She is currently working
Electron., vol. 24, no. 5, pp. 11981208, May 2009. toward the Ph.D. degree in the Graduate School
[33] J. J. Soon and K.S. Low, Photovoltaic model identi cation using particle for Integrative Sciences and Engineering and
swarm optimization with inverse barrier constraint, IEEE Trans. Power the Solar Energy Research Institute of Singa-
Electron., vol. 27, no. 9, pp. 39753983, Sep. 2012. pore at the National University of Singapore,
[34] A. Laudani, F. R. Fulginei, and A. Salvini, High performing extracton Singapore.
procedure for the onediode model of a photovoltaic panel from exper After graduation, she worked at Nanyang
imental I V curves by using reduced forms, Sol. Energy, vol. 103, Technological University, Singapore, as a
pp. 316326, May 2014. Project Of cer, studying the resistive switching
[35] F. Ghani, M. Duke, and J. Carson, Numerical calculation of series and property of thin- lm materials. Her research interests include indoor
shunt resistances and diode quality fator of a photovoltaic cell using the characterization of photovoltaic (PV) modules and outdoor performance
Lambert W function, Sol. Energy, vol. 91, pp. 422431, May 2013. analysis for PV systems.
[36] F. Ghani, M. Duke, and J. Carson, Numerical calculation of series and
shunt resistance of a photovoltaic cell using the Lambert W function:
Experimental evaluation, Sol. Energy, vol. 87, pp. 246253, Jan. 2013.
Dazhi Yang received the B.Eng. and M.Sc.
[37] Q.G. Wang and Y. Zhang, Robust identi cation of continuous systems
degrees from the Department of Electrical and
with deadtime from step responses, Automatica, vol. 37, no. 3, pp. 377
Computer Engineering, National University of
390, Mar. 2001.
Singapore, Singapore, in 2009 and 2012, re-
[38] M. Hamdy and R. Call, The effect of the diode ideality factor on the
spectively, where he is currently working toward
experimental determination of series resistance of solar cells, Sol. Cells,
the Ph.D. degree.
vol. 20, no. 2, pp. 119126, Mar. 1987.
He is employed full-time by the Solar Energy
[39] M. C. Di Piazza and G. Vitale, Photovoltaic Sources: Modeling and
Research Institute of Singapore. His research
Emulation. Berlin, Germany: SpringerVerlag, 2012.
interests include statistical modeling for solar
[40] T. Easwarkhanthan, J. Bottin, I. Bouhouch, and C. Boutrit, Nonlinear
irradiance, forecasting methodologies, and en-
minimization algorithm for determining the solar cell parameters with
vironmental data mining.
microcomputers, Int. J. Sol. Energy, vol. 4, no. 1, pp. 112, Jan. 1986.

Li Hong Idris Lim (M14) received the B.Eng. Hui Du received the Diploma degree in mecha-
and Ph.D. degrees in electrical engineering from tronics engineering from Nanyang Polytechnic,
the National University of Singapore, Singapore. Singapore, in 2010. He is currently working to-
From 2008 to 2012, she was with Vestas ward the Bachelors degree in electronics engi-
Technology R&D. Since 2013, she has been a neering at the National University of Singapore,
full-time Assistant Professor with the University Singapore.
of Glasgow, Glasgow, U.K. Her research inter- He is also a full-time employee at the Solar
ests include control of wind energy systems, Energy Research Institute of Singapore.
solar forecasting and system identi cation, and
smart grids.

Anda mungkin juga menyukai