1

Power Optimization for Photovoltaic

Micro-Converters using Multivariable
Gradient-Based Extremum-Seeking
Azad Ghaffari, Sridhar Seshagiri, and Miroslav Krstić

Abstract—It is well-known that distributed architectures such [12], [13], [15]. It offers the advantages of fast convergence
as micro-converters and micro-inverters for photovoltaic (PV) and guaranteed stability over a range of environmental con-
systems can recover between 10%-30% of annual performance ditions, and yet is simple to implement, and hence very cost
loss or more that is caused by partial shading and/or module mis-
match. In this work, we present a novel multivariable gradient- effective in terms of processing/hardware requirements.
based extremum-seeking (ES) design to extract maximum power With the exception of [2], all existing work on ES applies
from an arbitrary micro-converter configuration of PV modules, the technique to PV systems whose cells receive the same
that includes cascade and parallel connections. Conventional irradiance level, i.e., have unimodal power characteristics.
maximum power point tracking (MPPT) schemes for micro- Recent works (for example, [5]) concentrate on designing
converters (where each PV module is coupled to its own DC/DC
converter) employ a distributed control, with one peak seeking MPPT methods to track multiple peaks (non-unimodal power)
scheme per each PV module, thereby requiring one control loop under rapidly changing irradiance conditions, and the issues
and two sensors per module (one each for current and voltage). of partial shading and module mismatch. These studies have
By contrast, the scheme that we present employs a single control led to a growing interest in distributed architectures (also
loop with just two sensors, one for the overall array output referred to as distributed power electronics), such as micro-
current and the other one for the DC bus voltage. This centralized
design provides more flexibility in tuning the parameters of the converters (distributed DC/DC converters) and micro-inverters
controller, and also takes into account interactions between PV (distributed DC/AC converters) [4]. While [2] deals with ES
modules. The computational effort of our design is not higher design for micro-converters (one DC/DC converter for each
than that of the conventional scheme, and simulation and exper- module), it employs a scalar ES loop for each PV module. Two
imental results show that our proposed design outperforms the problems arise here. First, this scheme requires two sensors
conventional one. Thus, our proposed design offers two benefits:
(i) the balance-of-system (BOS) cost reduction as a result of the per module, current and voltage, which increases the levelized
lower number of sensors, and (ii) improved performance; both energy cost. Second, the coupling effect between PV modules
contributing towards reduced average cost/watt, and enhancing is not addressed by this distributed control. Our current work
the economic viability of solar. shows that employing a multivariable MPPT algorithm instead
Index Terms—Photovoltaic systems, distributed power gener- of separate scalar ones solves these problems.
ation, power control, adaptive systems. To the best of our knowledge, there are a limited number
of multivariable MPPT schemes described in the literature,
I. I NTRODUCTION among which we refer the reader to [14], [18], [19]. The
last of these references [19] uses a multivariable version of
M AXIMUM power point tracking (MPPT) is a technique
for maximizing the energy extracted from PV modules.
Over the years, many MPPT methods have been developed
the popular Perturb and Observe (P&O) algorithm. Unlike
scalar designs which require one current sensor for each
module, the algorithm only requires a single current sensor
and implemented [2], [3], [5], [6], [8]–[10], [12]–[19]. These
on the DC bus. The operating point of the DC/DC converters
methods vary in complexity, convergence speed, cost, range
are perturbed asynchronously, to minimize the possibility of
of effectiveness, implementation hardware, and popularity.
converter interaction having a detrimental effect on the other
Comprehensive comparative analyses of currently available
modules. Closely related to [19] is the work in [18], where
techniques can be found in [6], [8], [9].
“extra variables” are employed in the classical P&O algorithm
Extremum-seeking (ES) is a non-model-based real-time
to overcome the limitation of scalar designs, which the authors
optimization algorithm [1], [11], [21], [22] for systems with
say fail when the feasibility region is nonconvex. It is unclear
unknown dynamics that has been applied to a wide range of
how [18] compares with distributed architectures, with respect
technical applications, including MPPT in PV systems [2], [3],
to power loss recovery in the case of module mismatch. Ref-
Azad Ghaffari is with Joint–Doctoral Programs (Aerospace and Mechan- erence [14] uses Particle Swarm Optimization (PSO), which is
ical) between San Diego State University and University of California at an algorithm that employs multiple agents to “search” for the
San Diego, La Jolla, CA 92093-0411, USA, aghaffar@[Link].
Sridhar Seshagiri is with Department of Electrical and Computer Engi- peak power. The paper does not describe the specific criteria
neering, San Diego State University, San Diego, CA 92182-1309, USA, used to select the number of agents and parameters of the
seshagir@[Link]. PSO, nor the conditions on the voltage and power boundary
Miroslav Krstić is with Department of Mechanical and Aerospace Engi-
neering, University of California, San Diego, La Jolla, CA 92093-0411, USA, limits to stop the algorithm at Maximum Power Point (MPP).
krstic@[Link]. Also, in a PV system with a higher number of PV modules, the
 2

(a) (c)
cell 1 6 6
+
+ Rs

Current(A)
Iph VD Rp V 4 4
ns
I
−
− 2 2

0 0 ◦C 1000 W/m 2 0
S cell 2 +
+ 0 10 20
25 ◦ C
30 40 50 0 10 20
800 W/m2
30 40 50
+ Rs 50 ◦ C 600 W/m2
Iph V (b) (d)
VD Rp
ns 75 ◦ C 400 W/m2
240 240
− V
−

Power(W)
160 160
.. .. ..
. . . − 80 80

cell ns +
0
0 10 20 30 40 50 0 10 20 30 40 50
0
+ Rs Voltage(V) Voltage(V)
Iph VD Rp V
ns
− Fig. 2. Characteristic (a) I-V and (b) P -V for varying irradiance, T =25◦ C.
−
Characteristic (c) I-V and (d) P -V for varying temperature, S=1000W/m2 .

Fig. 1. Equivalent circuit of a PV module.

remarks in Section VI. A preliminary version of this paper
was presented at the 2012 ACC [7]. The primary contribution
process of reinitialization and the tracking performance depend of this work over [7] is the addition of experimental results to
strongly on variable conditions like environmental factors, the the simulation results that were presented therein.
nature of the PV modules, and the shading area. The authors
claim that the required number of sensors are reduced to two,
but to compute the pulse duration, the output voltage of each II. P HOTOVOLTAIC M ODULES AND P OWER E XTRACTION
boost converter needs to be monitored by a separate sensor. Our design and analysis are based on the standard PV
We present a multivariable gradient-based ES schemes with module model described for example in [20], and shown
the following features: schematically in Fig. 1. Each PV cell is modeled as an ideal
• It is applied to micro-converter systems, and hence deals current source of value Iph in parallel with an ideal diode
with the case of non-unimodal power characteristics, and with voltage VD . Electrical losses and contactor resistance
deals specifically with the issue of module mismatch (for are accounted for by the inclusion of the parallel and series
example, possibly different irradiance levels as a result resistances Rs and Rp respectively. The amount of generated
of partially shaded conditions). current Iph is dependent on the solar irradiance S and the
• The use of the non-model-based ES technique makes temperature T through the following equation
the design robust to partial knowledge of the system
S
 
parameters and operating conditions. Iph = Iphr
+ ki (T − Tr )


, (1)
• As opposed to scalar designs, our multivariable design 1000
only requires 2 sensors in all, for the overall PV system r
where Iph is a reference short-circuit current, Tr a reference
current, and the DC bus voltage. This is a significant
temperature, and ki the short-circuit temperature coefficient.
reduction in hardware cost.
The diode models the effect of the semiconductor material and
• Moreover, interactions between PV modules are inher-
its I-V characteristics are given by
ently part of the multivariable design, and hence the
transient performance is less-sensitive to environmental
   
VD
variable variations than a corresponding scalar design. ID =I0 exp −1 , (2)
N Vt
• The computational burden is of the same order as a scalar  3
T
    
Eg 1 1 kT
design, but with a slightly faster transient response than I0 =I0r exp − , Vt = (3)
scalar ES designs, and significantly faster than non-ES Tr N Vt Tr T q
based designs such as [14]. where ns , I0r , Eg and N are respectively the number of
The rest of this paper is organized as follows: The math- series PV cells in the module, the diode reference reverse
ematical model of a PV module, along with a discussion saturation current, the semiconductor bandgap energy (barrier
of the DC/DC converter power electronics, is presented in height), and the emission coefficient, the later three being cell
Section II. Section III introduces the scalar gradient-based ES material/construction dependent, Vt is the thermal cell voltage,
scheme, presented for clarity for the case of a single module and k = 1.38 × 10−23 J/K and q = 1.6 × 10−19 C are
first, followed by how this is conventionally extended to the Boltzman’s constant and the charge on an electron respectively.
distributed micro-converter case. Our proposed multivariable The cell model described by the above equations along with
gradient-based ES is presented and discussed in Section IV, KCL/KVL: I = Iph −ID −VD /Rp , VD = V /ns +Rs I, is then
along with some simulation and experimental results in Sec- represent a PV module by considering ns cells in series (each
tion V, and a summary of our design and some concluding having cell voltage Vt ), so that the terminal I-V relationship
 3

DC/DC Converter DC Bus

d V P =V ×I
I Idc DC/DC Converter J(·)
(T , S) +
V Vdc
2 s
PV − a sin(ωt) sin(ωt)
a s + ωh
d
dˆ kg ĝ ωl
MPPT + ×
s s + ωl

Fig. 3. PV module supplying power to a DC bus via a DC/DC converter. Fig. 4. Scalar extremum seeking for MPPT of a PV module.

for the PV module is given by Also we have ∂V /∂d = −Vdc /(η̄d2 ) then
" V
! # " # ∂J ∗ Vdc
ns + Rs I ( nVs + Rs I) ḡ = (d ) = − g=0 (9)
I = Iph−I0 exp −1 − . (4) ∂d η̄(d∗ )2
N Vt Rp
∂2J ∗ Vdc2
h̄ = (d ) = h < 0. (10)
∂d2 η̄ 2 (d∗ )4
For the sake of model development and performing simula-
tions (done using the SimPowerSystems toolbox of Simulink), Many MPPT techniques, including the classical perturb-and-
we pick the PV module 215N from Sanyo, with the following observe (P&O) class of methods, and extremum-seeking (ES)
numerical values derived from the manufacturer’s datasheet: techniques, are based on detecting the sign of the power
Eg = 1.16 eV, N = 1.81, I0r = 1.13 × 10−6 A, Iph r
= 5.61 A, gradient. The next section discusses scalar gradient-based ES
ki = 1.96 mA/K, Tr = 298.15 K, Rs = 2.48 mΩ, in more detail.
Rp = 8.7 Ω, and the number of PV cells connected in series
is ns = 72. The resulting I-V and P -V curves are shown
III. S CALAR G RADIENT-BASED E XTREMUM S EEKING
in Fig. 2. As is clear from Fig. 2(b,d), the power-voltage
(P -V ) characteristic has a unique but (T , S) dependent peak Several authors have considered to use scalar gradient-based
(V ∗ , P ∗ ). It is the job of the MPPT algorithm to automatically ES for the MPPT problem [2], [3], [12], [13], [15]. Fig. 4
track this peak. In many grid-tied PV systems (including our shows the basic setup of the scheme for the case of a single PV
current work), this is done by means of a separate DC/DC module, and its principal features have been explained fairly
power electronics stage that serves two functions: (i) regulating clearly in the aforementioned references, but we reproduce
the output DC voltage at a (near) constant value, and (ii) them here for the sake of completeness/clarity.
extracting maximum power by forcing the PV module output The injection of the small periodic perturbation a sin(ωt)
V to equal V ∗ . Fig. 3 shows this setup for a DC/DC converter to the estimate dˆ of the optimal pulse duration d∗ results in
stage, whose output voltage is maintained constant as Vdc . a periodic power output P , whose DC component is removed
The ratio between the input voltage V and output voltage Vdc by the wash-out filter s/(s + ωh ), with the resultant signal
can be controlled by changing the duty cycle of the transistor being in phase or out of phase with the perturbation according
switch, which serves as the control input d. We use a step- to whether dˆ is greater than or less than d∗ respectively.
down or buck converter in our design. Under the assumption Multiplication of this signal by 2 sin(ωt)/a and extracting
that the buck converter is working in Continuous Current the DC component of the product using the lowpass filter
Mode (CCM), and that the switching Pulse Width Modulation ωl /(s + ωl ) results in an estimate of the gradient of the cost
(PWM) frequency fs is significantly higher than the bandwidth function. Defining d˜ = dˆ − d∗ , and expanding P about its
of the control loop, the buck converter input-output voltage optimal value and using (7), (8), we see that the ES design of
and current relationships are given by the following (averaged) Fig. 4 implements the gradient update law
relations:
˙
d˜ = kg ĝ, ĝ = h̄d˜
Vdc
V = (5) V2
η̄d = kg 2 dc∗ 4 hd˜ (11)
I = dIdc , (6) η̄ (d )
where h is the Hessian. Design guidelines for selecting the
where η̄ shows the power efficiency of the buck converter. parameters a, ω, ωh , ωl , and kg can be found in [11], but
From (4), (5) and Fig. 2(b,d), it follows that at the MPP are mentioned here for sake of completeness. The frequency
def ω must be chosen small enough to ensure that the plant
(V ∗ ,P ∗ ), the power P = IV = f (V )V = J(V ), satisfies
dynamics appear as a static nonlinearity from the viewpoint
∂J ∗ of the ES loop, and the filter frequencies chosen such that
g = (V ) = 0 (7) ωh ≤ ωl ≪ ω, so that the low pass filter attenuates the
∂V
2
∂ J ∗ perturbation frequency, whereas the highpass filter does not.
h = (V ) < 0. (8) The adaptation gain kg and the amplitude a of the probing
∂V 2
 4

Converter
I1 Io1
(T1 , S1 )
P1 = V1 I1

Idc
= = = PV1 V1 Vo1
PV11 PV21 PV11

(T11 , S11 ) = (T21 , S21 ) = (Tm1 , Sm1 ) =

D1
MPPT
+

Im
I1

I2
Converter Inverter
I2 Io2 Igrid
(T2 , S2 )

PV2 V2
P2 = V2 I2
Vo2 Vdc ∼

DC Bus
∼
PV12 = PV22 = PV12 =

(T12 , S12 ) = (T22 , S22 ) = (Tm2 , Sm2 ) =

Vdc
AC grid
D2 Uinv
MPPT

− In
Converter
Ion Idc
(Tn , Sn )
Pn = Vn In
PVn Vn Von
PV1n = PV2n = PVmn =

(T1n , S1n ) = (T2n , S2n ) = (Tmn , Smn ) =

Dn
MPPT

Fig. 6. Distributed MPPT for one string. One scalar ES loop is used for each
Fig. 5. PV system including m parallel strings. Each string has n PV modules
PV module.
in cascade

Converter
I1 Io1
signal need to be “sufficiently small”. Define (T1 , S1 )

PV1 V1 Vo1

ω=ǫω ′ (12)

P = Vdc Idc
ωl =ǫδωl′ (13) D1

ωh =ǫδωh′ (14) Converter Inverter

I2 Io2 Igrid
Centralized MPPT

(T2 , S2 )
kg =ǫδkg′ , (15)
PV2 V2 Vo2 Vdc ∼
where ǫ and δ are small positive real numbers, and ω ′
, ωh′ , ωl′
∼ AC grid

and kg′ are O(1) positive real parameters. The analysis of [11] D2 Uinv

shows that for sufficiently small ǫ, a, and δ, the output P Converter

converges to an O(ǫ + δ + a)-neighborhood of the MPP P ∗ . (Tn , Sn )

In Ion Idc

The above design for a single module can be extended PVn Vn Von

to the PV system shown in Fig. 5, that has m parallel

strings, with each string having n modules in cascade (series). Dn

Since irradiance (and temperature to a lower extent) may

vary between the modules, the peak power is not necessarily
the same for all of them. This “module mismatch” therefore
Fig. 7. Our proposed centralized MPPT for PV system, just one multivariable
results in maximum powers for string architectures that are ES loop is employed for all PV modules.
lower than the sum of the individual maximum powers of the
modules, which in turn has led to the use of micro-converters,
where each module is coupled with its own DC/DC converter. IV. M ULTIVARIABLE G RADIENT-BASED E XTREMUM
Micro-converter architectures can recover between 10%-30% S EEKING
of annual performance loss caused due to module mismatch. A block schematic of our proposed multivariable gradient-
The conventional way to implement the MPPT algorithm in based ES is shown in Fig. 7. As is clear from the schematic,
micro-converters is to simply extend the preceding (scalar) the design employs just one ES loop with two sensors for the
MPPT design to each PV module, as shown in Fig. 6 for one overall system, one each for the DC bus voltage Vdc and the
string. In each string we therefore have n separate control overall current Idc .
loops, with no consideration to the interaction between the Fig. 8 shows the multivariable extension of the ES design
series modules. We also still have two sensors per module, that is described by Fig. 4, and its principal features are essen-
that measure the module voltage and current. The multivariable tially the same as discussed in Section III. The perturbation
control algorithm that we present in the next section alleviates signals are defined as
both these issues; on the one hand, it considers the interaction  T
between modules, resulting in better performance, and in S(t) = a sin(ω1 t) · · · sin(ωn t) (16)
addition, uses just two sensors for the overall system, resulting 2 T
M (t) = sin(ω1 t) · · · sin(ωn t) , (17)
in hardware reduction cost. The details of the actual design are a
presented in the next section. where ωi /ωj are rational for all i and j, and a is a real number,
 5

Parameter Value Unit Parameter Value Unit

D1 DC/DC V1 PV 1 P1
fs 100 kHz Vdc 300 V
D D2 DC/DC V2 PV 2 P2 P = Vdc Idc Ci 3 µF Co 220 µF
. . . . .
.
.
.
.
.
.
.
.
.
.
L 220 µH d0 0.5 –
Dn DC/DC Vn PVn Pn Kg 0.01I6×6 – kg 0.01 –
s ω 7000 rad/s a 0.01 –
S(t) M (t) ωl 50 rad/s ωh 45 rad/s
s + ωh ω1 4500 rad/s ω4 5500 rad/s
ω2 6500 rad/s ω5 7500 rad/s
Kg ωl Pac = P − Pdc
+ D̂ Ĝ × ω3 8500 rad/s ω6 9500 rad/s
s s + ωl m 2 – n 3 –
TABLE I
PARAMETERS USED IN THE SIMULATIONS .
Fig. 8. Multivariable ES for MPPT of a PV system.
In addition, the diagonal structure of H in the scalar case,
coupled with the fact that this varies with irradiance, means
with the frequencies chosen such that ωi 6= ωj and ωi + ωj 6= that in the scalar design, the convergence rate of the parameters
ωk for distinct i, j, and k. In particular, the design derives an is very sensitive to partial shading, where the irradiance
estimate Ĝ of the gradient vector by adding a probing signal varies strongly from one module to another. The multivariable
to the estimate scheme, on the other hand, is less sensitive to the changes in
h iT the power-voltage characteristic of a specific module which
D̂ = D̂1 D̂2 · · · D̂n results from variation of temperature or irradiance.
of the pulse duration vector (of all the DC/DC converters), and
“filtering” the resultant power P through the process described V. S IMULATION AND E XPERIMENTAL R ESULTS
before. With no additional information on the Hessian (and To show the effectiveness of the proposed centralized MPPT
also for simplicity), we choose the amplitudes of the probing algorithm in Fig. 6, we present both simulation as well as
signals to all be the same value a. As before, smallness of the experimental results.
probing frequencies and the matrix gain Kg are ensured by
selecting these as
A. Simulation Results
ωi = ǫωi′ , i ∈ {1, 2, · · · , n} (18) For the simulations, we consider a PV system with m = 2
Kg = ǫδKg′ (19) parallel strings and n = 3 cascade modules in each string.
The PV modules are model 215N from Sanyo, with datasheet
where ǫ and δ are small positive constants, ωi′ is a rational
parameters presented in Section II. Guidelines for the selection
number, and elements of Kg′ are O(1) positive real parameters.
of the ES parameters were presented in [7], and are reproduced
The filter coefficients ωl and ωh are defined by (13) and (14).
here for completeness.
As before, it can be shown that for sufficiently small ǫ, δ, and
Selecting all the frequencies in a narrow range creates large
a, and with Kg > 0, the estimate D̂ of the pulse duration vec-
overshoots and steady state errors in parameter estimation.
tor and the output P converge to O(ǫ + δ + a)-neighborhoods
However, choosing the frequencies in a wide range causes
of the optimal pulse duration D∗ = [D1∗ D2∗ · · · Dn∗ ]T and
very different convergence rates in each channel. Since we
the MPP P ∗ respectively.
set the lowpass filter frequency equal for all the channels, the
The differences between the scalar and multivariable designs
amplitude of the perturbation signal with the lowest frequency
become clear when one considers the update equations for the
reduces less than that with the highest frequency, which in
estimation error D̃ = D̂ − D∗ . In the multivariable case, we
turn results in a higher feedback gain for the low frequency
have
∂2P channel, which derives the parameter faster to the optimal
D̃˙ = Kg H D̃, H := (D∗ ) (20) value. It is possible to tune the matrix gain elements with
∂D2
respect to the selected frequencies. What this means is that in
where H is the (negative definite) Hessian and P = Vdc Idc .
order to have the same convergence rate for a wide range of
In the scalar ES design of Fig. 6 however, the above equation
selected frequencies, we can choose a higher gain for higher
is replaced by
frequencies to compensate the effect of lowpass filter. We
D̃˙ = kg HD̃ (21) prefer to select the frequencies in a reasonable range, between

h̄P V1 0 0 ··· 0
 50% up and down of the central frequency. We remind the
 0 h̄P V2 0 ··· 0  reader that the central frequency should be small enough in
H :=

 .. .. .. . . .. ,

(22) comparison to the PWM frequency. We suggest that this be of
 . . . . .  the order of less than 1% of the PWM frequency. The transient
0 0 0 ··· h̄P Vn for the estimate of the gradient vector contains frequencies
where that include harmonics of ωi − ωj , for all distinct i and j.
The bandwidth of the lowpass filter needs to be designed
h̄P Vi = ∂ 2 Pi /∂d2i , Pi = V i I i ∀i (23)
with respect to these values. We suggest selecting ωl to be
(see Fig. 6), so that the equations are decoupled, and there is no of the order of 5% of the least difference between the probing
way to affect the power extraction in one module by changing frequencies. The final step is selecting the cut-off frequency of
the pulse duration of the DC/DC converter of another module. the highpass filter, which we simply choose to be smaller than
 6

I1 IRF640 IL1 L1 LA 25-NP Io1

◦
1.3 + +
PV1 C
IC1
V1 MUR2020 Vo1
Power (kW)

+
IR2127
(T1 , S1 ) − rC1
1.1 −

Electronic load

Vdc = 5V
Centralized I2 IRF640 IL2 L2 LA 25-NP Io2
0.9 Distributed ◦

PV2 + C2 +
0.5 0.75 1 1.25 IC2

–
Time(s) V2 MUR2020 Vo2
IR2127
(T2 , S2 ) − rC2
−
Fig. 9. Simulation results in a partial shading scenario. Extracted power by Idc
(solid red) centralized and (dashed blue) distributed MPPT schemes.
IL2
PWM1 PWM2 ADC5
IC2
or equal to ωl . Based on the preceding remarks, the numerical ADC6
DS1104 Vo2
values of the design parameters are as presented in Table I. MPPT ADC8
The temperature T is assumed to be equal to 25 ◦ C for ADC7
Vdc

all modules throughout. The irradiance S is assumed to be

equal to 1000 W/m2 initially, with a step change to 500 W/m2 Fig. 10. Hardware configuration of the experimental setup.
for modules PV12 and PV23 at t = 0.5 s and then back to
1000 W/m2 at 1 s, so as to simulate partial shading on some
Voltage (V)
modules. The output power of the entire system is shown in
0 1 2 3 4 5 6
Fig. 9. It is clear from Fig. 9 that the multivariable algorithm 7
1.5
performs a uniform and faster transient against step up or step 6
down changes in the generated power. 1.2 5
At the beginning all PV modules and converters have the
Current (A)

Power (W)
same settings, gains, and initial conditions. Also all modules 0.9 4
are under the same irradiance and temperature. Hence, the 3
0.6
transient of the scalar ES for all parameters is the same. 2
On the other hand, multivariable ES shows different transient 0.3
1
for each parameter which is happening because of different
0 0
frequencies of the perturbation function in each channel. The 0 1 2 3 4 5 6
lowest frequency shows the fastest response, along with a Voltage (V)
correspondingly larger overshoot. It is possible to tune matrix
Kg such that all transients look the same. Fig. 11. Power and current maps of the custom-made PV modules used in
When the modules in each string are partially shaded, the the experiments for T = 25◦ C. (solid line) S = 1000 W/m2 , (dashed)
S = 520 W/m2 , and (dash-dot) S = 190 W/m2 .
overall power level decreases. The multivariable ES design
recovers from this power level change faster than the scalar
version. As clear from Fig. 9, the power goes to the MPP in interact with the DC/DC converters through Connector Panel
less than half the time needed for the scalar scheme. CP1104. Also we use the “Power-Pole Boards” developed by
The irradiance level of the partially shaded modules is the University of Minnesota for educational purposes, that are
returned to 1000 W/m2 at t = 1 s. At this point both schemes general purpose DC-DC converter boards, configured here as
show a similar transient. It is concluded that the convergence DC/DC buck converters, with external PWM signals generated
rate of the multivariable scheme does not vary largely from by the DS1104. Each Power-pole board has a current sensor
step up to step down in power generation, which is not true LA 25-NP to measure the inductor current which we use along
for the scalar ES. It is clear that in the step down situation with the capacitor ripple current measurement to calculate
the scalar scheme shows a slower performance than the step the DC bus current. We employ the DC bus current and
up case. DC bus voltage to measure the power supplied to the DC
bus. The hardware setup is shown in Fig. 12. The numerical
B. Experimental Results values of the parameters are as follows: ω = 100π rad/s,
Our hardware setup consists of 2 cascade PV modules ω1 = 0.9ω, ω2 = ω, ωl = ωh = ω/20, kg = 2, Kg =
connected to an active load which plays the role of the DC kg I2×2 , a = 0.05, and D0 = [0.7 0.7]T . We use high order
bus with Vdc = 5V, as shown in Fig 10. The PV modules Bessel filters instead of the first order lowpass and highpass
are custom-made using 12 PV cells, with P -V and I-V char- filters to separate the AC and DC parts of the power signals
acteristics shown in Fig. 11. We use dSPACE Control Desk and to create a smooth estimate of the gradient vector. The
Next Generation software and the DS1104 R&D Controller PWM frequency is 100 kHz and the sampling time of the
Board to implement our MPPT algorithms inside Simulink and MPPT algorithm is 0.3 ms. The temperature of PV modules
 7

80
Centralized
Distributed
70

D̂1 (%)
60

50
40 60 80 100 120 140 160

90

D̂2 (%)
Fig. 12. Experimental setup. 80

Centralized
70 Distributed
10
Centralized 40 60 80 100 120 140 160
Distributed Time(s)
9

Fig. 14. Experimental results of the adaptation of the pulse duration. (solid
red) Centralized, and (dashed blue) distributed MPPT algorithms.
Power (W)

7
individual powers causes different transients in the parameter
updates, particularly in response to sudden irradiance changes
6 caused by partial shading. The multivariable extremum seeking
design that we present removes these drawbacks. Since the
5 Hessian of the entire system (and not individual modules)
40 60 80 100 120 140 160
Time(s) defines the performance of the parameter update, this leads
to more uniform transients in response to irradiance and
Fig. 13. Experimental results of the generated power in a partial shading temperature changes, lower power ripple than the scalar de-
scenario. (solid red) Centralized, and (dashed blue) distributed MPPT algo- sign, and improved overall performance. The scheme also
rithms.
only uses two sensors for the overall system, resulting in
lower hardware cost. The dual advantages contribute towards
is 25◦ C and the modules are fully exposed to the sun from reduced average cost/watt, enhancing the economic viability
time 0–60 s and from 120–180 s. To simulate the effect of solar. The effectiveness of the proposed design is validated
of partial shading, PV1 is covered with a plastic mat from by both analysis as well as experimental results.
time 60–120 s. When one module is partially shaded, the
overall power level decreases. As clear from Fig. 13, the R EFERENCES
centralized design recovers from this power level change faster [1] K. Ariyur and M. Krstić, Real-Time Optimization by Extremum Seeking
than the distributed version. Furthermore, Fig. 14 shows that Feedback. Wiley-Interscience, 2003.
the adaptation process of the pulse duration in the centralized [2] A. Bratcu, I. Munteanu, S. Bacha, D. Picault, and B. Raison, “Cascaded
DC-DC converter photovoltaic systems: power optimization issues,”
method is faster than the distributed design with the same IEEE Transactions on Industrial Electronics, vol. 58, pp. 403–411, 2011.
MPPT gain in both algorithms. [3] S. Brunton, C. Rowley, S. Kulkarni, and C. Clarkson, “Maximum
The irradiance level of the partially shaded module is re- power point tracking for photovoltaic optimization using ripple-based
extremum seeking control,” IEEE Transactions on Power Electronics,
turned to normal level at t = 120 s. At this point both schemes vol. 25, pp. 2531–2540, 2010.
show a similar transient. It is concluded that the convergence [4] C. Deline, B. Marion, J. Granata, and S. Gonzalez, “A performance
rate of the centralized scheme does not vary largely from step and economic analysis of distributed power electronics in photovoltaic
systems,” in Technical Report, National Renewable Library, 2011.
up to step down in power generation, which is not true for the [5] S. Dhople, J. Ehlmann, A. Davoudi, and P. Chapman, “Multiple-input
distributed MPPT. It is clear that in the step down situation the boost converter to minimize power losses due to partial shading in
distributed scheme shows a slower performance than the step photovoltaic modules,” in Proc. of IEEE Energy Conversion Congress
and Exposition (ECCE), 2010.
up case. As expected, the experimental results are in keeping [6] T. Esram and P. Chapman, “Comparison of photovoltaic array maxi-
with the analytical and simulation results. mum power point tracking techniques,” IEEE Transactions on Energy
Conversion, vol. 22, pp. 439–449, 2007.
[7] A. Ghaffari, S. Seshagiri, and M. Krstić, “Power optimization for photo-
VI. C ONCLUSIONS voltaic micro-converters using multivariable gradient-based extremum-
Using extremum seeking in a micro-converter configuration seeking,” in Proc. of the American Control Conference, 2012.
[8] D. P. Hohm and M. E. Ropp, “Comparative study of maximum power
is a promising way to extract maximum power from a PV point tracking algorithms,” Progress in Photovoltaics: Research and
system. Conventionally used scalar gradient-based designs do Applications, vol. 11, p. 4762, 2003.
so based on the generated power of each module. On the one [9] S. Jain and V. Agarwal, “Comparison of the performance of maximum
power point tracking schemes applied to single-stage grid-connected
hand, this requires two sensors per module, and on the other, photovoltaic systems,” IET Electric Power Applications, vol. 1, pp. 753–
the dependence on the level and direction of changes of the 762, 2007.
 8

[10] R. Kadri, J.-P. Gaubert, and G. Champenois, “An improved maximum Miroslav Krstić holds the Daniel L. Alspach en-
power point tracking for photovoltaic grid-connected inverter based on dowed chair and is the founding director of the
voltage-oriented control,” IEEE Transactions on Industrial Electronics, Cymer Center for Control Systems and Dynamics
vol. 58, pp. 66–75, 2011. at UC San Diego. He also serves as Associate
[11] M. Krstić and H.-H. Wang, “Stability of extremum seeking feedback for Vice Chancellor for Research at UCSD. Krstic is
general nonlinear dynamic systems,” Automatica, vol. 36, pp. 595–601, a recipient of the PECASE, NSF Career, and ONR
2000. Young Investigator Awards, as well as the Axelby
[12] P. Lei, Y. Li, Q. Chen, and J. Seem, “Extremum seeking control based and Schuck Paper Prizes. Krstic was the first re-
integration of MPPT and degradation detection for photovoltaic arrays,” cipient of the UCSD Research Award in the area
in Proc. of American Control Conference, 2010. of engineering (immediately following the Nobel
[13] R. Leyva, C. Alonso, I. Queinnec, A. Cid-Pastor, D. Lagrange, and laureate in Chemistry Roger Tsien) and has held
L. Martinez-Salamero, “MPPT of photovoltaic systems using extremum the Russell Severance Springer Distinguished Visiting Professorship at UC
seeking control,” IEEE Transactions on Aerospace and Electronic Sys- Berkeley and the Harold W. Sorenson Distinguished Professorship at UCSD.
tems, vol. 42, pp. 249–258, 2006. He is a Fellow of IEEE and IFAC and serves as Senior Editor in IEEE
[14] M. Miyatake, M. Veerachary, F. Toriumi, N. Fuji, and H. Ko, “Maximum Transactions on Automatic Control and Automatica. He has served as Vice
power point tracking of multiple photovoltaic arrays: a PSO approach,” President of the IEEE Control Systems Society and chair of the IEEE CSS
IEEE Transactions on Aerospace and Electronic Systems, vol. 47, pp. Fellow Committee. Krstic has coauthored nine books on adaptive, nonlinear,
367–380, 2011. and stochastic control, extremum seeking, control of PDE systems including
[15] S. Moura and Y. Chang, “Asymptotic convergence through Lyapunov- turbulent flows, and control of delay systems.
based switching in extremum seeking with application to photovoltaic
systems,” in Proc. of American Control Conference, 2010.
[16] F.-S. Pai, R.-M. Chao, S. H. Ko, and T.-S. Lee, “Performance evaluation
of parabolic prediction to maximum power point tracking for PV array,”
IEEE Transactions on Sustainable Energy, vol. 2, pp. 60–68, 2011.
[17] H. Patel and V. Agarwal, “MPPT scheme for a PV-fed single-phase
single-stage grid-connected inverter operating in CCM with only one
current sensor,” IEEE Transactions on Energy Conversion, vol. 24, pp.
256–263, 2009.
[18] G. Petrone, G. Spagnuolo, and M. Vitelli, “A multivariable perturb-
and-observe maximum power point tracking technique applied to a
single-stage photovoltaic inverter,” IEEE Transactions on Industrial
Electronics, vol. 58, pp. 76–84, 2011.
[19] C. A. Ramos-Paja, G. Spagnuolo, G. Petrone, M. Vitelli, and J. Bastidas,
“A multivariable MPPT algorithm for granular control of photovoltaic
systems,” in Proc. of IEEE International Symposium on Industrial
Electronics, 2010.
[20] G. Vachtsevanos and K. Kalaitzakis, “A hybrid photovoltaic simulator
for utility interactive studies,” IEEE Transactions on Energy Conversion,
vol. EC-2, pp. 227–231, 1987.
[21] H.-H. Wang and M. Krstic, “Extremum seeking for limit cycle minimiza-
tion,” IEEE Transactions on Automatic Control, vol. 45, pp. 2432–2436,
2000.
[22] H.-H. Wang, S. Yeung, and M. Krstic, “Experimental application of
extremum seeking on an axial-flow compressor,” IEEE Transactions on
Control Systems Technology, vol. 8, pp. 300–309, 1999.

Azad Ghafari received his B.S. degree in Electrical

Engineering and M.S. degree in Control Engineer-
ing from K. N. Toosi University of Technology in
Tehran, Iran. He is now working towards his PhD
degree in the Joint Doctoral Program in Mechanical
and Aerospace Engineering at University of Califor-
nia, San Diego, and San Diego State University. His
research interests include extremum seeking, con-
trol of photovoltaic arrays, wind energy conversion
systems, induction machines, power electronics, and
sliding mode control.

Sridhar Seshagiri received the [Link] degree from

the Indian Institute of Technology, Madras, and
the M.S. and Ph.D degrees from Michigan State
University, in 1995, 1998, and 2003 respectively,
all in electrical engineering. He joined the Electrical
& Computer Engineering Department at San Diego
State University in 2003, where he is currently an
Associate Professor. His research interests are non-
linear control with applications to energy systems,
particularly control of power electronic converters.