Sunteți pe pagina 1din 11

Available online at www.sciencedirect.

com

ScienceDirect
Solar Energy 132 (2016) 321331
www.elsevier.com/locate/solener

Maximum power point tracking algorithms for single-stage


photovoltaic power plants under time-varying reactive power injection
M. Carrasco , F. Mancilla-David
Department of Electrical Engineering, University of Colorado Denver, Denver, CO 80217, USA
Received 2 October 2015; received in revised form 18 February 2016; accepted 12 March 2016
Available online 25 March 2016
Communicated by: Associate Editor Bibek Bandyopadhyay

Abstract
The power electronic interface of traditional grid-connected photovoltaic (PV) systems includes a boost dc/dc converter and an inverter. In large PV power plants, the boost converter is usually eliminated by connecting an appropriate number of PV panels in series to
achieve the desired voltage. This architecture, referred to as single-stage topology, is the subject of this paper. The paper presents a unied
dynamic model which is able to reproduce the behavior of any single-stage PV power plant with an arbitrary PV array conguration
through a single circuital representation. A methodology based on this unied model is developed to enhance the performance of standard iterative maximum power point tracking (MPPT) algorithms under time-varying reactive power injection. In particular, enhanced
versions of the perturb & observe and incremental conductance algorithms are discussed. Furthermore, a predictive MPPT algorithm
based on a neural network is proposed to improve the dynamic of the dc-link voltage. The unied model and MPPT control schemes
are validated via detailed PSCAD/EMTDC computer simulations in a 450 kW PV system, designed according to a typical 1000 V dc
architecture. Real solar irradiance and cell temperature data collected in a partially cloudy day are utilized in the simulations, providing
comparative performance under rapidly changing atmospheric conditions.
2016 Elsevier Ltd. All rights reserved.

Keywords: Photovoltaic panels; MPPT; Single-stage PV systems; Grid-connected PV systems

1. Introduction
Photovoltaic (PV) power generation is increasingly
attracting the attention of industry and academia mainly
motivated by the urgent need to depart from fossil fuelbased electricity generation. As the cost of PV panels production continues to decrease, it is expected that bulk solar
power generation will be competitive with other forms of
renewable energy, and hence massively deployed.
As illustrated in Fig. 1(a), traditional PV systems are
connected to the ac grid via a power electronic interface
Corresponding author.

E-mail address: miguel.carrascolopez@ucdenver.edu (M. Carrasco).


http://dx.doi.org/10.1016/j.solener.2016.03.023
0038-092X/ 2016 Elsevier Ltd. All rights reserved.

that includes a boost dc/dc converter, a dc-link capacitor,


an inverter, and a low-pass lter. The boost converter lifts
the PV array voltage level to a constant voltage across the
dc-link which is then converted to ac by the inverter. In this
architecture, the maximum power point tracking (MPPT)
control is performed by the dc/dc converter. The inverter,
in turn, transfers the power collected from the PV array
to the ac grid with independent control of reactive power.
However, in large PV power plants the dc/dc converter
can be eliminated if a number of PV panels are stacked
in series so that the dc-link voltage is large enough for
proper operation of the inverter, resulting in a single-stage
topology. See Fig. 1(b). A single-stage architecture is more
ecient and economical than its two-stage counterpart

322

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

Fig. 1. Grid-connected PV power plants. Double-stage (a) and single-stage (b) architectures.

(Kadri et al., 2011). In this topology, the inverter performs


both the MPPT function, as well as the power transfer to
the ac grid. Research on the topological realization of the
inverter includes: H-bridge multilevel inverters (Cecati
et al., 2010), yback inverters with center-tapped secondary
(Kasa et al., 2005), Z-source inverters (Huang et al., 2006),
and current-source converters (Sahan et al., 2008). However, the topology realized through a two-level voltagesource inverter (VSI) is widely adopted as the conguration
of choice by most major PV system manufacturers
(Yazdani et al., 2011; Figueres et al., 2009). Several works
have studied dynamic models for thesingle-stage VSI-based
PV system. See (Kadri et al., 2011; Figueres et al., 2009;
Yazdani and Dash, 2009) and references therein. In
Nousiainen et al. (2013), a low-frequency model of a PV
panel is presented based on the operating-point-dependent
output impedance of a single-diode model. This low-frequency model is used to study the eect of the PV nonlinear
characteristic on the operation of dcdc and dcac converters. In Messo et al. (2014), the aforementioned PV model
along with a linearized model of the three-phase PV inverter cascaded with a L-lter is utilized to determine the minimum value of the dc-link capacitance required for stable
operation. In the present work, a unied dynamic model
for a single-stage VSI-based PV system cascaded with a
LC-lter making use of a modied PV array model is
described.
Due to the absence of a dc/dc converter, the MPPT
function is realized by altering the dc-link voltage including
an additional outer loop in the inverter control scheme. As
a result, the inverter is operated at variable dc-link voltage,
where the reference for the voltage is provided by the
MPPT algorithm. Once the dc-link voltage reference is
computed, the inverter is modulated through an inner loop
based on standard current control methods (Yazdani et al.,
2011; Yazdani and Iravani, 2010). Variations of this control structure have also appeared in the literature. In
Yazdani and Dash (2009), a feedforward compensation

mechanism is proposed for the dc-link voltage-control loop


in order to make the PV system dynamics immune to the
PV array nonlinear characteristic. Other approaches for
controlling single-stage PV systems such as nonlinear control strategies (Bao et al., 2013) and resonant controllers
(Mastromauro et al., 2012) have also been investigated.
MPPT plays an important role in increasing the eciency of PV systems and therefore some of the ongoing
eorts are focused on developing and improving MPPT
algorithms. Since there are many dierent MPPT techniques, several comparative studies have been reported.
While some of these studies were performed without focusing on a specic topology (Subudhi and Pradhan, 2013;
Esram and Chapman, 2007), most make use of a dcdc
converter (de Brito et al., 2013; Chen et al., 2015). A comparison of MPPT techniques for single-stage PV systems is
reported in Jain and Agarwal (2007), but it is limited to single-phase systems. In Azevedo et al. (2008), a comparison
of the most popular iterative optimization algorithms,
namely perturb & observe (P&O) and incremental conductance (IC), applied to single-stage three-phase PV systems
was reported. However, neither the PV array nor VSI
model are properly modeled. As Fig. 2 shows, power vs

Fig. 2. Typical PV performance curves for dierent values of G


(G1 < G2 < G3 < G4 < G5 ) (top), and T (T 1 < T 2 < T 3 < T 4 < T 5 )
(bottom).

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

voltage (PV) curves for a PV array change with variations


in the solar irradiance, G, and temperature, T. It is well
documented in the literature that P&O- and IC-based algorithms perform poorly when these PV curves are moving
(Kadri et al., 2011; Masters, 2013; Maity and Sahu, 2016),
which translates into an increased diculty to track the
maximum power point (MPP). A rich literature is available
on enhancements for these algorithms (Kadri et al., 2011;
Dousoky et al., 2013; Libo et al., 2007; Li, 2015; Tsang
et al., 2013; Zhang et al., 2015; Raj and Jeyakumar,
2014). However, most of these solutions are either not tailored to single-stage operation in three-phase systems or do
not consider voltage regulation on the ac side through variable reactive power injection. Alternatively, solutions
based on predictive neural networks (NNs) to overcome
this issue have also been proposed (Lin et al., 2011), but
only in the case of architectures operating at a constant
dc-link voltage.
Although it is customary (and in some countries
required) to set the power factor to unity, the VSI can synthesize an arbitrary value of reactive power and hence provide voltage regulation as well. As the installed power of
PV plants grows, grid codes will be modied to consider
the participation of PV plants in the provision of ancillary
services, such as reactive power injection. For example, the
German medium-voltage grid code introduced in 2008
(BDEW, 2008) states that if required by the network operator, PV plants must participate in voltage control. Therefore, control strategies for VSIs which enhance MPPT
algorithms while injecting variable reactive power will nd
increased deployment in the near future.
In this work, a methodology is developed for the early
detection of changes in the PV curves, enhancing the algorithms response against abrupt variations in G. Changes in
the curves are quantied through the measurement of the
dc voltage steady-state error, and a small-signal model of
the single-stage PV system. Variable reactive power injections are considered in the small-signal model as disturbances, which makes the algorithms robust when the PV
system is performing dynamic voltage regulation.
In view of the above discussion, the main contributions
of this paper are:

323

 A unied dynamic model is provided for single-stage PV


power plants including a modied PV array model. The
proposed unied model is able to reproduce the behavior of a single-stage architecture with a PV array of an
arbitrary size through a single circuital representation.
The model is also provided via a rigorous mathematical
formulation through a set of dierential algebraic equations (DAEs).
 Based on this model, a methodology is proposed to
improve iterative optimization- and prediction-based
MPPT algorithms which are able to operate in concert
with a single-stage architecture under time-varying reactive power injection.
 To perform a comparative evaluation among the proposed MPPT algorithms while operating under rapidly
changing G and time-varying reactive power injection.
The rest of the paper is organized as follows. Section 2
describes the realization of a single-stage grid-connected
PV power plant, followed by its unied modeling in terms
of DAEs and a single equivalent circuit in Section 3. The
control approach is discussed in Section 4. In Section 5,
three MPPT algorithms are proposed, and in Section 6, a
comparative evaluation is performed based on detailed
computer simulations. The conclusions of Section 7 close
the paper.

2. Single-stage PV system
The circuit schematic of Fig. 3 illustrates the power
stage of the single-stage grid-connected PV power plant
considered in this paper. It includes the cascaded connection of a dc capacitor that connects to the output terminals
of a PV array of an arbitrary size; a two-level three-phase
VSI; and a LC-lter in series with a coupling transformer
that connects to the ac grid.
As suggested in Fig. 3, the overall control is realized
through a two-layer strategy. An outer loop tracks the
PV arrays MPP utilizing one of the three MPPT strategies
proposed in this paper. An inner loop regulates the VSI in
current control mode in order to transfer the power

Fig. 3. Schematic of typical grid-connected single-stage PV power plant.

324

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

generated by the PV array into the ac grid with the capability of reactive power injection.
3. Unified dynamic model of a single-stage PV power plant
In this section a unied dynamic model to reproduce the
behavior of a grid-connected single-stage PV power plant
with a PV array of an arbitrary size is proposed. The unied model allows the representation of the overall system
through a single equivalent circuit, as shown in Fig. 4.
The circuital model for a single PV cell and its generalization to a number of cells in series is well established in
terms of a current source, an anti-parallel diode, a series
and a shunt resistance (Masters, 2013). However, this
paper makes use of a modied circuit introduced in Tian
et al. (2012) that replaces the anti-parallel diode by a controlled current source. The controlled current source is governed by a modied diode equation which allows the
generalization of the model to an arbitrary number of cells
connected in series, N S , and in parallel, N P . A single circuital representation can therefore be used to reproduce
the behavior of an arbitrary PV array including all details
of each cell. The generalized model is presented in the left
side of Fig. 4 and takes the form of




qvpv ipv R0S
vpv ipv R0S
0
0
ipv I irr  I 0 exp
;
1 
N S nkT
R0P
1
where q 1:602  1019 C is the electrons electric charge,
k 1:3806503  1023 J/K is the Boltzmann constant, and
n is the ideality factor or the ideal constant of the diode. The
various ``prime values are related to their cell-specic counterparts according to
I 0irr : N P I irr ; I 00 : N P I 0 ; R0S :

NS
NS
RS ; R0P :
RP ;
NP
NP

where I irr is the photo current or irradiance current


generated when the cell is exposed to sunlight; I 0 is the
diode saturation current or cell reverse saturation current;
and RP and RS represent the cell shunt and series resistance,
respectively (Masters, 2013). The cell-specic parameters
and n depend on G; T , and certain reference parameters
measured at standard reference condition (SRC), namely

Gref ; T ref ; I irr;ref ; I 0;ref ; RP;ref ; RS;ref , and nref , as described


by (2)(6),
G
I irr;ref aT T  T ref ;
Gref

3


T
Eg;ref Eg
;
exp

I 0 I 0;ref
T ref
kT ref kT


Gref
RP RP;ref
;
G

RS RS;ref ;

n nref :

I irr

In (2), aT is the absolute temperature coecient of the


short-circuit current, which represents the rate of change
of the short-circuit current with respect to T. In (3), Eg is
the bandgap energy for silicon in eV. Analytical expressions for Eg are obtained through curve tting from experimental measurements and varies from author to author.
Herein the expression from Kim et al. (2009) is used,
Eg 1:17  4:73  104 

T2
:
T 636

The IV curves described by (1) plus the breakdown of the


various constituent parameters described by (2)(7) provide a detailed representation for the performance of a
PV array. The reference parameters at SRC are PV cellspecic and may be obtained from information available
on manufacturerss datasheets following the methodology
introduced in Tian et al. (2012). A PV power plant contains
a large number of cells in series and in parallel in order to
match particular levels of voltage and power. For instance,
if 0.5 V/2 W PV cells are used as a building block to realize
a 450 kW power plant, N S and N P would be in the order of
1300 and 180, respectively. The key advantage of this
model is that an arbitrary number of cells connected in
series and/or paralleland therefore a PV array with an
arbitrary congurationcan be modeled through a single
circuital representation including the details of each cell.
The cascaded connection of the dc capacitor, VSI,
LC-lter and ac grid is modeled in the dq reference frame
using dynamic phasors. The right side of Fig. 4 illustrates
the dynamic equivalent circuit in terms of controlled
sources (Yang et al., 2003; Yazdani and Iravani, 2010).

Fig. 4. Unied dynamic equivalent circuit schematic of a grid-connected single-stage PV power plant.

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

325

The equations which describe the states evolution are


given by (8)(11),


d
1
3 ~
~i ;
vpv
ipv  m
8
dt
C
4

d~
1  vpv
~
 R Rf ~i Rf~ig ~
vf ;
i jx~i
m
9
dt
L
2
d
1 ~ ~
~
vf
vf jx~
i  ig ;
10
dt
Cf



d~
1
~
11
ig jx~ig
vf  Rg Rf ~ig Rf~i ~
vg :
dt
Lg

gle-stage PV power plant. The DAE formulation can be


utilized to develop more rigorous feedback control algorithms. Furthermore, the equivalent single circuit shown
in Fig. 4 may be utilized to study PV power plants in ac systems by incorporating the circuit into power ow, optimal
power ow and dynamic simulation programs. In this work
the presented unied model is used to develop a methodology to enhance iterative MPPT algorithms for single-stage
architectures.

In (8)(11), the vector quantities relate to their abc counterparts considering a generic vector quantity ~
x xd jxq ,
with

The VSIs control scheme utilized herein corresponds to


a conventional current control (Yazdani and Iravani, 2010)
plus a variety of three proposed MPPT algorithms specically tailored to perform in concert with a single-stage
architecture. The overall scheme is presented in Fig. 5.
As observed in the gure, the control is broken down into
four blocks: (i) a PLL utilized for synchronization with the
ac grid; (ii) a dc voltage control to track the PV arrays
MPP; (iii) a power-to-current transformation; and (iv) a
conventional current control.

xd

xq  T dq
abc xa

xb

xc  ;

where T dq
abc is the Park transformation,



!
cosh
cos h  2p
cos h  4p
2
dq
3
3



:
T abc
3  sinh  sin h  2p
 sin h  4p
3
3
In (8), the operator  denotes the dot product between
~ is the
two vectors, yielding a dc value. In (8) and (9), m
VSIs modulating function. Furthermore, x and h are the
synchronizing frequency and phase angle, respectively.
The synchronizing quantities are measured at the LC-lters output and are synthesized through a phase-lockedloop (PLL) which is explained in Section 4. In (11), Lg
and Rg represent the grid inductance and resistance, respectively, as seen from the LC-lters output. These values correspond to lumping the coupling transformers impedance
and the (uncertain) Thevenin impedance of the ac grid seen
from the point of common coupling (PCC). Finally, in (11)
~
vg is the Thevenin voltage of the ac grid. Refer to Fig. 3 for
the interpretation of the remaining parameters.
The state space model readily follows from (8) plus the
six equations which result from separating (9), (9)(11) into
real and imaginary parts, yielding a seven-order state space
model. Moreover, the nonlinear relationship between vpv
and ipv established by (1) plus (2)(7) leads to a dierential
algebraic equations (DAEs) formulation. Thus, (1)(11)
provide a unied dynamic model of a grid-connected sin-

4. Control scheme

4.1. Synchronization and control of power & current


The current control is synchronized with the ac grid by
means of a simple PI-based PLL. The grid voltage is measured, transformed into the dq reference frame, and its q
component is forced to zero through a PI controller. The
PLL block also provides a measure of vd and the grid phase
angle. Further information about the PLL design and
gains tuning is readily available in the literature (Chung,
2000).
The grid current is controlled according to active and
reactive power references. The active power reference is
generated by the dc-link voltage control, whereas the reactive power reference may be supplied directly. In steadystate, the power-to-current transformation reduces to
P ref 3=2vd I d ref and Qref 3=2vd I q ref . According
to the computed current references, a conventional current
control loop generates a modulating signal through two decoupled compensators (Yazdani and Iravani, 2010).
Finally, the modulating signal in the abc domain is com-

Fig. 5. Block diagram for the VSI control.

326

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

Fig. 6. Operation at variable dc voltage.

pared against a triangular carrier signal at F S using doubleedge pulse width modulation (PWM) in order to generate
the IGBTs pulses.
4.2. Control of dc-link voltage
A main feature of the single-stage architecture described
in this paper is the operation at a variable dc-link voltage.
This may be explainedand justiedconsidering Fig. 6.
The gure illustrates that within most values of G and T
the MPP voltage remains within a relatively narrow range.
Only when G falls considerably (below 200 W/m2), the
MPP voltage becomes too low and thus unacceptable for
the operation of the inverter. The inverter can only buck
the voltage, hence there is a minimum voltage at
which the inverter can operate when connected to the ac
grid. In practice, inverters manufacturers design the control system specifying a range vpv 6 vpv 6 ^vpv for which
the VSI will perform MPPT.
The dc-link voltage has to follow the PV arrays MPP,
which is controlled by the MPPT algorithm through changing the dc-link voltage reference. As Fig. 5 suggests, once
the dc-link voltage reference, V pv ref , is computed, the value
is processed through a voltage limiter and fed to a PI controller yielding a power command. Alternatively, the dclink voltage control may be performed by regulating the
energy ow in and out of the capacitor controlling the
square of voltages. Energy-based controllers are reported
to feature faster dynamics than those of their voltage-based
counterparts (Yazdani and Iravani, 2010). However, both
controllers are fast enough within the MPPT timescale,
and hence the voltage-based approach is selected here for
the sake of simplicity.

5.1. Iterative optimization algorithms: P&O and IC


Tracking heuristics can be described as iterative optimization algorithms that, starting from an initial feasible
solution, progressively improve the solution by applying
a series of local modications. At each iteration, the search
moves to an improved solution that diers only slightly
from the current one. Following the methodology
proposed in this section, P&O- and IC-based heuristics
are modied to improve their MPPT capabilities when
operating in concert with single-stage architectures. Both
algorithms stem from a straightforward manipulation of
(12), leading to the following equivalent incremental
expressions,
vpv

~ipv
ipv ;
~vpv

~ipv ipv

;
~vpv vpv

5. MPPT under time-varying reactive power


The aim of a MPPT algorithm is to keep the PV array
operating at the MPP as the atmospheric conditions
change with time. As Fig. 2 suggests, the atmospheric conditions which have an eect on the MPP are G and T. The
MPP (vmpp ; impp ) can be found numerically if G and T are
measured (Zhao et al., 2015). It follows from solving (1)
plus (2)(7), and (12) which holds at the MPP:
dppv
0;
dvpv

where ppv vpv ipv is the PV arrays output power. As


observed in Fig. 2, if the PV array is not under partial shading (Daz-Dorado et al., 2014; Liu et al., 2014; Boukenoui
et al., 2016), the concave shape of the PV curve guarantees a unique voltage (for a given G and T) at which the
output power is maximum. However, G and T are usually
unavailable for measurement, and MPPT algorithms thus
rely on iterative optimization and/or prediction
approaches.
This paper proposes a methodology to enhance iterative
optimization algorithms for single-stage architectures.
These algorithms solve an optimization problem iteratively
in real time. They are able to detect a change in the MPP
due to a variation in G or T without actually sensing them.
The operation of these heuristics relies solely on the concave shape of the PV curve, which is common to all PV
cells. Therefore, they are able to track the MPP without
being trained for a specic PV cell.
On the other hand, the prediction approach proposed
herein makes use of a NN architecture whose parameters
depend on the PV curves of a PV array. The NN has to
be trained for a specic PV array in order to accomplish
the MPPT function. In this case the optimization process
takes place oine in the training of the NN. The NN-based
algorithm needs T to be either measured through a dedicated sensor or estimated by means of other measurements.

12

13
14

which are equal to zero at the MPP. Throughout the rest of


the paper, all incremental variables are dened as
~x x  xk , where k denotes the current sampling period,
x is a generic time-varying variable evaluated at the end
of the k sampling period, and xk is the value of x at the
beginning of the k sampling period. The P&O and IC algorithms are based on the evaluation of the quantities dened
by (13) and (14), respectively. When the operating point on
the PV plane is placed to the right (left) of the MPP, these
quantities are negative (positive). Therefore, based on the

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

sign of these quantities, the iterative algorithms can provide a new voltage reference as a corrective action to
approach the MPP. In the case of the P&O algorithm,
(13) corresponds to ~
ppv =~vpv , i.e. the ratio between the increment in PV arrays power and voltage. For the IC algorithm, the sign of (14) is evaluated by comparing the
incremental conductance with the instantaneous
conductance.
The error introduced by using the incremental quantities
dened in (13) and (14) depends on the voltage step size
and the operating point. The closer to the MPP, the more
nonlinear the PV curve is, and therefore the worse this
approximation is. In order to decrease the error introduced
by this approximation, a variable step approach is considered in which the voltage step size decreases as the operating point approaches the MPP. Furthermore, this
approach reduces the oscillations around the MPP, which
in turn increases eciency. Finally, it allows a quicker
response when the operating point is far from the MPP
(Lee et al., 2006).
Under abrupt changes in G these algorithms perform
slowly and may even fail, which translates into a reduced
eciency (Masters, 2013). Changes in ~ipv can be tracked
~
~ T~ , and ~vpv . These algorithms become confused if G
to G;
~
or T are signicant enough to change the sign of the quantities dened in (13) and (14). This represents the main
drawback of the standard P&O and IC algorithms and
results in a reduced extracted power under rapidly changing atmospheric conditions. In Kadri et al. (2011), a modied version of the standard P&O algorithm was proposed
to overcome this issue. The algorithm was improved by
taking advantage of the characteristics of a single-stage
architecture. Specically, by using the signal error of the
dc voltage control loop to compute the change in power
due to changing atmospheric conditions. Following this
idea, a generalization of this approach is proposed in this
paper. A small-signal model based on the linearization of
the unied dynamic model for the single-stage PV system
presented in Section 3 is used to nd an expression for
the error of the dc voltage control loop. This approach
allows the development of a methodology to (i) improve
both P&O and IC; and (ii) consider a variable reactive
power injection, which would enable voltage regulation.

327

If the sampling period is suciently short, the following


assumptions can be made: (A1) the dynamics of the dc-link
voltage can be approximated through the linearization of
Eqs. (8), (1), plus (2)(7), around the operating point at
the beginning of the sampling period, dened as
ipv k ; vpv k ; Gk ; T k ; md k ; mq k ; vd k ; (A2) T remains constant; (A3) G varies linearly at a constant rate of change,
aG ; (A4) Qref varies linearly at a constant rate of change,
aQ ; and (A5) vd remains constant. As a consequence of
(A1) and (A2), ~ipv can be computed according to
~ K V~vpv ;
~ipv ~ipv ~ipv K G G
G
V

15

where ~ipvG and ~ipvV are the changes in PV arrays current

~ and ~vpv , respectively; K G @ipv

; K V @ipv

;
due to G
@G
@V
k

and ipv is dened by (1), plus (2)(7).


Fig. 7 shows the block diagram representing the smallsignal dynamics of the dc-link voltage including the voltage
and the current control loops. The incremental error can be
expressed as
~es V~ pv

ref s

AsV~ pv

 ~vpv s

ref s

~ CsQ
~ ref s Dsm
~ d s
BsGs

~ q s;
Esm
where V~ pv ref s V~ pv ref =s, since the voltage reference
remains constant during the sampling period. As a conse~
quence of (A3) and (A4), Gs
aG =s2 , and
~ ref s aQ =s2 , respectively. On the other hand, the curQ
rent loop is designed to be faster than the voltage loop.
Therefore, the current loop transfer functions can be con~ d s; m
~ q s can be
sidered as Fid s Fiq s 1, and m
~ d s m
~ d =s; m
~ q s m
~ q =s. The
approximated to steps, m
transfer functions As to Es can be readily obtained from
Fig. 7. The steady state error at the end of one sampling
period, ~ef , can be expressed using the nal value theorem
for Laplace transforms:
ai
mq k aQ
~ef lim s~es md kpv
;
16
s!0
K
md k K i
i
2vd k
where K i is the gain used for the integral action of the
voltage control loop; and aipv is the rate of change of the

Fig. 7. Block diagram representing the small-signal dynamics of the dc-link voltage during a sampling period. Note: the block diagram is based on
assumptions (A1), (A2) and (A5).

328

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

PV arrays current caused only by the irradiation variation,


expressed as aipv K G aG . As it can be seen, only the pertur~
~ ref s contribute to steady state error.
bations Gs,
and Q
The zero-input component of ~es is zero when the
steady-state is reached. Therefore, ~ef V pv ref  vpv , where
V pv ref and vpv are evaluated at the end of the sampling period. Because ~ef can be measured, (16) can be solved for aipv ,
and ~ipvG can be computed as
~
~ipv aipv~t md k K i ef  mq k aQ ~t;
G
2vd k
where ~t is the sampling period. According to (15), once ~ipvG
is determined, quantities (13) and (14) can be evaluated
considering only ~ipvV ,
vpv

~ipv  ~ipv
G
ipv ;
~vpv

~ipv  ~ipv
ipv
G

;
~vpv
vpv

17
18

and the algorithms can determine the right direction of the


next corrective action. Fig. 8 shows a generalized owchart
of the P&O and IC iterative algorithms.
5.2. Prediction algorithm: neural network
A dierent approach is to consider the MPPT as a prediction problem. The prediction problem consists in
approximating the unknown real mapping between the
inputs T ; vpv and ipv , and the output vmpp , which is considered to be V pv ref in Fig. 5. In this section a NN-based prediction algorithm is proposed. In this approach, the MPPT
and the reactive power control are two decoupled problems, and therefore a time-varying reactive power injection
does not compromise the eciency of the MPPT function.
NNs may be classied as traditional NNs and locally recurrent NNs (LRNNs). Traditional NNs have no memory of
the past inputs: the output is strictly a function of the
instantaneous input to the network (Dounis et al., 2015).
In contrast, in a LRNN the output may also be a function

Fig. 8. Generalized iterative MPPT algorithm owchart. k is dened by


the quantities (17) and (18) for the enhanced P&O and IC iterative
algorithms, respectively. The variable voltage step size is accomplished by
multiplying the quantity k by a proportionality constant K.

of delayed inputs and delayed outputs. This allows for the


creation of richer and more complex decision surfaces.
LRNNs have been successfully applied, for example, to
the electrical load forecasting problem (Khotanzad et al.,
1997). In the MPP tracking problem, the voltage at the
MPP is only a function of the instantaneous inputs, and
therefore a traditional multi-layer feed-forward network
is well suited. The optimization problem appears in the
design of the NN which relies on choosing the number of
layers, the number of neurons in the hidden layers, and
the appropriate interconnections. The objective function
to be minimized is the error on a test data set, reached after
a number of iterations or training epochs. In the general
case, all these parameterssome of which are discrete in
natureshould be considered in the optimization problem,
leading to a mixed integer programming problem. Metaheuristics such as simulated annealing, genetic algorithms
and tabu search have been proposed to solve this
optimization problem (Glover and Marti, 2006). In the
case of a traditional multi-layer feed-forward network,
the optimization problem can be simplied. Architectures
with a single hidden layer have been proved to be well
suited for the MPP tracking problem (Liu et al., 2013).
Furthermore, the number of neurons can be determined
by a trial and error process. Therefore, the only variables
left to be determined are all of continuous nature, and
the process through which are determined is called the
training process.
In this paper, a NN-based algorithm is proposed with a
two-fold objective: (i) to provide a continuous step-free
voltage reference to feed the inverters voltage control loop
therefore improving the dynamic across the dc capacitor,
and (ii) to track the MPP under changing G and T while
operating at a variable dc-link voltage. As illustrated in
Fig. 9, the considered NN features a three-layer structure:
3 input neurons; 8 neurons for the hidden layer; and 1 output neuron. The methodology makes use of the modied
PV array model described by (1) and (2)(7), for nding
a suitable matrix of synaptic weights. The backpropagation
(BP) method is an eective algorithm based on the steepest
descent direction and it is the most widely used optimization method for accomplishing the training of the NN
(Glover and Marti, 2006). The NN has to be trained to
predict vmpp for each characteristic curve of the PV array
at particular G and T. In this work, the training process
was executed using the LevenbergMarquardt BP

Fig. 9. Structure of the considered NN.

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

algorithm (Glover and Marti, 2006), and the goal for mean
square error (MSE) on the training set was xed equal to
107 . Fig. 10 shows the MSE during the NN training. After
33 epochs the training process is stopped.
6. Simulation results
In order to validate the control scheme proposed in Section 4 and evaluate the performance of the three MPPT
algorithms explained in Section 5, a detailed computer simulation in PSCAD/EMTDC of a single-stage 450 kW gridconnected PV system was implemented. The system is
designed according to a typical 1000 V dc architecture.
The unied dynamic model of Section 3 along with the control scheme of Section 4 were simulated considering real G
and T data collected on PV panels in the city of Denver,
Colorado, USA. The collected data is presented in the
top and middle plots of Fig. 11. Measurements were collected every 800 ms over 40 s on a partially cloudy day,
which helps evaluate the performance of the MPPT control
schemes under shading conditions. The PV array is realized
using 40 panels in series and 150 in parallel. The simulated
panel corresponds to the model KC85TS manufactured by
Kyocera, with datasheet MPP at 17.4 V/87 W at SRC. The
panel was characterized and tested in the laboratory using
the procedure from Tian et al. (2012) and the actual MPP
at STC was found to be 16.67 V/75.93 W. The panel features two parallel strings of 36 cells in series. Therefore,
in the simulation the total number of cells in series, N S , is

Fig. 10. MSE during the NN training.

Fig. 11. Actual G and T proles, and reactive power reference used during
the simulation.

329

1440 and the total number of cells in parallel, N P , is 300.


The utility grid is rated at 12.47 kV/60 Hz, with a short circuit power of 100 MVA and a unity R/X ratio. The coupling transformer is rated at 600 kVA with a
transformation ratio of 270 V/12.47 kV and an impedance
of 6%. The lter inductor, capacitor and damping resistance were selected to be L 38:675 lH, C f 363:87 lF
and Rf 64:2 mX respectively. The VSIs switching frequency is F S 6 kHz, and the dc-link capacitor was
selected 5000 lF. Finally, the xed sampling period for
the iterative MPPT algorithms was set to 0.2 s. In the simulation, a time-varying reactive power injection was considered, following a ramp starting with unity power factor and
reaching 100 kVAr after 40 s, as shown in the bottom plot
of Fig. 11. According to Figs. 3 and 5, the reactive power
reference for the two-layer control can be computed as
Qref Qinj  Qc , where Qinj is the desired reactive power
injection before the coupling transformer, and Qc is the
reactive power injected by the lter capacitors. The latter
can be calculated as Qc 1:5v2d xC f , assuming that the
PLL forces vq 0.
The overall performance of the MPPT algorithms is
illustrated in Fig. 12. The voltage plots show: (i) the voltage
reference dictated by the algorithms; (ii) actual dc-link voltage; and (iii) the trajectory of the theoretical voltage to
extract the maximum power. The dierence between the
theoretical MPP voltage and the actual dc-link voltage is
clearly lower in the case of the enhanced algorithms, while
the NN-based algorithm outperforms the iterative algorithms. It is noteworthy that the NN performs the MPPT
function using only one measurement and in one computation step. The response time is determined solely by the frequency at which vpv ; ipv and T are sensed, and hence in
practice the NN will track the MPP in a continuous manner. This is in contrast to the P&O and IC algorithms,
which require a number of measurements and computational steps to iteratively approach the voltage at the
MPP, never settling on the MPP voltage. Each measurement-computational step pair generates a dc voltage which
translates into a step-like reference that is applied to the
voltage control loop. The VSI dynamic response to changes
on the voltage reference dictated by the MPPT algorithms
is illustrated in the zoomed areas of the voltage trajectories.
From the gure it can be seen that the NN algorithm provides a smooth reference to the dc-link voltage control loop
and the actual dc-link voltage is able to follow the reference
with an improved dynamic response. In the case of the
P&O and IC algorithms, the dc-link voltage is forced to follow a step-like reference. Fig. 12 also shows the actual
power being extracted from the array along with the theoretical maximum power available in the PV array. The difference between the theoretical and the actual power is
lower in the case of the enhanced iterative algorithms,
showing the NN algorithm the best performance. This is
reected in the corresponding tracking factors (TF),
dened as the ratio between the actual amount of energy

330

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331

Fig. 12. Simulation results showing the performance of the enhanced P&O, standard P&O, enhanced IC, standard IC, and the NN-based predictive
algorithm (from left to right). The interpretation of the legend is as follows. V pv ref : voltage reference dictated by the algorithm; vpv : actual dc-link voltage;
vmpp : theoretical voltage at the MPP; P dc : actual extracted power; P mpp : theoretical power at the MPP. Di is the normalized dierence between the
theoretical voltage (power) at the MPP and the actual dc-link voltage (extracted power). The TF values for the enhanced and standard P&O are 99.74%
and 99.29%, respectively. The TF values for the enhanced and standard IC are 99.77% and 99.31%, respectively. Finally, the NN-based predictive
algorithm features a TF value of 99.99%.

extracted from the PV panel, and the theoretically available


power. See Fig. 12.
7. Conclusions
A unied dynamic model for single-stage PV power
plants including a modied PV array model was presented.
The unied model is able to reproduce the behavior of any
single-stage PV power plant with an arbitrary PV array
conguration through a single circuital representation. A
methodology based on this unied model was developed
to enhance iterative MPPT algorithms for single-stage
architectures. The methodology increases the MPPT eciency under rapidly changing atmospheric conditions
while considering time-varying reactive power injections.
This, in turn, would enable voltage regulation without
compromising the MPPT performance. Following this
methodology, improved P&O- and IC-based heuristics
were presented. Furthermore, a NN-based prediction
MPPT algorithm was proposed to improve the dynamic
of the dc-link voltage. Finally, a performance evaluation
of the three MPPT strategies was carried out based on
detailed computer simulations. Although the NN-based
algorithm outperforms the iterative algorithms in tracking
the theoretical voltage at the MPP, all of them feature a
very similar performance in terms of extracted power. It
can be therefore concluded that, if properly designed and
parameterized, any of the three aforementioned algorithms
can be utilized for the MPPT operation of single-stage
three-phase PV systems under time-varying reactive power
injection. The authors are currently working on augmenting the proposed control strategies for operation under

abnormal grid operating conditionssuch as sustained


voltage unbalance or ickerwhich may reveal further differences among the proposed MPPT strategies. Results will
be reported as they become available.
References
Azevedo, G.M.S., Cavalcanti, M.C., Oliveira, K.C., Neves, F.A.S., Lins,
Z.D., 2008. Evaluation of maximum power point tracking methods for
grid connected photovoltaic systems. In: IEEE Power Electronics
Specialists Conference, June, pp. 14561462.
Bao, X., Zhuo, F., Tian, Y., Tan, P., 2013. Simplied feedback
linearization control of three-phase photovoltaic inverter with an
LCL lter. IEEE Trans. Power Electron. 28 (6), 27392752.
BDEW, 2008. BDEW Technical Guideline. Generating Plants Connected
to the Medium-Voltage Network. <http://http://www.bdew.de/>.
Boukenoui, R., Salhi, H., Bradai, R., Mellit, A., 2016. A new intelligent
MPPT method for stand-alone photovoltaic systems operating under
fast transient variations of shading patterns. Sol. Energy 124, 124142.
Cecati, C., Ciancetta, F., Siano, P., 2010. A multilevel inverter for
photovoltaic systems with fuzzy logic control. IEEE Trans. Industr.
Electron. 57 (12), 41154125.
Chen, P.-C., Chen, P.-Y., Liu, Y.-H., Chen, J.-H., Luo, Y.-F., 2015. A
comparative study on maximum power point tracking techniques for
photovoltaic generation systems operating under fast changing environments. Sol. Energy 119, 261276.
Chung, S., 2000. A phase tracking system for three phase utility interface
inverters. IEEE Trans. Power Electron. 15, 431438.
de Brito, M.A.G., Galotto, L., Sampaio, L.P., de Azevedo e Melo, G.,
Canesin, C.A., 2013. Evaluation of the main MPPT techniques for
photovoltaic applications. IEEE Trans. Industr. Electron. 60 (3),
11561167.
Daz-Dorado, E., Cidras, J., Carrillo, C., 2014. Discrete IV model for
partially shaded PV-arrays. Sol. Energy 103, 96107.
Dounis, A.I., Konas, P., Papadakis, G., Alafodimos, C., 2015. A direct
adaptive neural control for maximum power point tracking of
photovoltaic system. Sol. Energy 115, 145165.

M. Carrasco, F. Mancilla-David / Solar Energy 132 (2016) 321331


Dousoky, G.M., Ahmed, E.M., Shoyama, M., 2013. MPPT schemes for
single-stage three-phase grid-connected photovoltaic voltage-source
inverters. In: IEEE International Conference on Industrial Technology, February, pp. 600605.
Esram, T., Chapman, P.L., 2007. Comparison of photovoltaic array
maximum power point tracking techniques. IEEE Trans. Energy
Convers. 22 (2), 439449.
Figueres, E., Garcera, G., Sandia, J., Gonzalez-Espin, F., Rubio, J.C.,
2009. Sensitivity study of the dynamics of three-phase photovoltaic
inverters with an LCL grid lter. IEEE Trans. Industr. Electron. 56
(3), 706717.
Glover, F., Marti, R., 2006. Metaheuristic Procedures for Training
Neutral Networks. Springer, New York.
Huang, Y., Shen, M., Peng, F.Z., Wang, J., 2006. Z-source inverter for
residential photovoltaic systems. IEEE Trans. Power Electron. 21 (6),
17761782.
Jain, S., Agarwal, V., 2007. Comparison of the performance of maximum
power point tracking schemes applied to single-stage grid-connected
photovoltaic systems. IET Electr. Power Appl. 1 (5), 753762.
Kadri, R., Gaubert, J.P., Champenois, G., 2011. An improved maximum
power point tracking for photovoltaic grid-connected inverter based
on voltage-oriented control. IEEE Trans. Industr. Electron. 58 (1), 66
75.
Kasa, N., Iida, T., Chen, L., 2005. Flyback inverter controlled by
sensorless current MPPT for photovoltaic power system. IEEE Trans.
Industr. Electron. 52 (4), 11451152.
Khotanzad, A., Afkhami-Rohani, R., Lu, T.-L., Abaye, A., Davis, M.,
Maratukulam, D.J., 1997. ANNSTLF a neural-network-based
electric load forecasting system. IEEE Trans. Neural Netw. 8 (4),
835846.
Kim, S.K., Jeon, J.H., Cho, C.H., Kim, E.S., Ahn, J.B., 2009. Modeling
and simulation of a grid-connected PV generation system for electromagnetic transient analysis. Sol. Energy 83 (5), 664678.
Lee, J.H., Bae, H., Cho, B.H., 2006. Advanced incremental conductance
MPPT algorithm with a variable step size. In: International Power
Electronics and Motion Control Conference, September, pp. 603607.
Li, S., 2015. A variable-weather-parameter optimization strategy to
optimize the maximum power point tracking speed of photovoltaic
system. Sol. Energy 113, 113.
Libo, W., Zhengming, Z., Jianzheng, L., 2007. A single-stage three-phase
grid-connected photovoltaic system with modied MPPT method and
reactive power compensation. IEEE Trans. Energy Convers. 22 (4),
881886.
Lin, W.-M., Hong, C.-M., Chen, C.-H., 2011. Neural-network-based
MPPT control of a stand-alone hybrid power generation system. IEEE
Trans. Power Electron. 26 (12), 35713581.
Liu, Y.-H., Liu, C.-L., Huang, J.-W., Chen, J.-H., 2013. Neural-networkbased maximum power point tracking methods for photovoltaic
systems operating under fast changing environments. Sol. Energy 89,
4253.
Liu, Y.-H., Chen, J.-H., Huang, J.-W., 2014. Global maximum power
point tracking algorithm for PV systems operating under partially
shaded conditions using the segmentation search method. Sol. Energy
103, 350363.

331

Maity, S., Sahu, P.K., 2016. Modeling and analysis of a fast and robust
module-integrated analog photovoltaic MPP tracker. IEEE Trans.
Power Electron. 31 (1), 280291.
Masters, G.M., 2013. Renewable and Ecient Electric Power Systems.
John Wiley & Sons, New Jersey.
Mastromauro, R.A., Liserre, M., DellAquila, A., 2012. Control issues in
single-stage photovoltaic systems: MPPT, current and voltage control.
IEEE Trans. Industr. Inf. 8 (2), 241254.
Messo, T., Jokipii, J., Puukko, J., Suntio, T., 2014. Determining the value
of DC-link capacitance to ensure stable operation of a three-phase
photovoltaic inverter. IEEE Trans. Power Electron. 29 (2), 665673.
Nousiainen, L., Puukko, J., Maki, A., Messo, T., Huusari, J., Jokipii, J.,
Viinamaki, J., Lobera, D.T., Valkealahti, S., Suntio, T., 2013.
Photovoltaic generator as an input source for power electronic
converters. IEEE Trans. Power Electron. 28 (6), 30283038.
Raj, J.S.C.M., Jeyakumar, A.E., 2014. A two stage successive estimation
based maximum power point tracking technique for photovoltaic
modules. Sol. Energy 103, 4361.
Sahan, B., Vergara, A.N., Henze, N., Engler, A., Zacharias, P., 2008. A
single-stage PV module integrated converter based on a low-power
current-source inverter. IEEE Trans. Industr. Electron. 55 (7), 2602
2609.
Subudhi, B., Pradhan, R., 2013. A comparative study on maximum power
point tracking techniques for photovoltaic power systems. IEEE
Trans. Sustain. Energy 4 (1), 8998.
Tian, H., Mancilla-David, F., Ellis, K., Muljadi, E., Jenkins, P., 2012. A
cell-to-module-to-array detailed model for photovoltaic panels. Sol.
Energy 86 (9), 26952706.
Tsang, K.M., Chan, W.L., Tang, X., 2013. PLL-less single stage gridconnected photovoltaic inverter with rapid maximum power point
tracking. Sol. Energy 97, 285292.
Yang, Y., Kazerani, M., Quintana, V.H., 2003. Modeling, control and
implementation of three-phase PWM converters. IEEE Trans. Power
Electron. 18 (3), 857864.
Yazdani, A., Dash, P.P., 2009. A control methodology and characterization of dynamics for a photovoltaic (PV) system interfaced with a
distribution network. IEEE Trans. Industr. Electron. 24 (3), 1538
1551.
Yazdani, A., Iravani, R., 2010. Voltage-Sourced Converters in Power
Systems. Wiley-IEEE Press, New Jersey.
Yazdani, A., Di Fazio, A.R., Ghoddami, H., Russo, M., Kazerani, M.,
Jatskevich, J., Strunz, K., Leva, S., Martinez, J.A., 2011. Modeling
guidelines and a benchmark for power system simulation studies of
three-phase single-stage photovoltaic systems. IEEE Trans. Power
Delivery 26 (2), 12471264.
Zhang, F., Maddy, J., Premier, G., Guwy, A., 2015. Novel current sensing
photovoltaic maximum power point tracking based on sliding mode
control strategy. Sol. Energy 118, 8086.
Zhao, J., Zhou, X., Ma, Y., Liu, W., 2015. A novel maximum power point
tracking strategy based on optimal voltage control for photovoltaic
systems under variable environmental conditions. Sol. Energy 122,
640649.

S-ar putea să vă placă și