Energy Conversion and Management 48 (2007) 210–218

www.elsevier.com/locate/enconman

DSP implementation of a PV system with GA-MLP-NN based

MPPT controller supplying BLDC motor drive
R. Akkaya *, A.A. Kulaksız, Ö. Aydoğdu
Department of Electrical and Electronics Engineering, Faculty of Engineering and Architecture, Selcuk University, 42031 Konya, Turkey

Received 14 January 2005; received in revised form 20 October 2005; accepted 28 April 2006
Available online 27 June 2006

Abstract

This paper presents a brushless dc motor drive for heating, ventilating and air conditioning fans, which is utilized as the load of a
photovoltaic system with a maximum power point tracking (MPPT) controller. The MPPT controller is based on a genetic assisted,
multi-layer perceptron neural network (GA-MLP-NN) structure and includes a DC–DC boost converter. Genetic assistance in the neu-
ral network is used to optimize the size of the hidden layer. Also, for training the network, a genetic assisted, Levenberg–Marquardt
(GA-LM) algorithm is utilized. The off line GA-MLP-NN, trained by this hybrid algorithm, is utilized for online estimation of the volt-
age and current values in the maximum power point. A brushless dc (BLDC) motor drive system that incorporates a motor controller
with proportional integral (PI) speed control loop is successfully implemented to operate the fans. The digital signal processor (DSP)
based unit provides rapid achievement of the MPPT and current control of the BLDC motor drive. The performance results of the sys-
tem are given, and experimental results are presented for a laboratory prototype of 120 W.
 2006 Elsevier Ltd. All rights reserved.

Keywords: Photovoltaics; MPPT; DSP; Artificial neural networks; Genetic algorithms; Brushless dc motors

1. Introduction tems feeding brushless DC (BLDC) motor loads, despite

these motors’ favorable features such as high reliability,
In sun rich countries like Turkey, the abundance and high efficiency, less maintenance requirements and reduced
accessibility of photovoltaic (PV) energy make it one of environmental effects. In various contributions, brushed
the most promising energy alternatives. Since PV electricity DC motor or induction motor loads are considered. The
is expensive compared to electricity from the utility grid, performance of DC motors supplied from PV sources has
stand alone applications of PV energy are more favorable. been analyzed by Appelbaum [1]. Two low cost PV utiliza-
They are particularly utilized for applications like powering tion schemes for ventilation and air conditioning loads
residential appliances, electrification of villages in rural have been presented in Ref. [2]. In that study, permanent
areas, telecommunications, heating, ventilating and air magnet DC (PMDC) type and AC induction type motors
conditioning (HVAC) applications, water pumping etc. have been employed. Vector control of an induction motor
Although HVAC applications and PV energy show a com- fed by a PV generator is presented in Ref. [3]. Our work
plementary nature as the trends of supply and demand demonstrates that compared with other studies in the field,
loads match with time of the day, there is not enough liter- the use of a BLDC motor, which exhibits the highest effi-
ature about HVAC applications of PV energy. Also, it ciency among all conventional motors, provides an effective
seems that not enough attention has been given to PV sys- demand side energy management technique.
Because the energy conversion efficiency of PV genera-
*
Corresponding author. Tel.: +90 332 2232230/2233724; fax: +90 332
tors is low, a proper matching between the PV generators
2410635. and the electric load should be considered. Therefore, the
E-mail address: akkaya@selcuk.edu.tr (R. Akkaya). coupling between the motor load and the PV module is

0196-8904/$ - see front matter  2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.enconman.2006.04.022
 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218 211

implemented via a maximum power point tracking (MPPT) Marquardt (GA-LM) method. The use of genetic assis-
controller to operate the PV system at its maximum output tance helps in optimizing the size of the hidden layer of
power for any temperature and solar radiation level. Vari- the NN. By adopting an evolutionary algorithm, imple-
ous MPPT methods have been considered in PV power menting such a hybrid system helps use the advantages of
applications. Among the hill climbing methods [4–6], the both the Levenberg–Marquardt (LM) algorithm, which is
perturb and observe (P&O) method tracks the maximum a higher order adaptive algorithm for minimizing the mean
power point (MPP) by repeatedly increasing or decreasing square error (MSE) of a NN, and the robustness, speed,
the output voltage at the MPP of the PV module. The efficiency and flexibility features of genetic algorithms
implementation of the method is relatively simple, but it (GA). Since our neural network implementation is run in
cannot track the MPP when the irradiance varies quickly a recall mode, the complexity of the training algorithm
with time. In addition, it may cause system oscillation and the disadvantage of slow training cause no difficulties.
around the peak power points due to the effect of measure- This study aims to show that employing an effective
ment noise. The incremental conductance method is also MPPT controller and replacing a conventional motor like
often used in PV systems [7,8]. This method tracks the the brushed DC or induction motor with a BLDC motor
MPPs by comparing the incremental and instantaneous can accomplish a considerable energy saving in a stand
conductances of the solar array. This method requires alone PV power system operating HVAC fans. The overall
more conversion time, and a large amount of power loss diagram of the implemented system is shown in Fig. 1.
results. In addition, extra hardware circuitry is required Solar insolation is converted to DC electrical power, and
to implement the system. Neural network (NN) based the MPPT controller output supplies the BLDC motor
MPPT techniques have been proposed [9–11]. Compared drive to operate the fans. In the MPPT controller, the
to the conventional methods, these studies show that this GA-MLP-NN is used, and the system robustness and
modern control algorithm is capable of improving the insensitivity to intermittent weather conditions is
tracking performance. The method requires little or no enhanced. The BLDC motor controller is used to decode
knowledge of the PV parameters, and the trained NN position data coming from the Hall effect sensors and gen-
can provide a sufficiently accurate MPPT. A PC based erates switching signals for the three phase power stage to
MPP tracker for a PV system using NNs has been devel- electronically commutate the motor. The speed controller
oped by Ref. [9]. The system consisted of a PV module with also implements a simple proportional integral (PI) control
a MPP tracker and a DC motor driving an air fan. Solar loop. The power stage combines a three phase inverter cir-
array modeling and MPP tracking is studied in Ref. [10], cuit. The implementation of the overall system is based on
and two NNs, which are back propagation NN and radial a DSP unit. The matching of the supply and demand loads
basis function NN, are compared. It is shown that the back with time of the day makes the system widely applicable.
propagation NN needed less information for training. The
use of radial basis function networks (RBFN) for solar 2. Characteristics of PV array
array modeling and MPP prediction is presented in Ref.
[11]. It shows that an RBFN model can be trained faster The basic element of a PV system is the solar cell. A typ-
than a multi-layer perceptron (MLP) model. However, ical solar cell consists of a p–n junction formed in a semi-
the DSP implementation of this network is more difficult. conductor material similar to a diode. As shown in
In our study, the proposed MPPT technique utilizes Fig. 2, the equivalent circuit model of a solar cell consists
the genetic assisted, multi-layer perceptron neural network of a current generator (IL) and a diode plus series (Rs)
(GA-MLP-NN) trained by the genetic assisted, Levenberg– and parallel resistances (Rsh) [12].

Fig. 1. The diagram of the implemented system.

212 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218

acteristics of the solar array is nonlinear and vitally affected

by the solar radiation, temperature and load condition.
In order to maximize the output power from a solar
module, it has to be operated at a unique point with spec-
ified voltage and current values, or in other words, at a
specified load resistance. This requires a separate power
converter circuit for the MPPT. In our design, a boost type
DC–DC converter is employed to match the load to the PV
Fig. 2. The equivalent circuit of a solar cell. array to extract the maximum power.

3. Boost type DC–DC converter

The current–voltage characteristic of a single cell when
neglecting the internal shunt resistance is described by the In Fig. 3(a), the schematic of the boost converter power
Shockley solar cell equation [13]: stage is given. It consists of the power switch T, boost
    inductor L, filter capacitor C, output diode D and load
QV oc
I g ¼ I L  I sat exp 1 ð1Þ resistor R. Here, the DC input source voltage Vg is supplied
AKT
by the PV modules. The converter steady state waveforms
where Ig is the output current, IL is the generated current in the continuous conduction mode (CCM) where the
under a given insolation, Isat is the diode saturation cur- inductor current flows continuously (iL > 0) are presented
rent, Voc is the open circuit voltage, K is Boltzmann’s con- in Fig. 3(b).
stant, Q is the electron charge, T is the temperature (K) and When the switch T is in the on state, the current in the
A is the ideality factor for a p–n junction. The saturation boost inductor increases linearly, and at that time, the
current (Isat) of the solar array varies with temperature diode is in the off state. When the switch T is turned off,
according to the following equation: the energy stored in the inductor is released through the
 3    diode to the output RC circuit. The pulsating current pro-
T QEGO 1 1 duced by the switching action is smoothed by the capacitive
I sat ¼ I r exp  ð2Þ
Tr KA T r T filter and a DC voltage is provided to the load. The boost
converter transfer function is obtained by considering its
where Tr is the reference temperature, Ir is the saturation steady state operation [14]. The DC voltage transfer func-
current at Tr, T is the temperature of the solar array (K) tion is
and EGO is the band gap energy of the semiconductor used
in the solar array. IL in Eq. (1) is a function of incident so- Vd 1
¼ ð4Þ
lar radiation and cell temperature and is given as Vg 1D

S Assuming a lossless circuit, Pg = Pd

I L ¼ ½I scr þ K I ðT  25Þ ð3Þ Id
100 V dI d ¼ V gI g ) ¼ ð1  DÞ ð5Þ
Ig
where Iscr is short circuit current at Tr, KI is the short cir-
cuit current temperature coefficient and S is the insolation where Vg and Ig is converter input voltage and current (the
in mW/cm2. Eqs. (1)–(3) can be applied to simulate the output voltage and current of the PV array), respectively,
characteristics of solar cells. It is seen that the output char- Vd and Id is the output voltage and current, respectively,

Fig. 3. Boost converter: (a) power stage schematic; (b) waveforms.

 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218 213

and D is the duty ratio of the controllable switch. The topology (number of hidden layers, number of hidden
boost converter operates in the CCM for L > Lmin where nodes, interconnection pattern) for the ANN, which, in
2 turn, is trained using some training scheme [18]. The GA
ð1  DÞ DR
Lmin ¼ ð6Þ combines selection, crossover and mutation operators with
2f the goal of finding the best solution to a problem by search-
Here, f is the switching frequency and R is the load resis- ing until the specified criterion is met. The solution to a
tance. The minimum value of the filter capacitance that re- problem is called a chromosome, which is composed of a
sults in voltage ripple Vr is given by collection of genes. In hybrid neuro-genetic applications,
DV d genes are the NN parameters to be optimized. The GA cre-
C min ¼ ð7Þ ates an initial population and then evaluates this popula-
V r Rf
tion by training a NN for each chromosome. It then
As the name of the converter suggests, the output volt- evolves the population through multiple generations in
age is always greater than the input voltage. Therefore, in the search for the best network parameters. In our study,
our design, the boost converter is used to increase the volt- genetic assistance is used to optimize the number of hidden
age magnitude for the BLDC motor drive bus voltage and neurons. Also, a hybrid genetic assisted LM algorithm is
to control the MPPT. The control signal of the converter is accomplished to train the MLP-NN. Thus, it reduces the
obtained from the GA-MLP-NN based MPPT controller, chance of converging to local minima in which other algo-
of which the neural network structure and training algo- rithms might get trapped, and the global optimum of the
rithm is explained in the following section. problem can be approached with higher probability. This
training algorithm was compared with separately imple-
4. GA-MLP-NN structure of MPPT controller mented standard BP algorithm with delta bar delta (BP-
DBD) and LM algorithms.
An artificial neural network (ANN) is a massively paral- The configuration of the implemented NN to predict the
lel distributed processor that has a natural tendency for MPPs of the PV array is shown in Fig. 4. The network has
storing experimental knowledge and making it available three layers, i.e. an input, a hidden and an output layer.
for use [15]. The power of ANNs in system identification The input layer has two nodes for the irradiation and the
and development of adaptive controllerss makes them well temperature. The number of nodes in the hidden layer
suited to PV energy applications such as identification of has been determined by genetic assistance. The best result
the optimal operating point of PV arrays. In engineering has been obtained for 8 nodes. The output layer has two
applications, a multi-layer perceptron network trained by nodes, and the voltage (Vmp) and current (Imp) at the
the back propagation (MLP-BP) method is the most widely MPP are the outputs of the NN. In this feed forward
utilized technique. In MLPs, the units are structured into NN consisting of a single hidden layer, all the layers of
ordered layers, and connections are allowed only between the NN have a hyperbolic tangent transfer function.
adjacent layers in an input to output sense. Back propaga-
tion (BP) training is a complex gradient descent algorithm. 5. BLDC motor drives
It tries to improve the performance of the NN by reducing
the total error by changing the weights along its gradient A BLDC motor is a rotating electric machine where the
[16]. Standard gradient descent algorithms use only the stator is a classic three phase stator like that of an induc-
local approximation of the slope of the performance sur- tion motor and the rotor has surface mounted permanent
face (error versus weights) to determine the best direction magnets. In this respect, the BLDC motor is equivalent
to move the weights in order to lower the error. to a reversed DC commutator motor in which the magnet
The Levenberg–Marquardt (LM) method is one of the
most appropriate higher order adaptive algorithms known
for minimizing the MSE, which is one of the most widely
used error norms in a NN. The LM algorithm is an
approximation to the Newton method used also for train-
ing ANNs. The Newton method approximates the error of
the network with a second order expression, which con-
trasts to the BP algorithm that does it with a first order
expression. The weight update equations of the LM algo-
rithm are provided in Ref. [17]. The outputs offer very good
results compared to other systems utilizing MLP networks
trained by the standard gradient descent algorithms.
GAs have been increasingly applied in ANN design in
several ways, namely topology optimization, genetic train-
ing algorithms and control parameter optimization. In
ANN topology optimization, GAs are used to select a Fig. 4. Boost converter power stage schematic.
214 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218
2 3 2 32 3
rotates while the conductors remain stationary. In the DC va R 0 0 ia
commutator motor, the current polarity is altered by the 6 7 6 76 7
4 vb 5 ¼ 4 0 R 0 54 ib 5
commutator and brushes. On the contrary, in the BLDC
vc 0 0 R ic
motor, the polarity reversal is performed by power transis- 2 3 2 3 2 3
tors switching in synchronization with the rotor position. LM 0 0 ia ea
6 7d6 7 6 7
Therefore, BLDC motors often incorporate either internal þ 40 LM 0 5 4 i b 5 þ 4 eb 5
dt
or external position sensors to sense the actual rotor posi- 0 0 LM ic ec
tion [19].
ð8Þ
BLDC motor drives require variable frequency, variable
amplitude excitation that is usually provided by a three where va, vb, vc are the phase voltages, ia, ib, ic are the phase
phase, full bridge inverter as shown in Fig. 5(a). The inver- currents, ea, eb, ec are the phase back EMF waveforms, R is
ter is usually responsible for both the electronic commuta- the phase resistance, L is the self inductance of each phase
tion and current regulation [20]. In Fig. 5(b), the back and M is the mutual inductance between any two phases.
EMF and phase current waveforms for the three phase The electromagnetic torque is given by
BLDC motor with 120 bipolar current are shown [20]. T e ¼ ðea ia þ eb ib þ ec ic Þ=wr ð9Þ
The pulse width modulated (PWM) inverter topology where wr is the mechanical speed of the rotor. The equation
has a six switch voltage source configuration with constant of motion is
dc link voltage (Vd), which is identical with induction
motor drives and permanent magnet AC motor drives. d
wr ¼ ðT e  T L  Bwr Þ=J ð10Þ
The analysis is based on the following assumptions for sim- dt
plification [21]: where TL is the load torque, B is the damping constant and
J is the moment of inertia of the drive. The electrical fre-
• the motor is not saturated, quency is related to the mechanical speed by
• the stator resistances of all the windings are equal, and
we ¼ ðP =2Þwr ð11Þ
the self and mutual inductance are constant,
• the power semiconductor devices in the inverter are ideal where P is the number of motor poles.
• iron losses are negligible. Fig. 6 describes the basic building blocks of the BLDC
motor drive. The drive system consists of the PI speed con-
Under the above assumptions, a BLDC motor can be troller, reference current generator, PWM current control-
represented as [20] ler, Hall-IC position sensor, 3 phase inverter and the

Fig. 5. (a) Configuration of BLDC motor drive system and (b) back EMF pattern and reference current generation.

Fig. 6. Block diagram of BLDC motor drive system.

 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218 215

motor-load union. The speed of the motor is compared Table 2

with its reference value, and the speed error is processed Parameters of BLDC motor
in the PI speed controller. The output of this controller is Parameters Value Dimension
considered as the reference torque. A limit is put on the R 0.348 X
speed controller output depending on the permissible max- L 0.314 MH
imum winding currents. The reference current generator Ke 0.0419 V/(rad/s)
J 19.106 kg m2
block generates the three phase reference currents using B 0.000 N m/s
the limited peak current magnitude decided by the control- Kt 0.0419 kg m/A
ler and the position sensor. The PWM current controller
regulates the winding currents (ia, ib, ic) within the small
band around the reference currents. The motor currents active temperature compensation, and its irradiance mea-
are compared with the reference currents, and switching surement error with temperature compensation is ±6%,
commands are generated to drive the inverter devices [22]. and the temperature measurement accuracy at 25 C is
Because of its simple structure and easy implementation, ±1.5 C. The power switch of the boost converter is an
the PI controller is widely used in industrial control sys- IGBT device rated at 200 V, 12 A, and the fast recovery
tems. A speed error signal (e) is input to the PI controller, diode has 200 ns reverse recovery time. The calculated out-
and in the digital control system, an error expression for put capacitor value is 470 lF, and the boost inductor value
the nth sample is determined by the following: is 350 lH. The BLDC motor, which is coupled to the ven-
eðnÞ ¼ wref ðnÞ  wr ðnÞ ð12Þ tilating fan, is a 24 V, 106 W, 8 pole BLDC motor. The
parameters of the BLDC motor are given in Table 2.
The PI controller output value in terms of torque is In implementation of the MPPT controller, the control
given by signal for the DC–DC boost converter switching device is
T ðnÞ ¼ T ðn  1Þ þ K p feðnÞ  eðn  1Þg þ K I feðnÞg ð13Þ determined by the implemented GA-MLP-NN structure.
Genetic assistance in the MLP-NN is firstly used to deter-
where e(n) is the current error signal, e(n  1) is the previ- mine the number of hidden nodes in the hidden layer. The
ous error value, Kp is the proportional gain and KI is the GA parameters for training are Pm = 0.01, Pc = 0.9, Pop
integral gain. These constants are dependent on the process Size = 20, Maximum Generations = 100, Selection
and are determined in compliance with the desired starting type = Roulette, Selection basis = Rank, Progres-
and continuous time responses. sion = Generational, Crossover = One point. The best
result has been obtained for eight hidden nodes. Using
6. Simulation and experimental results eight neurons in the hidden layer, a relatively small net-
work is created. A network of this size is ideal for off line
The prototype system has been developed using the implementation on a DSP.
above explained method and tested in the laboratory. This NN topology has been trained for three different
The overall system is controlled by an ADSP-21992 Digital training algorithms to obtain the accuracies of each
Signal Processor. The DSP unit has a 160 MHz, ADSP- method. The algorithms used are BP training using the
219x DSP core and features an 8-Channel, 14-Bit, 20 delta bar delta (DBD) method, LM algorithm and genetic
MSPS ADC with On-Chip voltage reference. Also, it fea- assisted LM algorithm. The network is pre-trained by using
tures a three phase PWM generation unit with additional 400 data points randomly generated by Eqs. (1)–(3). The
auxiliary PWM outputs and three 32 bit general purpose number of training iterations is 10.000; the error goal is
timers. With the DSP based unit, the GA-MLP-NN based 1 · 1010; and the activation function of the neurons is
MPP tracking and current control of the BLDC motor the tangent hyperbolic function. The weight and bias val-
drive can be achieved rapidly. Two parallel connected PV ues obtained from off line training of these three methods
modules with a total maximum power of 120 W have been are used to validate the MPPT controller operation. In
used in the system. The specification of the PV modules Table 3, it is seen that comparing with other conventional
provided by the manufacturer is given in Table 1. methods, GA-LM training offers the best results.
For solar radiation and temperature measurements, a For testing the MPP tracking, another group of 70 ran-
silicon irradiance sensor has been used. The sensor has dom input–output data points were used. The error of the
maximum current and maximum voltage for the three
Table 1 methods are calculated and compared with the information
Specification of PV module at STC (25 C and 1 kW/m2) from the mathematical model as shown in Table 4. The
Parameters Value superiority of the estimation of the proposed GA-LM
Maximum power, Pmax 60 W hybrid algorithm is seen in Table 4.
Nominal current, IMPP 3.55 A The performance of the PV system is studied for a clear
Nominal voltage, VMPP 16.9 V sunny day with high radiation levels (obtained on 24th
Short circuit current, ISC 3.73 A August 2005) in Konya (location: 3751 0 1300 North,
Open circuit voltage, VOC 21.5 V
3234 0 1200 East, elevation: 1027 m). Fig. 7 shows the
216 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218

Table 3
The test results of the NN structures trained by three different algorithms
BP algorithm with DBD LM algorithm GA-LM hybrid algorithm
method
Vmp Imp Vmp Imp Vmp Imp
MSE 0.0216220 0.0027955 0.0008322 0.0001280 0.0004287 0.0000311
Normalized MSE 0.0196367 0.0019174 0.0007558 0.0000879 0.0003893 0.0000213
Mean absolute error 0.0982449 0.0334360 0.0171013 0.0088656 0.0096071 0.0041174
Max jabs errorj 0.7178296 0.1992776 0.1665330 0.0268288 0.1377690 0.0118656
Min jabs errorj 0.0010255 0.0013521 0.0002858 0.0002419 0.0002270 0.0001434

Table 4
The number of percent errors of maximum current and voltage predictions
using the three different algorithms
Error (%) BP algorithm LM GA-LM hybrid
with DBD algorithm algorithm
method
Vmp Imp Vmp Imp Vmp Imp
<1 6 25 32 45 55 68
1-10 45 38 37 25 14 2
11-20 14 7 1 0 1 0
21-30 2 0 0 0 0 0
31-40 1 0 0 0 0 0
41-50 1 0 0 0 0 0
50-100 1 0 0 0 0 0

Fig. 8. PV module output voltage and current waveforms (voltage: 5 V/

measured maximum power (MP), the maximum power pre- div, current: 5 A/div, time: 1 s/div).
dicted by the GA-MLP-NN based MPPT (MPGA) and the
normal operating power (NOP) (power delivered by PV
module when it is directly coupled to the load without
the MPPT). These values have been obtained for every neous PV module output voltage and current waveforms
15 minutes throughout the given day with a solar module and boost converter output voltage and current wave-
orientation of 33. It shows that NOP is much less than forms, respectively, for a step change of insolation (from
the power taken by the load in the case of using the MPPT 700 W/m2 to 900 W/m2 at 45 C).
controller and an average of 25–35% power increase is Fig. 10 shows the phase to phase voltage and phase cur-
gained by the proposed MPPT controller. rent waveforms of the BLDC motor obtained when it is
The laboratory implemented system has been tested, and coupled to an air fan and loaded from the MPPT control-
waveforms have been obtained to show the performance of ler. Fig. 11 shows the variation of rotor speed of the BLDC
the system. Figs. 8 and 9 present the variations of instanta- motor for step changes of insolation.

Fig. 7. The performance of the PV system studied for a clear sunny day with high radiation levels.
 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218 217

7. Conclusion

A prototype DSP implementation of a PV system with a

GA-MLP-NN based MPPT controller and 3 phase BLDC
motor drive has been designed and tested. From the results
acquired in a day with high radiation levels, it is seen that
the proposed MPPT controller provides an average power
increase of 25.35%. Also, the system could be operated
longer periods of time through the day by means of the
MPPT controller. The use of genetic assistance provided
optimization of the size of the hidden layer. Also, training
the NN with the GA-LM algorithm presented better results
compared with other systems trained by standard gradient
descent algorithms by reducing the chance of converging to
local minima. The presented BLDC drive system has been
Fig. 9. Boost converter output voltage and current waveforms (voltage:
10 V/div, current: 1 A/div, time: 1 s/div). successfully applied to a fan application, and from the
experimental results, it has been shown that the use of PI
controlled PWM inverter drive can smoothly control the
BLDC motor. From the results, it can be concluded that
the proposed system can provide an effective energy man-
agement technique in PV systems implemented in remote
areas.

Acknowledgement

This research is supported by Selcuk University Scien-

tific Research Projects Support Fund under contract num-
ber 2002/226.

References

[1] Appelbaum J. Starting and steady-state characteristics of DC motors

powered by solar cell generators. IEEE Trans Energy Convers
1986;1:17–24.
[2] Sharaf AM, AboulNaga MM, El Diasty R. Building-integrated solar
photovoltaic systems – a hybrid solar cooled ventilation technique for
Fig. 10. Phase to phase voltage and phase current waveforms of the hot climate applications. Renew Energy 2000;19:91–6.
BLDC motor (voltage: 20 V/div, current: 2 A/div, time: 1 ms/div). [3] Arrouf M, Bouguechal N. Vector control of an induction motor fed
by a photovoltaic generator. Appl Energy 2003;74:159–67.
[4] Koutroulis E, Kalaitzakis K, Voulgaris NC. Development of a
microcontroller-based photovoltaic maximum power point tracking
control system. IEEE Trans Power Electron 2001;16(1):46–54.
[5] Kim Y, Jo H, Kim D. A new peak power tracker for cost-effective
photovoltaic power systems. IEEE Proc Energy Convers Eng Conf
IECEC 1996;3(1):1673–8.
[6] Kuo YC, Liang TJ, Chen JF. Novel maximum power point tracking
controller for photovoltaic energy conversion system. IEEE Trans
Ind Electron 2001;48(3):594–601.
[7] Sullivan CR, Powers MJ. A high-efficiency maximum power point
tracker for photovolatic array in a solar-powered race vehicle. Proc
IEEE PESC 1993:574–80.
[8] Hussein KH et al. Maximum photovolatic power tracking: An
algorithm for rapidly changing atmospheric conditions. Proc Inst
Elect Eng 1995;142(1):59–64.
[9] Bahgat ABG, Helwa NH, Ahmad GE, El Shenawy ET. Maximum
power point tracking controller for PV systems using neural
networks. Renew Energy 2005;30:1257–68.
[10] Premrudeepreechacham S, Patanapirom N. Solar-array modelling
and maximum power point tracking using neural networks. In: Power
Tech Conference Proceedings, vol. 2, IEEE, Bologna, 2003. p. 5.
Fig. 11. Variation of rotor speed of BLDC motor for step changes on [11] Al-Amoudi A, Zhang L. Application of radial basis function
insolation (speed: 1500 rpm/div, time: 1 s/div). networks for solar-array modelling and maximum power-point
218 R. Akkaya et al. / Energy Conversion and Management 48 (2007) 210–218

prediction. IEE Proc Generat Transmission Distrib [17] Bishop CM. Neural networks for pattern recognition. Oxford,
2000;147(5):310–6. UK: Oxford Univ Press; 1995.
[12] Akkaya R, Kulaksız AA. A microcontroller-based stand-alone [18] Koza J, Rice J. Genetic generation of both the weights and
photovoltaic power system for residential appliances. Appl Energy architecture for a neural network. IEEE Int Joint Conf Neural
2004;78:419–31. Networks 1991;2:397–404.
[13] Hua C, Lin J, Shen C. Implementation of a DSP-controlled [19] Grasblum P. 3-phase BLDC motor contents control with Hall sensors
photovoltaic system with peak power tracking. IEEE Trans Ind using DSP56F80x. Motorola App Note AN1916/D, 2004.
Electron 1998;45(1):99–107. [20] Toliyat HA, Gopalarathnam T. AC machines controlled as DC
[14] Mohan N, Undeland TM, Robbins WP. Power electronics: convert- machines (Brushless DC machines/electronics). In: Skvarenina TL,
ers, applications, and design. 2nd ed. New York: John Wiley & Sons editor. The power electronics handbook. New York: CRC Press;
Inc; 1995. 2002.
[15] Haykin S. Neural networks: a comprehensive foundation. Engle- [21] Lee B, Ehsani M. Advanced simulation model for brushless dc motor
wood Cliffs: Macmillan College Publishing Company; 1994. drives. Electric Power Compon Syst 2003;31:841–68.
[16] Werbos PJ. Back propagation through time: What it does and how to [22] Singh B, Singh BP, Kumar M. PFC converter fed PMBLDC motor
do it. Proc IEEE 1990;78(10):1550–60. drive for air conditioning. IE(I) J-EL 2003;84:22–7.