Neural Network Efficiency Optimization Control of
A Variable Speed Compressor Motor Drive
Abdul Halim Mohd Yatim and Wahyu Mulyo Utomo
Energy Conversion Department -Electrical Engineering Faculty
Universiti Teknologi Malaysia
Johor Baru 813 10 Malaysia
Abstract- This paper presents a new method of improving the
energy efficiency of a Variable Speed Drive (VSD) for induction
into two categories [7,8,9,10]: loss-model-based controller
(LMC) and search controller (SC) method.
Keywords- Induction motor, efficiency optimization control,
neural network control
Basically, the LMC method uses analytical computation of
the motor losses to optimize the efficiency. The optimum flux
is determined by deriving the equation of the motor power
losses against to the flux. The previous work on LMC method
shows that the main advantage is simplificity of this method
i.e. does not require extra hardware. However, it is mandatory
that an accurate knowledge of motor parameters is known,
which change considerably with temperature, saturation and
skin effect. In real-time application, the difficulty in measuring
the motor parameters of the loss model does not permit the
implementation of the LMC [7]. In addition some of the motor
losses such as stray losses and iron losses are also very
complex to be determined.
motors. The efficiency of induction motor at low load operation
and at rated flux is very low compared when operating at full
load due to excessively increase in iron losses. To improve this
efficiency, it is essential to obtain the flux level that minimizes the
total motor losses. In this paper, the proposed controller is
designed to generate both the voltage and frequency reference
signals simultaneously. The proposed controller was simulated
for variable speed compressor application. The results obtained
clearly show that the efficiency at low speed is significantly
increased. Besides that the speed of the motor can be maintained.
The simulation results are also verified by experiment.
I. INTRODUCTION
World wide, approximately around 7000O of total electrical
energy is consumed by electric motor [1]. Around 96% of the
total electric motors are consumed by the induction motor [2].
In terms of the efficiency, operation of the induction motor at
rated flux results in good utilization of the motor iron hence
high efficiency and torque per stator ampere can be achieved
At rated flux the nominal electromagnetic torque can be
developed at all frequencies. However, at light load the motor
flux may be greater than necessary for development of required
load torque. In this condition the iron and stator copper losses
increase excessively hence the total losses become high and the
efficiency drops dramatically [2,3,4].
According to the load condition, the induction motor drive
efficiency can be increased by reducing the motor air gap flux.
In scalar control method, the flux can be indirectly controlled
by adjusting both stator voltage and frequency [5].
The main problem of the efficiency optimization control of
the induction motor drive system at variable load operation is
to obtain the optimum motor flux level that minimizes the total
motor losses and the maximum efficiency is achieved [6,7]. At
the same time it is also important to ascertain that the rotor
speed of the motor is still stable.
A number of methods have been published on efficiency
optimization control of the induction motor drive system. The
technique allowing the efficiency improvement can be divided
1 -4244-0743-5/07/$20.OO ©2007 IEEE
Search controller (SC) method is an induction motor
efficiency control technique based on the minimum input
power tracking approach. The principle of this method is that
the input power is measured and then the motor flux function is
gradually decreased to achieve the minimum input power
associated to the minimum power losses or maximum
efficiency. The previous works on the SC method show that to
achieve optimal efficiency, the flux is decremented in steps
until the measured input power for a certain load torque and
speed condition settles down to the lowest value. This method
does not require any knowledge of the motor parameters, is
completely insensitive to motor parameter variation and the
algorithm is applicable universally to any arbitrary drive
[4].Some implementations of the intelligent control method
such as fuzzy logic and neural network control for this method
have many advantages over classical search control methods
proposed in literatures [1 1, 12,13,14,15,16].
In this paper, a new neural network search controller with
real-time learning algorithm based on scalar control model is
proposed. The proposed controller is explained in Section II.
The simulation and experimental results, related to a 0.25hp
induction motor drive and dynamometer as a compressor load
prototype are given in section III. The results of the increasing
efficiency are compared to those obtained with the constant
Volt per Hertz approach. Finally, section IV summarizes the
conclusions.
1716
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on December 29, 2008 at 19:21 from IEEE Xplore. Restrictions apply.
11.
EFFICIENCY OPTIMIZATION CONTROL DESIGN
poad
A. Development ofNeural Network Efficiency Optimization
Controller
Based on the input power measurement, a direct neural
network control reference model of efficiency optimization is
developed. The block diagram of the proposed neural network
controller for efficiency optimization of a variable speed
compressor drive is presented in Fig. 1. The controller is
implemented through the use of a digital signal processor as
indicated by dashed outline in Fig. 1.
---
__ Z~~~~~~~~~~~~~~~L
------------------
+~~
(kL N )N
(3)
where: Pload is compressor load power
If efficiency of the motor drive is targeted with the
efficiency at nominal speed (1qnom) for all speed operation, the
input power reference model can be defined as:
ref -
~~~~P
Lk(N ef
~
Pload
(4)
!!nom
)7 nom
-
Iwhere: P*ef is the input power motor reference and 'nom is the
CONTROL
i ' i X X ; nominal motor efficient.
Input pDwer
ref nrudl
--_
_
__________
}
I'mB.FfStructure ofModel reference Neural Network Controller
z
L !
to design the neural network controller, the
Basically,
l tP T P, ; number of inputs and outputs neuron at each layer are equal to
the number of input and output signals of the system
e
X
respectively. Further the number of hidden layers and the total
neurons is depended on the complexity of the system and the
required training accuracy. Based on the type of the task to be
Pf
performed, the structure of the proposed neural network
controller is shown in Fig.2.
_ _ _ _ _-
Fig. 1.
Diagram block model reference neural network efficiency
optimization of a scalar control induction motor drive system
The controller will receive three input signal i.e. the speed
reference signal (a *), error speed signal () *- (om) and error
input power signal (Pref - Pd). The output of the controller that
consist of stator voltage reference signal or modulation index
(V=m,) and frequency reference signal or modulation
frequency( mf) is fed to the space vector PWM modulator.
In this scheme the input power reference model (Pref)
block is determined as follows. With the load torque
characteristic of the compressor assumed proportional to the
square of the speed as given by:
b2
XI
Fig. 2.
_
Diagram block of neural network efficiency optimization
controller
Tlad =k N2
L
(1)
The structure of the neural network controller consists of
three layers. Based on the neuron number in each layer this
structure is known as 2-3-2 network structure. The first layer is
the input, which consists of two input signals XI and X2. XI
received signal from the speed reference or speed command
e) *, while X2 received signal from the output layer Y, as a feed
back loop or recurrent structure model.
By using in-start model, each of the neuron signals in the
where: Tload is compressor load torque, kL is load torque
coefficient and N is motor speed=compressor speed.
The power of the compressor as a mechanical motor load
with friction and windage are not considered can be defined as:
Pla =load
Tload N
(2)
layer is feedforward to all neurons in the hidden layer via
connections between the input and the hidden
~ layers.
~input
~ ~the weight
The connections weight between neuron i and] in the Ith
neuron at 'tth layer respectively are represented by Wmji.
1717
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on December 29, 2008 at 19:21 from IEEE Xplore. Restrictions apply.
The second layer also known as hidden layer consists of six
all, a12, .. a16 respectively. Besides receiving signal
from input layer, it also receives the bias signal. A transfer
function of the neuron in the hidden layer at the jth neuron is
defined by:
neurons
n =
n
wl,iXi + bl
a
m
=i
a m
J
W m.
J,l
a
i
Wm
J,i
Where the first term on the right hand side is defined as the
Marquardt sensitivity is given by:
(6)
sm =
where: nj> is neuron transfer function in hidden layer, XI is
input value that has been normalized, wl,i is weight connection
parameter value between input layer to hidden layer and bjl is
bias parameter value in hidden layer.
At the hidden layer the tangent hyperbolic activation
function is employed. The neuron output function in this layer
is given by:
-
(12)
i
Derivative of the neuron output function against to the
weight parameter is given by:
man'
=
a
a
(13)
j,'
l
1nl
n
exp
l+ exp
Substitution
result in:
(7)
-n .e
>k
~w
The output layer consist of two neurons, the first neuron is
used as a reference signal frequency (Y1 =f *) and the second
neuron is used as a reference signal voltage (Y2=V,*). The
activation function employed in this layer is known as the
linear activation function. The neuron output function in this
layer is used as an output variable as given by:
2
n
2
1
Equations (12) and (13) into Equation (1) is
n
j,'
sSM a' 1
(14)
The updating neural network parameters can be written by:
Xk+j
= Xk - [JT (xk )JT (xk) + dukI] JT(x, )e(xk) (15)
III.
b2
SIMULATION AND EXPERIMENTAL RESULTS
(8)
r
i
" 'A. Simulation Results
='
Simulation of the efficiency optimization of the proposed
control scheme is carried out using various block developed to
= 2
(9) represent the actual system using the MATLAB/SIMULINK
program. The Simulink block consists of three major blocks,
i.e. the three phase induction motor and compressor load block,
three phase space vector PWM inverter block and the
C. Real-Time Levenberg-Marquardt Learning Algorithm
controller block. These blocks are designed in the S-function
The important step in Levenberg-Marquardt neural network
algorithm is the computation of the Jacobian matrix. For two block by employing Borland C++ program. The induction
motor data are given in appendix.
output neuron, the Jacobian matrix J of the neural network is
given by [1 7]:
To investigate the efficiency improvement of the proposed
controller, two Simulink controller blocks of the proposed
controller and neural network constant Volt per Hertz are
developed in parallel. In order to switch the controller from the
ael
ael
ael
proposed controller to neural network constant Volt per Hertz
awlll aWl
ab22
or vice versa, a switch selector block is added and fed to the
1,1J1,2
(10)
controller. At the start of the plot, the variable speed
_
compressor motor drive system was operated by neural
ae2 ae2
ae2
network
constant Volt per Hertz, after the system is stable at 3
awli,i aWlii,2
ab22
second the controller is switched to the proposed controller.
Back propagation derivation of Jacobian matrix weight
Fig. 3 shows the response of the input power, rotor speed and
parameters is described by the following function.
stator voltage of the motor when the control is switched from
the neural network constant Volt per Hertz to the proposed
controller at speed reference command of 500 rpm.
+
.
.
1718
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on December 29, 2008 at 19:21 from IEEE Xplore. Restrictions apply.
Thespdrncmaitgl2.53S4
Initially, the motor is run at reference speed command by using
neural network constant Volt per Hertz, and maintaining the
same load condition the controller is changed to the proposed
800-
controller after 3 seconds.
Fig. 4 show responses of the rotor speed, electromagnetic
torque, stator voltage and input power motor when the
controller is switched from the neural network constant Volt
per Hertz to the proposed controller at a speed reference
command of 500 rpm.
600400200
2
2.5
3
35
time 0s)
4
45
(
3~~~~~~~~~~~~~~~~~~~~~~~~~000
(a)
8050
(b)
100
022S
500''
20
3
3
60~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
time~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~tm
per Hertz2control
are developed in the DSP controller board.
2.5
02
35
3
(s)
4
4.5
3
S0
Fig. 3. Simulation results when the controller was switched from the~~~~~~~~400
0,2
Fingthi e
2.5
n
3
3.5
time s5
4
(b
4.5
(b) s
B.ExperimentalResults
to
v
t
fg
\~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ism
results when the controller was switched from the i
1407
3. Simulamenta results when the controller was switched from the
Ithseprmn,tveiyteefcecimrvmnofFig.
tepooecotolrthsaepoeueththvbenNNV/fto
the NNEOC
at 1s: (a) the speed response,
(b) the stator voltage
thesatohaebe
NN/othe
propose aotroller,athe speed
N/tteNOals()hsedepne()hsaovoltage
prcdrespneb
doei
h smlto
nscio II.
I. eeoe,response
s
power response.fl
and (c) the input powerresponse.
done In the
isdvlp.rsonad()hnput
spmltonsan ctsthes inseto
1719~~~~~~~4
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on December 29, 2008 at 19:21 from IEEE Xplore. Restrictions apply.
The experimental results show that, by using the proposed
Florida Power and Light Company" IEEE Transaction on Energy
controller, the input power consumption and stator voltage
reduce, and the speed of the motor can be maintained constant
[4] Bose, B. K.; Patel, N. R. and Rajashekara, K.. "A Neuro-Fuzzy-based
per Hertz scheme have been conducted to verify the
efficiency improvement of the proposed controller. It shows
[7]
Conversion, vol. 7, No.3, pp.396-404, September1992.
in accordance to the speed reference command.
On-line Efficiency Optimization Control of a Stator Flux-Oriented
Direct
Vector-Controlled Induction Motor Drive". IEEE Transaction on
Industrial Application. 44(2):270-273, 1997.
IV. CONCLUSION
[5] Zidani, F.; Said, M.S.N; Abdessemed, R.; Diallo, D. and Benbouzid,
The develop t oM.E.H.. "A Fuzzy Technique for Loss Minimization in Scalar
The development of the neural network control to optimize
Controlled Induction Motor." Electric Power Components and Systemthe efficiency of the compressor motor at low speed operation
Taylor & Francis. 30(6): 625-635,.2002.
has been presented. Simulation and experiments on the variable
[6] Abrahamsen, F.; Blaabjerg, F.; Pedersen, J.K.; Pedersen, J.K. and
speed compressor motor drive system with neural network
Thoegersen, P.B. "Efficiency-Optimized Control of Medium-Size
Induction Motor Drives". IEEE Transaction on Industrial Application.
efficiency optimization control and neural network constant
37(6): 1761-1767, 2001.
Volt
Kioskederis I. and Margaris N., "Loss Minimization in Scalar Control
Motor Drives with Search Controllers." IEEE Transaction on
that
forthe
at low
speed operation
operation theInduction
that
for
the proposed
controller at
lOW speedl
the
Power Electronics, vol. 11,,N.p.1
No.2, pp.213-220, Marchl1996.
proposedl controller
iS
increased
and
the
error
efficiency significantly
speed during
[8] Bernal, F. F.; Cerrada, A. G. and '20
Faure, R. "Model-based
Poe
Elcrois
vol
I
ac
96
andMutual
336
H
inductance: 00.336
andl
Mlutual inductance:
HOU
loss
minimization for DC and AC vector-controlled motors including core
saturation". IEEE Transaction on Industrial Application. 36: 755-763,
2000.
[9] Ta, C.M. and Hori, Y. "Convergence Improvement of Efficiency
Optimization Control of Induction Motor Drives". IEEE Transaction on
Industrial Application. 37(6): 1746-17, 2001.
[10] Chakraborty, C.; Uchida, T. and Hori, Y.. "Fast Search Controllers for
Efficiency Maximization of Induction Motor Drives Based on DC Link
Power Measurement". Proceedings of IEEE-Power Conversion
Conference. 2: 402-408, 2002.
[11] Sousa G.C.D., Bose B.K. and Cleland J.G., "Fuzzy Logic Based on-Line
Efficiency Optimization Control Of An Indirect Vector Controlled
Induction Motor Drives" IEEE Transaction on Industrial Electronics,
vol. 42, No.2, pp.192-198, April 1995.
Combined inertia: 0.00153 kg-m2
[12] Choy I.,
the efficiency optimization process can be compensated
immediately.
APPENDIX
Motor Parameters
Power: 0.25hp; Voltage: 120V; Frequency: 50 Hz and Speed:
1460 rpm.
Stator resistance: 5.2 Q. and Rotor resistance: 4.0 Q.
Stator self inductance: 0.347 H; Rotor self inductance: 0.347 H
Kwon S.H., Choi J.Y., Kim J.W., and Kim K.B.," On-Line
Efficiency Optimization Control of a Slip Angular Frequency Controlled
Induction Motor Drive Using Neural Networks." IECON Proceedings 13
annual Conference, pp. 1216-1221, 1996
[13] Yatim A.H.M., and Utomo W.M., " On-Line Optimal Control of
Variable Speed Compressor Motor Drive Using Neural Control Model."
PECon Proceedings Conference, pp. 83-87, 2004
[14] Hasan K.M., Zhang L., and Singh B.," Neural Network Control of
Induction Motor Drives for Energy Efficiency and High Dynamic
Performance." IECON Proceedings 13 annual Conference, pp. 488-492,
ACKNOWLEDGMENT
The authors would like to acknowledge the financial
support for this work by Ministry of Science, Technology and
Innovation (MOSTI) through Intensification of Research in
Priority Areas (IRPA) program.
1997
REFERENCES
Sen, P.C.; Namudiri, C.S. and Nandam, P.K. "Evaluation of Control
Techniques for Industrial Drives". Proceedings of Power Electronics,
Drives and Energy Systems for Industrial Growth Conference. 2: 869875,1996.
[2] Abrahamsen F., Blaabjerg F., Pedersen J.K., and Grabowski P.Z., "On
the Energy Optimized Control of Standard and High Efficiency
Induction Motors in CT and HVAC Application." IEEE Transaction on
Industrial Application. 34(4):822-831, 1998.
[3] Domijan A., Hanhock O., and Maytrott C," A Study and Evaluation of
Power Electronic Based Adjustable Speed Motor Drives for Air
Conditioners and Heat Pumps whit an Example Utility Case Study of the
[1]
[15] Abdin E.S., Ghoneem A.S., Diab H.M.M., and Desaz S.A.," Efficiency
Optimization of a Vector Controlled Induction Motor Drive an Artificial
Neural Network." IECON Proceedings annual Conference, pp. 25432548, 2003.
[16] Pryymak, B.; Moreno-Eguilaz, J.M. and Petracaula, J. "An Efficient
Energy Controller for Induction Motor Drive With Compensation of
Temperature Using Neural Networks". Proceedings of EPE. 1-10, 2005.
[17] Wilamowski, B.M., Iplikci, S., Kaynak, 0. and Efe, M.O.," An
algorithm for fast convergence in training neural networks." Neural
Networks, IJCNN '01 Proceedings International Joint Conference, pp.
1778 - 1782 vol.3, July 2001.
1720
Authorized licensed use limited to: UNIVERSITY TEKNOLOGI MALAYSIA. Downloaded on December 29, 2008 at 19:21 from IEEE Xplore. Restrictions apply.