Está en la página 1de 7

IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 20, NO.

3, SEPTEMBER 2005 497

ANN-Based Soft Starting of Voltage-Controlled-Fed


IM Drive System
Adel Gastli, Senior Member, IEEE, and Mohamed Magdy Ahmed, Member, IEEE

Abstract—Soft starters are used as induction motor controllers


in compressors, blowers, fans, pumps, mixers, crushers and
grinders, and many other applications. Soft starters use ac voltage
controllers to start the induction motor and to adjust its speed.
This paper presents a novel artifical neural network (ANN)-based
ac voltage controller which generates the appropriate thyristors’
firing angle for any given operating torque and speed of the motor
and the load. An ANN model was designed for that purpose.
The results obtained are very satisfactory and promising. The
advantage of such a controller are its simplicity, stability, and high
accuracy compared to conventional mathematical calculation of
the firing angle which is a very complex and time consuming task
especially in online control applications.
Index Terms—AC voltage controller, artificial neural network
(ANN), firing angle, induction motor, soft starter, thyristor.
Fig. 1. Symmetrical ac voltage controller.

I. INTRODUCTION Soft starters allow the machine to start, vary its speed, and
stop with minimum mechanical and electric stresses on the
L IKE induction motor (IM) variable speed drives, soft
starters are also essential components in every modern IM
drives and automation systems. Ac voltage-controller-based
equipment. This can be done by appropriate adjustment of the
IM terminal voltage. However, adjusting the voltage for a given
operating condition of speed and torque is not a very simple
soft starters offer many advantages over conventional starters task. To adjust the voltage, the firing angle of the thyristors
such as the following. shall be calculated for each operating condition. This firing
• Smooth acceleration, which reduces stress on the me- angle is a nonlinear function of the motor speed and torque and
chanical drive system due to high starting torque hence it is quite difficult to find the exact value of for any motor
increases the life and reliability of belts, gear boxes, chain speed and torque. Some methods of closed loop control of ac
drives, motor bearings, and shafts [1]. Smooth accelera- voltage regulators have been developed and applied, which
tion reduces also stress on the electrical supply due to high requires a speed sensor [9]. In [6] and [10], the authors have
starting currents meeting utility requirements for reduced proposed a method of optimal soft starting without a speed
voltage starting and eliminating voltage dip and brown out sensor but it requires sensing of the thyristors voltages.
conditions[2]–[4]. It reduces also the shock on the driven This paper proposes an artifical neural network (ANN)-based
load due to high starting torque [4], [6] that can cause selection of the thyristors firing angles of a voltage-con-
a jolt on the conveyor that damages products, or pump trolled-fed IM drive system. The controller operates in open
cavitations and water hammer in pipes. Thus, a fully ad- loop and does not require any speed or voltage sensing. The
justable acceleration (ramp time) and starting torque for only sensor that is needed is a current sensor, which in most of
optimal starting performance, provides enough torque to applications is used to protect the converter and the motor from
accelerate the load while minimizing both mechanical and over currents. The soft starter is designed to meet the industrial
electrical shock to the system [6]. requirements of compressors, blowers, fans, pumps, mixers,
• Energy savings at lightly loaded conditions. Energy sav- crushers and grinders, etc.
ings by voltage control is achieved by reducing the applied
voltage if the load torque requirement can be met with less II. SOFT STARTER
than rated flux. This way, core loss and stator copper losses A soft starter is an ac voltage controller in which the voltage
can be reduced [7], [8]. is adjusted through the setting of the thyristors firing angle .
Fig. 1 shows a typical configuration of a symmetrical voltage
Manuscript received June 15, 2001; revised April 7, 2004. Paper no. controller. The six thyristors in Fig. 1 are fired according to the
2001TR328. sequence shown in Fig. 2. Note that at least two thyristors must
The authors are with the Department of Electrical and Computer Engineering, conduct simultaneously to allow current to flow through the load
Sultan Qaboos University, Al-Khodh, Muscat 123, Sultanate of Oman (e-mail:
agastli@ieee.org; magdy@squ.edu.om). and that the firing angle is measured from the zero crossing
Digital Object Identifier 10.1109/TEC.2004.841522 of phase A voltage.
0885-8969/$20.00 © 2005 IEEE
498 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 20, NO. 3, SEPTEMBER 2005

Fig. 4. Torque versus speed characteristics for different values of the thyristors
firing angle.
Fig. 2. Thyristors firing logic.

Fig. 3. Matlab/Simulink simulation model used to generate data for ANN


training. Fig. 5. Block diagram of ANN model.

A soft-starter-fed three-phase induction motor was modeled of the firing angle smaller or equal to the load impedance angle
and simulated using Matlab/Simulink Power system blocksets , there is a continuous conduction.
as shown in Fig. 3. The asynchronous motor and all power Thus, changing the firing angle within that limit will have no
electronics switches were modeled according to their operating effect on the voltage applied to the motor; hence, the torque and
characteristics. The Simulink model was used for predicting speed values are kept unchanged.
the performance of the soft starter and has given satisfactory re-
sults. The program is run several times with a fixed firing angle III. ANN
and varying load torque. The steady state value of the speed
The ANN model, used for the calculation of the appropriate
is calculated after each run. The same procedure is repeated
thyristors’ firing angle as a function of the motor speed
for different values of the firing angle. The system parameters
and torque , has two input variables ( and ) and one
used in the simulation are very similar to those of actual lab
output variable . Since the angle is a nonlinear function of
equipment: 1/3 hp, 4 poles, 50-Hz, 220-V, star-connected,
the and , then the tansigmoidal function given by (1) is
squirrel-cage induction motor. The per-unit motor parameters
the most appropriate to model it [11]
obtained using the rated voltage as a base voltage and 375 W
as a base power are
(1)

The typical two-layer architecture used for the firing angle


calculation is shown in Fig. 5. It has a hidden layer of tan-
Fig. 4 shows the torque-versus-speed characteristics obtained sigmoidal neurons, which receive inputs (in this case the speed
for different values of thyristors’ firing angle. Note that for small and the torque) directly, then broadcast their outputs to a layer of
values of , from 0 up to 30 the speed torque characteristics linear neurons, which compute the network output (in this case
are very similar. This can be explained by the fact that for values the firing angle).
GASTLI AND AHMED: ANN-BASED SOFT STARTING OF VOLTAGE-CONTROLLED-FED IM DRIVE SYSTEM 499

TABLE I newly generated sets of speed and torque patterns are input to
PARAMETERS OF THE NEURAL NETWORK MODEL the Matlab neural net model and the corresponding pattern of
firing angle is calculated systematically. According to Fig. 6, it
is noticed that there is a very good fitting between both ANN
and actual results, which proves that the designed ANN model
is very precise.
It is important to note that this accuracy is valid only for
the input data that is within the boundaries specified during the
training of the ANN model (see Table I for maximum and min-
imum input values). Beyond those boundaries, the accuracy may
not be as good as expected.

IV. SIMULATION
A. Procedure
The ANN model and the drive system were implemented with
Simulink Power System Blockset and Simulink Neural Network
Toolbox as shown in Fig. 7.
The weight and biases used are those obtained from the pre-
viously trained network in Table I. The load was considered as
a pump or a fan with the following torque-speed characteristic:

(2)

where is a constant.
The input to the program is the reference speed . Based
on the setting of the initial conditions, the program can simulate
any operating condition of the system such as starting, breaking,
or speed control. The reference torque is calculated based on the
reference speed and load characteristic (2). Both the reference
speed and torque are input to the ANN model, which produces
Fig. 6. Comparison between actual and ANN results.
the corresponding firing angle. A triggering logic module gen-
erates pulses to the six thyristors according to the firing pattern
This architecture has been proven capable of approximating
in Fig. 2.
any function with finite number of discontinuity with arbitrary
The outputs of the program are:
accuracy [11]. If an input set of data corresponds to a definite
speed and torque pattern, the network can be trained to give a • instantaneous values of the line currents and voltages;
correspondingly desired firing angle pattern at the output. The • instantaneous values of the motor torque and speed;
network has the capability to learn from a sample set of input • rms value of the current and average value of the torque.
output data within a certain boundaries.
The back-propagation-training algorithm is most commonly B. Results
used for this type of feed-forward neural networks. The training The system was first tested during soft starting operation fol-
is automated with Matlab simulation program that uses a cer- lowing a linear speed ramp increase of 0.08 p.u./s. The reason
tain number of input–output sample patterns. The sample pat- behind using ramp increase and not step increase of the speed
terns can be derived by experiments or by simulations. In our is because, as it was mentioned before, one of the advantages
case, the patterns were derived by simulation using the com- of using soft starters is to produce less electrical and mechan-
puter program mentioned in Section II (Fig. 3). Part of the gen- ical stresses on the drive system during motor starting and speed
erated data is used to train the neural network. At the end of the variations. The simulation results are presented in Fig. 8. It is
training process (when the target mean squared error is reached), clear that, at steady state, the speed reaches the same value as
the model obtained consists of the weight and the bias vectors. the reference speed.
Table I summarizes the parameters of the ANN model and the To illustrate the benefits of using the ANN-based soft starter,
results of the training. The obtained weight and bias vectors are direct-on-line motor starting simulations were also carried out.
saved in a file that will be used during the simulation of the IM The results are shown in Fig. 9. Notice that, in the case of di-
drive system. rect-on-line starting (Fig. 9), the rms current and av-
To check the accuracy of the designed ANN model, the actual erage torque oscillate and reach values higher than three times
values of the firing angle obtained by simulation were compared their rated values before reaching their steady-state values. This
to those obtained by the ANN (see Fig. 6). sudden increase and oscillations of the current put stresses on
Note that the samples used for this comparison are different the electrical supply and create voltage dips. The oscillations of
from the samples used for the training of the ANN model. The the torque decrease the life and reliability of belts, gearboxes,
500 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 20, NO. 3, SEPTEMBER 2005

Fig. 7. Matlab/Simulink simulation model with ANN.

Fig. 9. Simulation results of direct-on-line motor starting.

Thus, it provides enough torque to accelerate the load while


minimizing both mechanical and electrical shock to the system.
The designed ANN-based control of the soft starter was also
tested for speed control. Fig. 10 shows the simulation results
when the speed varied linearly from 0.5 to 0.8 p.u., then back
to 0.5 p.u. The reference speed was varied following a ramp
of 0.1 p.u./s. Notice that the actual values of speed and torque
follow very closely the reference values. Moreover, the cur-
rent and torque variations are taking place very smoothly, thus,
avoiding high electrical and mechanical stresses on the system.

V. EXPERIMENTS
The proposed control algorithm was tested by experiments
using the hardware configuration shown in Figs. 11 and 12. The
experimental setup uses the LabVolt 0.2-kW electromechanical
Fig. 8. Simulation results of motor soft starting characteristics: (a) torque system (model 8006) that is controlled with a PC through a
[p.u.], (b) speed [p.u.], (c) firing angle, and (d) current rms [p.u.]. digital-to-analog converter. The system consists of the following
hardware equipment.
chain drives, motor bearings and motor shafts. However, in the • 1/3 hp, 4 pole, 50-Hz, 220-V, star-connected, squirrel-cage
case of soft starting (Fig. 8), the oscillations vanish and the cur- induction motor.
rent and torque vary smoothly toward their steady state values. • DC dynamometer as a load.
GASTLI AND AHMED: ANN-BASED SOFT STARTING OF VOLTAGE-CONTROLLED-FED IM DRIVE SYSTEM 501

Fig. 11. Photograph of the experimental system using the LabVolt 0.2-kW
electromechanical system, model 8006.

Fig. 10. Simulation results for speed control: (a) torque [p.u.], (b) speed [p.u.],
(c) firing angle, and (d) current rms [p.u.].

• Three-phase 220-V, 50-Hz balanced power supply.


• AC regulator unit using six power thyristors.
• Thyristor firing unit controlled by a dc voltage (Volts
Firing angle).
• Digital-to-analog converter which converts a digital
voltage value output from the PC parallel port into an
analog voltage that is fed to the thyristor firing unit.
• PC unit which calculates the appropriate firing angle for
each given speed and torque using the proposed control Fig. 12. Schematics of the experimental system configuration.
algorithm. The output data through the parallel port is the
equivalent dc voltage that is required by the thyristor firing • The designed PC program is run which asks to enter from
unit to generate the appropriate pulses at the required firing the keyboard the values of torque and speed.
angle. • The entered torque value is 0.3 and the speed value is en-
The experiments were conducted in order to check if the refer- tered according to the values used to train the ANN model.
ence speed of the motor set by the PC program is the same as the • The program calculates the corresponding firing angle
one measured at the motor shaft for different load torque values. based on the ANN model and converts it to an equivalent
These load torques were 0.3, 0.45, and 0.6 Nm. The experiments binary dc voltage value.
were conducted following the procedure described hereafter: • This binary dc voltage value is output to the parallel port
• The load torque of the motor is set to 0.3 Nm on the and converted to an analog value by the digital-to-analog
dynamometer side. converter.
502 IEEE TRANSACTIONS ON ENERGY CONVERSION, VOL. 20, NO. 3, SEPTEMBER 2005

Note also that these results validate the good accuracy of the
Matlab/Simulink simulation model since the ANN model was
built based on training data obtained from the simulation model
and not from the experiments.

VI. CONCLUSIONS
In this paper, a novel method for controlling soft-starter-fed
induction-motor-drive systems using ANN is introduced. The
method consists of training a two-layer ANN model on a set
of data generated by simulation or experiments. The generated
data are the speed and torque patterns as inputs and their corre-
sponding firing angle patterns as output of the ANN model.
The ANN model was trained successfully and the results of
comparison between the actual data and the ANN calculated
data were very satisfactory. To validate the effectiveness of the
proposed soft starter control scheme, an induction motor fan
Fig. 13. Reference and actual speed at T = 0:3 Nm. drive system, fed by the proposed soft starter, was implemented
by software program and hardware experimental setup. Several
simulations and experiments were carried out for different op-
erating conditions and the results were very satisfactory.
Thus, the ANN approach has resolved the problem of the
complexity of the online determination of the appropriate
thyristor firing angle for any operating condition. It is also
important to note that the controller operates in open loop
which has the advantage of being stable and does not require
any speed, or voltage sensing.
The proposed soft starter is designed to meet the industrial
requirements of compressors, blowers, fans, pumps, mixers,
crushers and grinders, etc.

REFERENCES
[1] R. F. McElveen and M. K. Toney, “Starting high-inertia loads,” IEEE
Trans. Ind. Appl., vol. 37, no. 1, pp. 137–144, Jan./Feb. 2001.
[2] A. J. William and M. S. Griffith, “Evaluating the effects of motor starting
on industrial and commercial power systems,” IEEE Trans. Ind. Appl.,
vol. IA-14, no. 4, pp. 292–299, Jul./Aug. 1978.
Fig. 14. Speed error versus actual speed for different load torque values. [3] F. M. Bruce, R. J. Craefe, A. Lutz, and M. D. Panlener, “Reduced-voltage
starting of squirrel-cage induction motors,” IEEE Trans. Ind. Appl., vol.
IA-20, no. 1, pp. 46–55, Jan./Feb. 1984.
• Finally, this dc voltage is converted to an equivalent [4] J. Nevelsteen and H. Aragon, “Starting of large motors—Methods and
thyristor firing signal that controls the thyristors with the economics,” IEEE Trans. Ind. Appl., vol. 25, no. 6, pp. 1012–1018,
Nov./Dec. 1989.
selected angle which in turns set the motor voltage to [5] A. A. Shaltoutn, “Analysis of tortional torques in starting of large
the desired value. squirrel cage induction motors,” IEEE Trans. Energy Convers., vol. 9,
• The same procedure is repeated for different speed and no. 1, pp. 135–141, Mar. 1994.
torque values. [6] G. Zenginobuz, I. Çadirci, M. Ermis, and C. Barlak, “Soft starting
of large induction motors at constant current with minimized starting
Fig. 13 shows the experimental results obtained for the three torque pulsations,” IEEE Trans. Ind. Appl., vol. 37, no. 5, pp. 137–144,
sets of load torque values: 0.3, 0.45, and 0.6 Nm (high, medium, Sept./Oct. 2001.
and light load). Comparing the actual speed (diamond points) to [7] F. Blaabjerg, J. K. Pedersen, S. Rise, H. H. Hassen, and A. M. Trzynad-
lowski, “Can soft-starters help save energy?,” IEEE Ind. Appl. Mag., vol.
the reference speed (solid line), one can clearly notice the good 3, no. 5, pp. 56–66, Sep./Oct. 1997.
fitting between these values. [8] N. Mohan, “Improvement in energy efficiency of induction motors by
Fig. 14 shows the absolute error between the reference speed means of voltage control,” IEEE Trans. Power App. Syst., vol. PAS-99,
(the value entered to the program) and the actual speed (the pp. 1466–1471, Jul./Aug. 1980.
[9] S. B. Dewan, G. R. Slemon, and A. Straughen, Power Semiconductor
reading in the dynamometer screen) as a function of the refer- Drives. New York: Wiley Interscience, 1984.
ence speed at different load torques. [10] V. V. Sastry, M. R. Prasad, and T. V. Sivakumar, “Optimal soft strating
It is clear from the results shown in Fig. 14 that the maximum of voltage-controller-fed IM drive based on voltage across thyristors,”
IEEE Trans. Power Electron., vol. 12, no. 6, pp. 1041–1059, Nov.
speed error is below 0.4%, which is relatively very small. This
1997.
result validates the previous statement that the proposed control [11] D. Howard and B. Mark, “Neural network toolbox for use with matlab,”
algorithm is very accurate. in User Guide: The Math Works Inc., 1992.
GASTLI AND AHMED: ANN-BASED SOFT STARTING OF VOLTAGE-CONTROLLED-FED IM DRIVE SYSTEM 503

[12] T. A. Lipo, “The analysis of induction motor with voltage controlled by Mohamed Magdy Ahmed (M’85) received the
symmetrically triggered thyristors,” IEEE Trans. Power App. Syst., vol. B.Sc. and M.Sc. degrees in electrical engineering
PAS-90, no. , pp. 515–525, Mar. 1971. from Alexandria University, Alexandria, Egypt, in
1974 and 1980, respectively, and the Ph.D. degree
from UMIST, Manchester, U.K., in 1985.
He is currently an Assistant Professor of Electrical
Adel Gastli (S’89–M’93–SM’00) received the B.Sc. Engineering at Sultan Qaboos University, Muscat,
degree in electrical engineering from Ecole Nationale Oman. His research interests are in the area of
des Ingénieurs de Tunis, Tunisia, in 1985, and the simulation and electric drives.
M.Sc. and Ph.D. degrees from Nagoya Institute of Dr. Ahmed is a member of the IEEE Industry
Technology, Nagoya, Japan, in 1990 and 1993, re- Applications, Power Electronics, and Industrial
spectively. Electronics societies.
He was with the R&D Department at Inazawa
Works (elevators and escalators) of Mitsubishi Elec-
tric Corporation, Japan, from April 1993 to August
1995. He is currently an Associate Professor of
Electrical Engineering at Sultan Qaboos University,
Muscat, Oman. His current research interests include electrical machines,
power electronics, drives, and control.