Monitoring the True Harmonic Current of a Variable Speed Drive Under Nonsinusoidal

Supply Conditions
J. Mazumdar, R. Harley, F. Lambert and T. Habetler
Georgia Institute of Technology
School of Electrical and Computer Engineering
Atlanta, GA 30332, USA
Email: joymazum@ece.gatech.edu

AbstractThe increased use of nonlinear devices in industry

has resulted in direct increase of harmonic distortion in the
industrial power system. Variable speed drives are an example.
With the widespread proliferation of nonlinear loads in a power
distribution network, the voltage at the point of common
coupling is rarely a pure sinusoid. It has become necessary to
identify accurately which load(s) is injecting the excessively
high harmonic currents. Simply measuring the harmonic
currents at each individual load is not sufficiently accurate since
these harmonic currents may be caused by not only the
nonlinear load, but also by a non-sinusoidal PCC voltage. This
paper proposes a neural network solution methodology for the
problem of measuring the actual amount of harmonic current
injected into a power network by a three phase variable speed
drive, and this technique can be extended to any nonlinear load
in general. The proposed method has been experimentally
verified by applying the scheme to a commercially available
variable speed drive. The scheme has been applied to each
phase individually as well as to all three phases together. The
goal of this paper is to quantify the difference in current
distortion of a load when supplied from a distorted source as
compared to a clean sine wave. A Multilayer Perceptron Neural
Network is used to estimate the true harmonic current
distortion of a load. Theory and practical results are presented.
This technology could be integrated into any commercially
available power quality instrument or be fabricated as a
standalone instrument.

I.

INTRODUCTION

The increased use of nonlinear devices in industry has

resulted in the increase of harmonic distortion in the
industrial power systems. Speed control for industrial ac
motors is one of the most important applications in today's
process/industrial control environment. This is achieved with
the help of a variable speed drive (VSD). VSDs present a
nonlinear admittance to the power grid feeding them. Most
VSDs have a diode bridge rectifier as a front end or
sometimes even SCRs are used. The switchings of these
nonlinear devices result in a pulsed input current with high
harmonic content.
The harmonic currents propagate within the power
distribution network and can cause extensive damage to
sensitive equipments connected to the same source. The input
current can certainly be made cleaner by using active filtering
techniques [1]. However that is not a standard option and is
expensive. When a VSD is connected to the distribution
network, as shown in Fig. 1, it gives rise to concerns.

G. Venayagamoorthy
University of Missouri - Rolla
Real-Time Power and Intelligent Systems Laboratory
Rolla, MO 65409, USA
Email: gkumar@ieee.org

Rs

Ls

v pcc
n

vs

Other
Loads

is = i1 + ih

VSD

h=2

Fig. 1: Variable speed drive connected to the distribution network

When the VSD is supplied is supplied from a sinusoidal

voltage source, its injected harmonic current ih is referred to
as contributions from the load, or load harmonics. These
harmonic currents interact with the line parameters
RS and LS and cause harmonic volt drops in the supply
network. Hence the voltage at the point of common coupling
(PCC) is no longer sinusoidal. Any other loads, even linear
loads, connected to the PCC, will have harmonic currents
injected into them by the distorted PCC voltage. Such
currents are referred to as contributions from the power
system, or supply harmonics.
Given the condition that the voltage at the PCC is distorted
and numerous loads are connected to the PCC, it is difficult to
predict the interactions between loads and the load and the
source. Under these circumstances, the determination of
harmonic current distortion of a load by simply measuring its
input current may not be an accurate representation of its
harmonics. IEEE 519 provides guidelines for utilities and
customers for controlling harmonics [2], [3].
It has been observed through actual measurements that the
distortion in the measured value of the input current is
different when it is supplied from a distorted source and a non
distorted source. The goal of this paper is to present a method
to predict the true current distortion of a variable speed drive
when it is supplied from a distorted source. This technique
can also be applied to any other nonlinear load in general to
determine its true current distortion.
II.

HARMONIC CURRENT DETERMINATION

The issue presented above is certainly not a new area of

research or concern. However, with the widespread
proliferation of nonlinear loads, it has become critical to
identify accurately which load(s) is injecting any excessively
high harmonic currents. Simply measuring the harmonic
currents at each individual load is not sufficiently accurate
since these harmonic currents may be caused by not only the
nonlinear load, but also by a non-sinusoidal PCC voltage.

Methods like active harmonic power flow direction [4], [5]

and online impedance measurements [6], [7] have been
proposed to offer a solution to the problem. However most
methods assume a radial feeder supplying a single load
through a known feeder impedance, or multiple loads
connected to a PCC which has a sinusoidal voltage and with
zero impedance in the supply feeder.
This paper proposes a method based on Artificial Neural
Networks (ANN) to determine the true harmonic current of a
nonlinear load in a three phase power system. An ANN is
characterized by the ability to learn or modify its behavior in
response to the environment. The greatest advantage lies in
the fact that a trained network can extract essential features
from unfamiliar inputs through generalization and
recognition. ANN based load identification techniques are
increasingly being used in power system applications.
Figure 2 shows the configuration of the proposed scheme
for determination of true harmonic current. A Multilayer
Perceptron Neural Network (MLPN) architecture is used for
the implementation of the scheme. Details about the neural
network architecture, governing equations as well as
implementation of this scheme with single phase loads can be
found in [8], [9].
THD Instrument

vsin e

Estimation
Neural
Network
ANN2

iabc distorted

Weights

Delay

Identification
Neural
Network
(ANN1)

iabc

iabc

vs

RS

LS

vabc

the three line currents iabc at the k th moment in time. The

neural network is designed to predict one step ahead line
current iabc as a function of the present and delayed voltage
vector values vabc ( k ) , vabc (k 1) and vabc ( k 2) . When the
k + 1th moment arrives (at the next sampling instant), the
actual instantaneous values of iabc are compared with the
previously predicted values of iabc , and the difference (or
error e ) is used to train the ANN1 weights. Initially the
weights have random values, but after several epochs, the
training soon converges and the value of the error e
diminishes to an acceptably small value. This process is
called identifying the load admittance.

B. Estimation Neural Network- ANN2

ANN2 is supplied with a mathematically generated sine
wave to estimate its output. The output of ANN2, called
iabc disto rted , therefore represents the current that the
nonlinear load would have drawn had it been supplied by a
sinusoidal voltage source. Any distortion present in the
iabc disto rted waveform can now truly be attributed to the
nonlinearity of the load admittance. Due to the nature of the
sigmoidal transfer function, the outputs of the neurons in the
hidden layer are limited to values between 0 and 1 . The
inputs to the neural networks are therefore scaled and limited
to values between 1 .

e
Supply
Network

A. Identification Neural Network- ANN1

The proposed method measures the instantaneous values
of the three voltages vabc (line or phase) at the PCC, as well as

EXPERIMENTAL RESULTS

III.

For illustrative purposes, the scheme has been applied to a

commercially available VSD, as shown in Fig. 3.

Nonlinear Load

vsin e

Other Loads

( ANN 2 )

Fig. 2: Schematic diagram of the proposed scheme

The nonlinear load injects distorted line current iabc into

the network. The Identification Neural Network (ANN1) is
trained to identify the nonlinear characteristics of the load.
The Estimation Neural Network (ANN2) predicts the true
harmonic current that would be injected by the load into the
network, if it were possible to isolate the load and supply it
from a pure sinusoidal source. ANN2 is structurally an exact
replica of the trained ANN1. The function of ANN2 can very
well be carried out by ANN1; however that would disrupt the
continual online training of ANN1 during the brief moments
of estimating.
The supply configuration at the PCC for a load could be
wye connected neural grounded, wye connected neural
floating or delta connected. Furthermore, the system could be
balanced or unbalanced. The determination of harmonics in a
three phase is done on a per phase basis. Hence the network
inputs are chosen, depending on the supply configuration.

Estimation
1dist
Neural Network

Weights

v pcc

z 1

Identification
Neural Network

error

Acquisition

vs

i1

( LABVIEW )
1 S

v pcc

( ANN1 )

Data

Utility Supply
Network

i1

i1

ACS500
(VSD )

Induction
Motor

Clean Power
Supply

Fig. 3: Experimental setup with VSD. The z-1 represents two time delays.

The clean power source used is a California Instruments

5001 iX harmonic generator which is capable of outputting
voltages with programmable distortion levels and zero
internal impedance. The scheme has to be applied to each

phase individually. The method of using online trained

ANNs to identify the load admittance and utilizing the
trained neural network to estimate the harmonic current of
the VSD, is now demonstrated for phase A.
With switch S in position 1, the VSD is supplied from the
utility source which has a back ground THD of 4.5%. The
measured three phase line to neural voltages and the phase
currents are recorded, and Fig. 4 shows the phase A voltage
and current waveforms.
Acquired Phase A PCC voltage
4
2
0
-2
-4

Fig. 6: FFT spectrum of current with S in position 1

1

1.01

1.02

1.03

1.04

1.05

1.06

1.07

1.08

1.09

1.1

1.07

1.08

1.09

1.1

Acquired Phase A current

5

-5

1.01

1.02

1.03

1.04

1.05 1.06
Time (s)

Fig. 4: Measured voltage and current with S in position 1

Now with switch S in position 2, the drive is supplied

from the clean power source and the measured phase A
voltage and current waveforms are shown in Fig. 5. The total
harmonic distortion (THD) of the utility voltage is 4.5% and
the THD of the CI 5001 iX voltage is 0.2%.

The data obtained with switch S in position 1 is used to

train the neural network ANN1 until the training error
converges to near zero, and the output of ANN1 correctly
tracks the actual current i1 as shown in Fig. 8.

Acquired Phase A PCC voltage

4
2
0
-2
-4

1.01

1.02

1.03

1.04

1.05

1.06

1.07

Fig. 7: FFT spectrum of current with S in position 2

1.08

1.09

Neural network predicted current superimposed on actual current

1.1

Actual current
ANN1 output

Acquired Phase A current

0.8

0.6

0.4

-2
-4

1.01

1.02

1.03

1.04

1.05 1.06
Time (s)

1.07

1.08

1.09

1.1

Currents

0.2
0
-0.2

Fig. 5: Measured voltage and current with S in position 2

The frequency spectrum of the phase A current with S in

position 1 is plotted in Fig. 6 with a current distortion
measure of 74.27%. Figure 7 shows the frequency spectrum
of the phase A current with S in position 2 with a current
distortion measure of 68.5%. Clearly, there is a difference in
the current THDs of the phase A current depending on
whether S is in position 1 or 2. The goal of this paper is to
identify and quantify this difference with the help of neural
network techniques.

-0.4
-0.6
-0.8

1.01

1.02

1.03
Time (s)

1.04

1.05

1.06

Fig. 8: ANN1 current tracking result for phase A

Figure 8 indicates how well the training of ANN1 has

converged since its output coincides with the actual
i1 waveform. The convergence of the training can also be
verified by looking at the absolute value of the tracking error

Te defined as

Te = (i1 i1 )

(1)

and the Mean Squared Error MSE is defined as

2
1 r
(2)
MSE = (i1 i1 )
r 1
where r is the number of epochs. Figure 9 shows the MSE
of the ANN1 current tracking for phase A.
MSE vs Epochs

-1

Mean Squared Error

10

Fig. 11: FFT spectrum of output of ANN2

-2

Similar to phase A, the scheme has also been applied to

phases B and C of the VSD. The training starts with a new set
of random weights. Figure 12 shows the training result for
Phase B. Figure 13 shows the MSE of the ANN1 current
tracking for phase B.

10

-3

10

10

10

10
Epochs

10

Neural network predicted current superimposed on actual current

10

Fig. 9: MSE in phase A current tracking

0.6
0.4
0.2
Currents

Once ANN1 has learned the phase A-to-neutral

admittance of the VSD, the weights of ANN1 are transferred
to ANN2. The output of ANN2 is i1 dist and is obtained by
using a mathematically generated sine wave voltage with
zero distortion as its input.

Actual current
ANN1 output

0.8

0
-0.2
-0.4
-0.6
-0.8
-1

1.01

1.02

1.03
Time (s)

1.04

1.05

1.06

Fig. 12: ANN1 current tracking result for phase B

MSE vs Epochs

10

Figure 10 shows what Fig. 8 would have looked like if it

were possible to isolate the VSD and supply it from a pure
sine wave. Figure 11 shows the FFT spectrum of the
waveform shown in Fig. 10. The true current distortion of
i1 dist in Fig. 11 turns out to be 67.49% (instead of the
74.27% of Fig. 6).
This result agrees well with the measured value of 68.5% of
Fig. 7 where the VSD was supplied by a 0.2% distorted
voltage. In other words, the current waveform shown in Fig.
10 is the true harmonic current that would be injected by the
VSD into the network if the PCC voltage had been
sinusoidal.

Mean Squared Error

Fig. 10: Estimated current waveform obtained from the output of ANN2

10

-1

10

-2

10

10

10

10
Epochs

10

10

Fig. 13: MSE in phase B current tracking

The convergence in ANN1 training for the phase C current

is shown in Fig. 14. Figure 15 shows the MSE of the ANN1
current tracking for phase C.

where THDd is iTHD from a distorted v pcc ,and THDs is

Neural network predicted current superimposed on actual current

1

Actual current
ANN1 output

0.8

iTHD from a mathematical sine wave, i.e. the output of ANN2.

em can be depending on the type of load and the condition
of the network. A positive sign indicates that the current THD
of the load is higher when supplied from a clean source. A
negative sign indicates that the current THD of the load is
lower when supplied from a clean source. The results for the
three phases are summarized in Table I.

0.6
0.4

Currents

0.2
0
-0.2
-0.4

TABLE I
SUMMARY OF RESULTS

-0.6
-0.8
-1

1.01

1.02

1.03
Time (s)

1.04

1.05

1.06

Phase

THDd

THDs

THDCI

em

74.27%

67.49%

68.5%

-10.05%

59.67%

49.02%

47.72%

-21.73%

152%

132.27%

132.47%

-14.92%

Fig. 14: ANN1 current tracking result for phase C

MSE vs Epochs

-1

10

THDCI is the distortion in current with the switch S in

position 2. This value is used for validation of the results
obtained using the proposed scheme. In an actual
implementation of the proposed scheme, the value of
THDCI will not be required since it is not used in the training
algorithm, nor will such a value be available in any real
power system application.
Some of the other experimental details of the neural
network implementation are given below:

Mean Squared Error

-2

10

-3

10

-4

10

10

10

10
Epochs

10

10

Fig. 15: MSE in phase C current tracking

The same procedure is followed to obtain the true current

distortion for phases B and C. The results are summarized in
the next section.
The sampling rate for data acquisition is set at 128 samples
per cycle. The number of neurons used in the hidden layer of
ANN1 is 20. Data acquisition is carried out by a National
Instruments data acquisition system. The voltage transducers
used are LEM LV 25-P and the current transducers used are
LEM LAH 25-NP. FFT of the acquired waveforms are
computed using the powergui block of SIMULINK.
IV.

QUANTIFICATION OF RESULTS

The above experiment has shown that there is a difference

in the current distortion of a load depending on whether the
loads are served by a clean supply or a distorted supply. Any
load serviced by a utility is designed and optimized to operate
at 60 Hz. For the purpose of quantification of this difference,
a new parameter em , known as the resultant error in
measurement, is introduced and is defined as:
em = (

THDs THDd
)%
THDs

(3)

MLP network implemented in MATLAB.

FFT computation : powergui block of SIMULINK
Number of Neurons in the hidden layer: 25
Time delayed inputs : 2
Learning gain: 0.05. Momentum gain not used.
Sampling frequency for data acquisition: 8 kHz.
Power quality instrumentations require ~ 28
samples/cycle.
Computation time for the MATLAB code to compute
the output weights (with 2 sec of acquired data) run on
a 1.8 GHz PC: 5 sec

The accuracy of neural network computations can be

further increased by increasing the sampling rate and number
of neurons. However that puts additional computational
demands on the processor and might make the actual
hardware implementation more difficult.
V.

MODIFIED SCHEME

Up to now, the implementation of the proposed scheme

required one identification network and one estimation
network for each phase. A modified implementation of the
scheme requires only one identification network and one
estimation network for all the three phases. Essentially, the
neural network structure now becomes three inputs and three
outputs. The present and two time delayed values of the
voltage vector, as well as a bias, are used for the training,

The frequency spectrum of the three line currents of Fig.

18 is shown in Fig. 19a, 19b and 19c.
Neural network (ANN2) estimated current
Phase A

1
0
-1

Phase C

Phase A

1.02

1.03

1.04

1.05

1.06

1.07

1.08

1.09

1.03

1.04

1.05

1.06

1.07

1.08

1.09

1.1

1.01

1.02

1.03

1.04

1.05

1.06

1.07

1.08

1.09

1.1

1.01

1.02

1.03

1.04

1.05 1.06
Time (s)

1.07

1.08

1.09

1.1

0
-1

1.01

1.02

1.01

0
-1

Neural network predicted current superimposed on actual current

-1

1
Phase B

making the actual number of inputs to the neural network

equal to ten. The training process begins with the
identification neural network predicting the line current as a
function of present and delayed voltage vector values.
Initially the weights have random values. Identification
network training results for the line currents are shown in Fig.
16. Convergence in training is demonstrated by the fact that
the neural network predicted current waveforms coincide
with the actual current; they practically lie on top of one
another.

1.1

Phase B

Fig. 18: Estimated line currents from the estimation network

0
-1

1.01

1.02

1.03

1.04

1.01

1.02

1.03

1.04

1.05

1.06

1.07

1.08

1.09

1.1

1.05 1.06
Time (s)

1.07

1.08

1.09

1.1

Phase C

1
0
-1

Fig. 16: Training result for the three line currents

The value of the MSE in Fig. 17 shows how the training

convergences for all the line currents. For each phase the
value is lower than 10 3 which is a sufficiently low value to
indicate that the identification neural network has correctly
learned the load characteristics. At this point, the
identification neural network therefore represents the
admittance of the VSD.

(a)

MSE vs Epochs

Phase A

10

-5

10

2000

4000

6000

8000

10000

12000

14000

16000

2000

4000

6000

8000

10000

12000

14000

16000

2000

4000

6000

8000
Epochs

10000

12000

14000

16000

Phase B

10

-2

10

(b)

-4

10

Phase C

10

-2

10

-4

10

Fig. 17: MSE in ANN1 tracking for all three line current

The estimation network is now supplied with a balanced

and symmetrical three phase mathematically generated sine
wave voltage. The output of the estimation network, shown
in Fig. 18, therefore represents the true harmonic current that
would be injected by the VSD into the power distribution
network, if the voltage at the PCC had no distortion.

(c)
Fig. 19: FFT of line currents, (a) Phase a, (b) Phase b, (c) Phase c

Table II provides the comparison of results obtained

using the modified scheme of only two neural networks for
all three phases, and the original scheme of six neural
networks, two per phase. The original scheme refers to the
results obtained from section III and the modified scheme
refers to the results obtained from section V.
TABLE II
COMPARISON OF RESULTS

THDs

THDs

(original)

(modified)

67.49%

66.42%

49.02%

48.63%

132.27%

134.07%

Phase

The results from Table II show that the performance of the

modified scheme is comparable with that of the original
scheme. In any real life situations, the three phases of a load
are rarely balanced and symmetrical. The feasibility of using
a single neural network to learn the nonlinear behavior of an
unbalanced load is therefore established.
VI.

CONCLUSION

This paper has demonstrated the ability of MLP neural

networks to learn the admittance of a VSD and utilize the
trained neural network for estimating the true harmonic
distortion caused by that load. The proposed method has been
successfully applied to a specific three phase load. The
advantages of the proposed method are that it can be
implemented online without disrupting the operation of any
load, only voltages and currents need to be measured. It does
not require any special instruments and it does not need to
make any assumptions about any quantities, e.g. the
impedance of the source.
This paper has clearly shown that there is a difference in
the current distortion of a load depending on whether the load
is served by a clean supply or a distorted supply. It has also
shown, contrary to popular belief, that only a linear load has a
higher current distortion when supplied from a distorted
supply. Even a nonlinear load like a VSD can have higher
current distortions when supplied from a distorted supply.
This information about the true current distortion of a load
could be used to persuade an offending load to take steps to
mitigate an unacceptably high level of distortion. On a
practical system the neural network computations can be
carried out on a DSP, together with a suitable A/D interface.
Utilities stand to benefit from this work, since it provides a
tool to model a load under distorted supply conditions and
may help the utility check the accuracy of the load model
provided by a customer during commissioning of a new
service.
ACKNOWLEDGMENT
Financial support by the National Electric Energy Testing
Research and Applications Center (NEETRAC), Atlanta,

USA; and from the Duke Power Company, Charlotte, North

Carolina, USA, is greatly acknowledged.
REFERENCES
[1]

[2]
[3]
[4]
[5]
[6]

[7]

[8]

[9]

M. McGranagham, and D. Mueller, Designing Harmonic Filters for

Adjustable Speed Drives to Comply with IEEE- 519 Harmonic Limits,
IEEE Transactions on Industry Applications, Vol. 5, Issue 2,
pp.312-318, March/April 1999.
IEEE Standard 519-1992, IEEE Recommended Practices and
Requirements for Harmonic Control in Electric Power Systems.
Electromagnetic Compatibility (EMC)-Part 3: Limits-Section
6:Assessment of Emission Limits for Distorting Loads in MV and HV
Power Systems, IEC 1000-3-6,1996.
W. Xu, and Y. Liu, A Method for Determining Customer and Utility
Harmonic Contributions at the PCC, IEEE Transactions on Power
Delivery, Vol. 15, Issue 2, pp.804-811, April 2000.
T. A. George, and D. Bones, Harmonic power flow determination
using the fast Fourier transform, IEEE Transactions on Power
Delivery, Volume: 6 Issue: 2, April 1991, pp. 530 535.
B. Palethorpe, M. Sumner, and D.W.P. Thomas, Power system
impedance measurement using a power electronic converter, in
Proceedings of IEEE International Conference on Harmonics and
Quality of Power, 2000, Orlando, Florida,Vol. 1, pp. 1035 1040, Oct
2000.
G. Moreau, H.H. Le, G. Croteau, G. Beaulieu, and E. Portales,
Measurement system for harmonic impedance of the network and
validation steps, in Proceedings of the IEEE International Symposium
on Quality and Security of Electric Power Delivery Systems,
CIGRE/PES 2003, Montreal, Canada, pp. 69 73, Oct 2003.
J. Mazumdar, R. Harley, F. Lambert and G.K. Venayagamoorthy,
Using a neural network to distinguish between the contributions to
harmonic pollution of nonlinear loads and the rest of the power
system, in Proceedings of the IEEE Power Electronics Specialist
Conference (PESC 2005), Recife, Brazil, June 12 16, 2005.
J. Mazumdar, R. Harley and F. Lambert, System and method for
determining harmonic contributions from non-linear loads, in
Proceedings of the IEEE Industry Applications Society Annual
Meeting (IAS 2005), Kowloon, Hongkong, pp. 2456 - 2463, Oct 2 6,
2005.