Journal of Electrical Engineering

www.jee.ro

SPEED DEVIATION AND MULTILAYER PERCEPTRON NEURAL

NETWORK BASED TRANSIENT STABILITY STATUS PREDICTION
SCHEME
Emmanuel A. Frimpong1, Johnson Asumadu2 and Philip Y. Okyere3
1,3
Kwame Nkrumah University of Science and Technology, Kumasi, Ghana
2
Western Michigan University, Kalamazoo, Michigan, USA
1
[email protected], [email protected], [email protected]

Abstract: In this paper, a rotor speed deviation and rotational speed and may lose synchronism.
multilayer perceptron neural network (MPLNN) based Generators losing synchronism must be taken off line
transient stability prediction scheme is presented. The to avoid catastrophic problems. Whenever generators
scheme uses the sum of the maximum rotor speed deviations are taken off line, capacity decreases, thus introducing
(MSDs) of the individual generators in a power system as
another large disturbance leading to cascading system
inputs to an MPLNN. The proposed scheme predicts
transient stability one cycle after the tripping of a bus or line failures. This could cause equipment damage, pose
following a disturbance. The trained MLPNN responded to safety hazards to personnel, contribute to cascading
56 transient unstable cases with 100% accuracy. The outages, and the shutdown of large areas of a power
response to 41 transient stable cases was also 100% system or the entire system [2, 3]. Control measures
accurate. The IEEE 39-bus test system was used for the such as out-of-step blocking and tripping, fast-valve
study. control of turbines, dynamic braking, superconducting
magnetic energy storage systems, system switching,
Key words: Power system stability, Stability prediction, modulation of high voltage direct current (HVDC) link
Transient stability, Artificial Neural networks power flow, and load shedding are employed to
mitigate the effect of cascading system failures [4].
1. Introduction The effectiveness of the aforementioned control
Industrialization, population growth, and measures are improved with the prediction of transient
modernisation have led to a huge demand for electric instability [4]. To this end, researchers have come up
power. This huge demand coupled with inadequate with a number of transient instability prediction
generation capacity has resulted in most power systems schemes [4-21]. These schemes employ decision trees
operating their generators with reduced stability [5], neural networks [4, 6-9], neural networks and
margins, thus making such systems weak [1].
fuzzy logic [10-14], support vector machines [15],
Consequently, the occurrence of a disturbance
wavelet analysis [16], apparent impedance [17],
endangers the stability of such systems.
numerical routines or state space techniques [18-20]
Under normal operating conditions, an electrical
and autoregression [21].
power system is near equilibrium, with only minor
The methods [18-20] are computationally
deviations from true steady-state conditions caused by
demanding for on-line application. Also, the Decision
small, nearly continuous, changes in load. When a
tree based technique has limited forecast accuracy.
large disturbance such as a three-phase short circuit
Additionally, the proposed techniques have delayed
occurs in a power network, there are significant, nearly
prediction times
instantaneous rise in power requirement from some
In this paper, a generator speed deviation and
generators. Instead of the power system returning to a
multilayer perceptron neural network based transient
steady-state condition after the disturbance, one or
stability status prediction scheme is proposed. The
more generators may encounter sufficient variations in
scheme predicts transient stability status 20ms (1 cycle

1
 Journal of Electrical Engineering
www.jee.ro

for a 50Hz) after the tripping of a bus or line in d 2

response to a disturbance. The scheme can be easily M Pm Pe (1)
realized with the aid of phasor measurement units dt 2
(PMUs) which can communicate time-tagged phasor where is the rotor angle, Pm is mechanical power,
measurements such as rotor speed deviations to a Pe is electrical power and M is the inertia
central location. This enables the tracking of the coefficient.
dynamic state of a system in real time. The proposed Rotor angles have been extensively used for
scheme uses the sum of the individual maximum rotor transient stability studies. Rotor angles need to be
speed deviation of each generator as input to a trained expressed relative to a common reference. This
MLPNN which does the prediction. reference cannot be based on a single generator, since
any instability in the reference generator makes the
2. Used power system configuration relative angles meaningless. In order to overcome this
The transient stability status prediction scheme was difficulty, the concept of system center of inertia (COI)
developed using the IEEE 39-bus test system which is angle, co defined in equation 2 is used to obtain a
also known as the New England test system. The IEEE
reference angle.
39-bus test system is a standard test system that is
n H
widely used for small and large signal stability studies co i n1 i i (2)
[4]. The test system consists of 10 generators, one of i 1 H i
which is a generator representing a large system. Data where i and Hi are the rotor angle and inertia constant
for the modeling of the test system was obtained from of the ith generator, respectively. The angle, i is
[22]. The test system is shown below as Fig. 1. usually approximated by the phase angle of the
respective generator bus voltage [25 and 26]. Many
researches however discourage the use of rotor angles
in algorithms. This is because the COI values, in
practice require continuous updates using real time
measurements. This requires extra pre-processing and
has significant errors [25]. Rotor angles, thus best
serve as the reference parameter for telling stability
status of a system in a simulation. Other electrical
parameters whose use in algorithms, do not have
practical constraints may then be employed for
algorithm development.
The time derivative of rotor angle is the rotor
speed deviation in electrical radians per second [23,
and 26]. Mathematically,
d
s (3)
dt
where is the rotor speed deviation, is the rotor
speed at a particular time, and s is the synchronous
speed. It follows from equations (1) and (3) that the
swing equation can be written as
Fig. 1 IEEE 39-bus Test System d
M Pm Pe (4)
dt
3. Rotor speed deviation
Equation 1 shows the fundamental equation
governing rotor dynamics. This equation is commonly
referred to as the swing equation [23 and 24].

2
 Journal of Electrical Engineering
www.jee.ro

It can also be shown that

1
d 0 2
0 Pa d (5)
dt H
where H is the inertia constant and Pa is the difference
between input mechanical power and output
electromagnetic power. For stability to be attained after
d
a disturbance, it is expected that 0 in the first
dt
swing. This equation gives rise to the equal area
criterion which is a well-known classical transient
stability criterion. From equations (3) and (5), it can be
written that
1
2
0 0 Pa d (6)
H Fig. 2 Speed deviations for a stable swing for a three-phase
fault between buses 11 and 6
Equation (6) then suggests speed deviation as a good
input parameter for the prediction of transient stability
status. The maximum speed deviation at some time
during a disturbance can be used to predict transient
instability or otherwise. The best time is within the first
swing, like the equal area criterion.
This paper demonstrates the validity of rotor speed
deviations for transient stability prediction. Fig. 2 and
Fig. 3 show waveforms of rotor speed deviations for
transient stable and transient unstable cases. It is noted
from Fig. 2 and Fig. 3 that unstable swings are
characterised by higher rotor speed deviations
compared with stable swings. Consequently, the sum
of the maximum speed deviations of the generators for
an unstable swing is greater than the sum of the
maximum speed deviations of the generators for a
stable swing. Fig. 3 Speed deviations for an unstable swing for a three-
Thus, the sum of the maximum speed deviation of phase fault between buses 11 and 6
the individual generators of a system following a
disturbance can be a suitable input data for the 4. Multilayer perceptron neural network
prediction of transient stability status. The input data Artificial Neural networks (ANNs) are constructed to
used for the proposed scheme is given as follows: make use of some organizational principles resembling
N those of the human brain [27]. They represent a
x Maxi , i 1, 2, 3,..., N (7) promising new generation of information processing
i 1 systems. Neural networks are good at tasks such as
Where x is the input data, i is rotor speed pattern-matching and classification, function
deviation, Max(i) is the maximum rotor speed approximation, optimization and data clustering [27].
ANNs can be used to extract patterns and detect trends
deviation, and N is the number of generators.
that are too complex to be noticed by either humans or
other computer techniques [28]. ANNs are made up of

3
 Journal of Electrical Engineering
www.jee.ro

a number of simple and highly interconnected yx (9)

processing elements called neurons, as shown in Fig. 4. The hidden layer had two neurons with tangent
The mathematical model of a neuron is expressed as sigmoid transfer functions. The output, y of a tangent
[28]: sigmoid transfer function for a given input x is given
N
O j f j w jk x k , j k 1, 2, 3,..., N (8) as:
k 1
y (10)
where, O j is the output of a neuron, f j is a transfer 1 e x
function, which is differentiable and non-decreasing, The output had one neuron with a transfer function
usually represented using a sigmoid function, w jk is an which is purelin. The MLPNN was trained to give an
output, O , of 0 for a swing that will be transient
adjustable weight that represents the connection stable and an output of 1 for a swing that will be
strength, and xk is the input of a neuron. transient unstable. The MLPNN was trained using the
Levenberg-Marquardt back-propagation technique.
The Levenberg-Marquardt algorithm trains a
neural network 10 to 100 times faster than the
usual gradient descent back propagation method.
This algorithm is an approximation of Newtons
method and it computes the approximate Hessian
matrix [30]

5. Simulations
Fig. 4 Mathematical model of a neuron The modeling and simulation of the test system
were carried out using the Power System Simulator for
A three-layer feed forward multilayer perceptron Engineers (PSSE) software [31]. Three-phase faults
neural network (MLPNN) with no bias was used for were created at various buses and on various lines. The
this study. Fig. 5 shows the architecture of the MLPNN simulations were carried out for four different loading
used. The choice was informed by the fast decision levels. The levels are base load, base load increased by
making capability of MLPs [29].
5%, base load increased by 7%, and base load
increased by 10% [25].
A total of 107 fault cases were simulated out of
which 61 resulted in transient instability while 46 were
transient stable cases. A system was seen as being
transient unstable if the rotor angle difference between
any two generators exceeds 180 degrees 1 second after
fault clearing time [4]. The output data (for analysis)
from the simulations were generator speed deviations
sampled using a sampling frequency of 6kHz. All
Fig. 5 A three-layered MLPNN
transient stable cases had fault durations of 0.1s while
The input layer had one neuron with a transfer transient unstable cases were realized for faults lasting
function, which is purelin. The input data x , was the between 0.7s and 0.9s. These times are similar to that
sum of the maximum speed deviations of the 10 reported in [4].
generators, in one cycle after the tripping of a bus or
line. 6. Data analysis
The output, y of a purelin transfer function for a The analysis of the output data (rotor speed
given input x is given as:
deviations) was done using the MATLAB software

4
 Journal of Electrical Engineering
www.jee.ro

[32]. In MATLAB, the speed deviations of each conditions after the disconnection of a bus following a
generator after the tripping of a line or bus were further bus fault. Table 4 shows sum of MSDs for stable
sampled using a time window of 20ms. For each cycle conditions after the disconnection of a bus following a
(sample window), the maximum speed deviation bus fault. A study of the sum of MSDs for the various
(MSD) was obtained for each generator. The obtained cases show that cases which led to transient instability
maximum speed deviations of each generator were had a much higher value compared to cases which did
then added. Table 1 shows sum of MSDs for unstable not result in transient instability. Thus the sum of the
conditions within one cycle after the tripping of maximum speed deviations of the individual generators
various lines following three-phase faults. Table 2 one cycle after the tripping of a line or bus can be used
shows sum of MSDs for stable conditions within one as an input data for the prediction of transient stability
cycle after the tripping of various lines following three- status.
phase faults. Table 3 shows sum of MSDs for unstable

Table 1: Sum of maximum speed deviations for various line faults (unstable cases)
Bus Bus 11 6 10 13 13 10 13 14 22 21 22 23 26 25 29 26 28 26
Sum of MSDs 0.0324 0.0326 0.0313 0.0312 0.0283 0.035 0.0283 0.0165 0.0155

Table 2: Sum of maximum speed deviations for various line faults (stable cases)
Bus Bus 11 6 10 13 13 10 13 14 22 21 22 23 26 25 29 26 28 26
Sum of MSDs 0.0047 0.0047 0.0045 0.0045 0.0041 0.0041 0.0036 0.0036 0.0019

Table 3: Sum of maximum speed deviations for various bus faults (unstable cases)
Bus 11 13 28
Sum MSD 0.0324 0.0312 0.0137

Table 4: Sum of maximum speed deviations for various bus faults (stable cases)
Bus 11 13 28
Sum of MSD 0.0047 0.0045 0.0019

7. Transient stability status prediction scheme The output of the MLPNN like any other neural
The proposed transient stability status prediction network in the testing phase usually has an error with
scheme uses a feedfoward multilayer perceptron respect to its actual binary value. A similar situation is
artificial neural network with 1 input neuron, two observed in digital communication networks, where the
hidden layer neurons and 1 output neuron [33]. The received bits have some deviation with respect to the
input to the neural network is the sum of the maximum sent bits. In these networks, the TTL standard is
speed deviation (SMSD) of the individual generators of usually used in the receiving equipment to detect the
the power system one cycle after the tripping of a line received bits. This standard is also used to determine
or bus following a disturbance. The network was the output status of the MLPNN [4].
trained with sum of maximum speed deviation (SMSD) O j 0.8 O j 1 (Transient unstable) (3)
data obtained from five transient unstable cases and
SMSD data obtained from five transient stable cases. O j 0.2 O j 0 (Transient stable) (4)
The clear distinction between the transient stable data where O j is the output of a MLPNN.
as against the transient unstable data permitted the use In the digital communication networks, if the value
of a small volume of training data set. The output pair of a received bit is in the range of 0.2 to 0.8, it is
of each of the five transient unstable case data was 1 considered as a missing bit. Besides, if a 1 bit is
while that of the transient stable data was 0. received in the range of 0 to 0.2 or a 0 bit is received

5
 Journal of Electrical Engineering
www.jee.ro

in the range of 0.8 to 1 it is considered an error bit, References

which is a worse incorrect case than the missing bit. 1. Frimpong E. A., Okyere P. Y. and Anto E. K.:
This interpretation for the error and missing bits is also Adaptive Single-Pole Autoreclosure Scheme based
used for the output of the MLPNN. on Standard Deviation and Wave Energy, In:
A flowchart of the proposed transient stability Journal of Electrical Engineering, Vol. 9, 2009,
status prediction scheme is shown below as Fig. 6. No. 4 p. 61-68.
2. Power Systems Transient Stability A Grand
Computing Challenge. NPAC Technical Report
SCCS 549, August 1992.
3. Power Swing and OOS Considerations on
Transmission Lines, A report to the Power System
Relaying Committee of the IEEE Power
Engineering Society, 2005.
4. Amjady N. and Majedi S. F.: Transient Stability
Prediction by a Hybrid Intelligent System. In:
IEEE Transaction on Power Systems, Vol. 22, No.
3, Aug. 2007, p. 1275 -1283.
5. Rovnyak S., Kretsinger S., Thorp J., and Brown
D.: Decision trees for real-time transient stability
prediction, In: IEEE Transaction on Power
Systems, Vol. 9, No. 3, August 1994, p. 1417
1426.
6. Ebrahimpour R., Abharian E. K., Moussavi S. Z.
and Montie Birjandi A. A. M.: Transient Stability
Fig. 6 Flowchart of transient stability status prediction Assessment of a Power System by Mixture of
scheme Experts. In: International Journal of Engineering,
Vol. 4. No. 1, April 2010, p. 93-104.
The trained MLPNN, responded to the 56 transient 7. Ebrahimpour R. and Abharian E. K.: An Improved
unstable cases with 100% accuracy. The responses to Method in Transient Stability Assessment of a
the 41 transient stable cases were also 100% accurate. Power System Using Committee Neural Networks.
In: International Journal of Computer Science and
8. Conclusion Network Security, Vol. 9 No.1, January 2009,
A transient stability status prediction scheme has been p.119-124.
proposed in this paper. The proposed scheme is based 8. Wahab N. I. A, Mohamed A. and Hussain A.: An
on the rotor speed deviations of generators in a power Improved Method in Transient Stability
system 1 cycle after the tripping of a bus or line Assessment of a Power System Using Probabilistic
following a disturbance, and multilayer perceptron Neural Network. In: Journal of Applied Sciences
artificial neural network. The scheme sums the Research, Vol. 3 No. 11, 2007, p. 1267-1274.
maximum rotor speed deviations of the individual 9. Amjady N. and Banihashemi S.A.: IET Digital
generators of a system 1 cycle after the tripping of a Library Transient stability prediction of power
bus or line and uses this sum as input to a trained systems by a new synchronism status index and
multilayer perceptron neural network which predicts hybrid classifier. In: IET Generation, Transmission
the transient stability status. The proposed scheme & Distribution, Vol. 4, No. 4, April 2010, p. 509
predicts transient stability status with 0% error. 518.
10. Liu C. W., Su M. C., Tsay S. S., and Wang Y. J.:
Application of a novel fuzzy neural network to
real-time transient stability swings prediction
based on synchronized phasor measurements.

6
 Journal of Electrical Engineering
www.jee.ro

IEEE Transaction on Power Systems, Vol. 14, No. transient stability prediction. Proceedings of
2, May 1999, p. 685692. Institute of Electrical Engineering, Generation,
11. Amjady N.: Application of a new fuzzy neural Transmission and Distribution, Vol. 142, No. 4,
network to transient stability prediction. In: 1995, p. 335360.
Proceedins of IEEE PES Summer Meeting, San 20. Stanton S. E., Slivinsky C., Martin K. and
Francisco, CA, June. 2005, p. 6976. Nordstorm J.: Application of phasor measurement
12. Liu C. -W., Tsay S. -S., Wang Y.- J, M. C., and partial energy analysis in stabilizing large
Neuro-fuzzy approach to real-time transient disturbances. In: IEEE Transaction on Power
stability prediction based on synchronized phasor Systems, Vol. 10, No. 1, February 1995, p. 297-
measurements. In: Electric Power Systems 306.
Research Vol. 49, 1999, p. 123127. 21. Yamashita K. and Kameda H., Out-of-step
13. Rebizant W., Feser K.: Out-of-step protection with prediction logic for wide-area protection based on
AI methods. In: Proceedings of IEE Seventh an autoregressive model. In: Proceedings of IEEE
International Conference on Developments in PES Power Systems Conference and Exposition,
Power System Protection, Amsterdam, April 2001, New York, 2004, vol. 1, pp. 307-312.
p. 295-298 22. Song Y.: Design of Secondary Voltage and
14. Rebizant W.: ANN based detection of OS Stability Controls with Multiple Control
conditions in power system. In: Proceedings of Objectives. In: PhD. Dissertation, School. of
12th International Conference on Power System Electrical and Computer Engineering, Georgia
Protection PSP2000, Bled, Slovenia, and Institute of Technology, Georgia, USA, 2009.
September. 2000, p. 51-56. 23. Kundur P., et al: Definition and classification of
15. Gomez F. R., Rajapakse A. D. Annakkage U. power system stability, In: IEEE Transactions on
D. and Fernando I.T., Support Vector Power Systems, Vol. 19, No. 2, May 2004, p.
Machine-Based Algorithm for Post-Fault 1387-1400.
Transient Stability Status Prediction. Support 24. Machowski J., Smolarczyk A. and Bialek J. W.:
Vector Machine-Based Algorithm for Post- Damping of power swings by control of braking
resistors. In: Electrical Power and Energy Systems,
Fault Transient Stability Status Prediction
Vol. 23, 2001, p. 539-548.
Using Synchronized Measurements. In: IEEE 25. Rajapakse A. D., Gomez F., Nanayakkara
Transactions on Power Systems, Vol. 26, No. O.M.K.K., Crossley P. A., and Terzija V. V.: Rotor
3, August 2011, p. 1474-1483. Angle Stability Prediction Using Post-disturbance
16. Frimpong, E. A., Okyere, P.Y. and Asumadu, J.: voltage trajectory patterns. In: Power & Energy
Generator out-of-step prediction using wavelet Society General meeting, July 2009, p. 1-6.
analysis, In: 4th IEEE International Symposium on 26. Vaithianathan M. V. and Venkataramana A.:Real-
Power Electronics for Distributed Generation time security assessment of angle stability using
Systems (PEDG), July 2013, p. 1 4. synchrophasors, Final project report, PSERC
17. Li M. , Pal A., Phadke A. G., Thorp J. S., document 10-10, May 2010.
Transient stability prediction based on apparent 27. Uyar M, Yildirim S., Gencoglu M. T.: An effective
impedance trajectory recorded by PMUs. In: wavelet-based feature extraction method for
International Journal of Electrical Power & Energy classificationof power quality disturbance signals.
Systems Vol. 54, January 2014, p. 498504. In: Electric Power Systems Research Vol. 78 2008
18. Bai X., Jiang T., Guo Z., Yan Z., and Ni Y.: A 17471755.
unified approach for processing unbalanced 28. Frimpong E. A. and Okyere P. Y.: Forecasting the
conditions in transient stability calculations. In: Daily Peak Load of Ghana Using Radial Basis
IEEE Transaction on Power Systems, Vol. 21, No. Function Neural Network and Wavelet Transform.
1, February 2006, p. 8590. In: Journal of Electrical Engineering, Vol. 10, No.
19. Liu C. W. and Thorp J. S.: Application of 1, March 2010, p. 15-18.
synchronized phasor measurements to real-time

7
 Journal of Electrical Engineering
www.jee.ro

29. Frimpong E. A., Okyere P. Y. and Anto E. K.:

Adaptive Single-Pole Autoreclosure scheme based
on Standard Deviation and Wave Energy, In:
Journal of Electrical Engineering, Vol. 9, No. 4,
2009, p. 61-68.
30. Rodriguez C. P. and Anders G. J.: Energy price
forecasting in the Ontario competitive power
system market. In: IEEE Trans. on Power
Systems, Vol. 19, No. 3, 2004, p. 366-374.
31. Power System Simulator for Engineers, PSSE
University Edition, 2012.
32. MATLAB R2013a, MathWorks Company, 2013.
33. Frimpong, E. A., Asumadu, J. and Okyere, P.Y.:
Prediction of transient stability status using speed
deviation and multilayer perceptron neural
network. In: 2014 CIGRE Canada Conference,
Toronto, Ontario, September 2014.