Critical Machine Identification for Power Systems

Transient Stability Problems using Data Mining

D. E. Echeverría, J. C. Cepeda D. G. Colomé

Research and Development Department Instituto de Energía Eléctrica
CENACE Universidad Nacional de San Juan
Quito, Ecuador San Juan, Argentina
decheverria@cenace.org.ec, jcepeda@cenace.org.ec colome@iee.unsj.edu.ar

Abstract—This paper presents a new methodology based on Conventionally, critical machines are identified by examining
data mining to identify the cluster of critical machines, i.e. the the relative swing curves between machines. However, it will
machines responsible for the loss of synchronization in a power not be practical even in off-line applications if the critical
system after the occurrence of a disturbance. Since only the post- cluster is searched thoroughly with all the possible combination
fault trajectory is required, the proposed method is independent
of machines. Significant reduction of the number of machine
of system modeling and could be extended for multi-swing
stability assessment. Numerical results obtained by applying the combinations has been obtained by using the sorted trajectories
approach on the New England test system demonstrates the approach [5], [6].
feasibility and effectiveness that could be achieved in identifying In this paper, a data-mining-based clustering for
the critical machines, which is also of great value for assessing identification of critical machines is presented. Based on post-
transient stability problems and defining suitable emergency fault trajectory from time domain simulation or a set of
control actions. measurements acquired in real time (PMU) after occurrence
and clearing of a disturbance or any sudden outage (i.e.
Index Terms — Critical machines, transient stability, data
immediate post-fault conditions), the methodology provides
mining, phasor measurement unit.
relevant information regarding the number and alignment of
critical machines.
I. INTRODUCTION The outline of the paper is as follows: Section II gives some
There are many instances in which power systems are background information on power system modeling issues.
operated close to their security limits. Moreover, under such Section III provides a brief overview of the classical
operating conditions, some faults and contingencies may lead methodology for identification of critical machines. The
to major consequences, including widespread disruptions or proposed approach is described in Section IV. Section V
even blackouts, whose root causes have been occasionally presents test results of the proposed algorithm. Finally, Section
attributed to large-disturbance (i.e. transient) stability problems VI gives brief conclusion of the paper.
[1].
The most commonly used methodologies, in order to assess II. MODEL OF MULTI-MACHINE OF POWER SYSTEM
the transient stability problems, include [2]: the time domain Mathematically, for each synchronous generator in a power
simulation, transient energy function, extended equal area system, the rotor angle δi (i=1,2,...n) is determined by the
criterion (EEAC). Among these methods, time domain swing equation [7]:
simulation is generally regarded as the best available analysis d δ i (t )
tool for its accuracy, reliability and modeling capability. = ωi (t ) − ωo
dt (1)
However, it is computational intensive and is incapable of
providing a quantitative stability margin. As a remedy, various d ωi (t ) 1
hybrid methods, which combine the time domain simulation
= [ Pmi (t ) − Pei (t )] (2)
dt Mi
with either transient energy function or equal area criterion, where Mi is the moment of inertia, Pmi is the mechanical
have been proposed and received considerable attention. For power input, Pei is the electrical power output, and ωi is the
instance, Single Machine Equivalent (SIME) methodology has speed of the generator rotor, corresponding to machine i.
shown excellent performance for assessing the post- Equations (1) and (2) can be approximated by using the
contingency transient stability system status [3]. Recently,
trapezoidal rule. Thus, the values of speed and rotor angles for
detection of loss of synchronism using phasor measurements
each generator, at each time instant tk, are determined as
(PMUs) has also given a new perspective of transient stability
assessment [4]. follows [8]:
The identification of cluster of critical machines is of most Δt = tk − tk-1
(3)
important in transient stability assessment and control.

978-1-4799-6251-8/14/$31.00 ©2014 IEEE

 Pmi (tk ) − Pei (tk )
Δωi (tk ) = Δt (4)
Mi
ωi (tk ) = ωi (tk-1 ) + Δωi (tk ) (5)
1
δi (tk ) = δi (tk-1 ) + [ωi (tk-1 ) + ωi (tk )] Δt (6)
2
III. CLASSICAL METHOD FOR DETERMINATION OF CRITICAL
MACHINES
The notion of critical machines is intimately related to
unstable scenarios. By definition, on such an unstable multi-
machine trajectory, the critical machines are those which go out
Fig. 2. Variation of critical machines in the time
of step i.e., which cause the system loss of synchronism.
The methodology used in [9] to determine the cluster of
critical machines is presented in this section. IV. NEW METHOD FOR DETERMINATION OF CRITICAL
On an unstable multi-machine trajectory, the identification MACHINES
of critical machines uses the following pattern: (i) at each time The methodology uses the data of rotor angles of
step, consider the post-fault time evolution of the system synchronous machines, which can be obtained from the time
machines; (ii) sort these machines in decreasing order of their domain simulations or real-time measurements PMU. First, an
rotor angles, identify the largest angular deviation (largest angle dissimilarity matrix (D) is built by applying the
“gap”) between any two adjacent machines thus sorted, and Recursive Method for Online Coherency Calculation
consider the candidate critical machines to be those which are (RMOCC) presented in [10], whose general procedure is
“above this largest gap”. summarized as follows.
In Fig. 1, an example of the identification of the cluster of A value Xi(t) on bus i at time t is defined by (7), in this way:
critical machines from the "largest gap" can be appreciated.
X i ( t ) = xi ( t ) − xi 0 − X (7)

where xi(t) represents the rotor angle on machine i at time t;

xi0 is its initial value prior to the perturbation; and X is the
average value over the number of generators Ng at time t.

Ng

∑ x (t )
t
X= i=1
(8)
Ng

Xij(t) is then defined as the difference of X(t) between

generators buses i and j.
Fig. 1. Identification of critical machines [9]

The presented methodology is very robust when the cluster X ij ( t ) = X i ( t ) − X j ( t ) (9)

of critical machines does not change over time, i.e., when the
"largest gap" does not change over time. Then, a value that represents the electric distance between
buses i and j is computed as follows:
Fig. 2 presents an example where there are changes in the
"largest gap", and therefore also changes in the cluster of T
1
critical machines. For this reason, the methodology is not d ij = ∫ X ij 2 ( t ) dt (10)
robust when there are oscillations that may change the cluster T tcl
of critical machines over time.
In the next section a new methodology is presented to
identify the set of critical machines in order to overcome this where tcl is the fault-clearing time, and T is the end of the
drawback. analyzed time window.
Then, a dissimilarity matrix (D) is built by applying the
described RMOCC.
 ⎡0 d12 ... d1n ⎤ M C1 = ∑ M i (15)
⎢d 0 ... d 2 n ⎥⎥ i∈C 1
D = ⎢ 21 (11)
⎢ .
⎢
. ... . ⎥
⎥
MC 2 = ∑M
j∈C 2
j (16)
⎣ d n1 dn2 ... 0 ⎦
Where MC1 and MC2 are the partial centers of inertia, Mi and
Mj are the inertia of each machine belonging to each cluster. In
where n is the number of generators, and dij is the distance
order to determine which cluster is critical or non-critical, the
between bus i and bus j determined by equation (10).
following inequality is written.
After building the dissimilarity matrices, the corresponding
Similitude Matrices (Q) can be obtained by (12), where I is the
if M C1 ≥ M C 2 ⇒ C1is the critical machine cluster
identity matrix and 1 is the all-ones vector [11], [12]. These (17)
matrices represent the variability between elements (similar to if M C1 < M C 2 ⇒ C1is the non-critical machine cluster
Covariance Matrices representing the variability between
variables). Next, in order to include the historical information of post-
fault trajectory, two membership functions (ucl={cm,ncm}) that
1⎡ 1 ⎤ ⎡ 1 ⎤ reflect the possibility of pertinence in the critical or non-critical
Q = − ⎢ I − 11′⎥ D ⎢I − 11′⎥ (12) cluster throughout the entire trajectory are defined. These
2⎣ n ⎦ ⎣ n ⎦
membership functions represent the cumulative possibility of
Using eigenvalues (Λr) and eigenvectors (Vr) of the matrix machine i belongs to the critical (cm) or non-critical (ncm) sets,
Q, it is possible to calculate their corresponding principal as shown by (18).
coordinates [11], [12], [13], and put them together to form a jcl ={cm ,ncm}
multi-dimensional space data matrix (X), as shown by (13) ucl ={cm ,ncm} = (18)
and (14). Yr is a matrix that contains the new orthogonal n
variables that keep the original distances (principal where, n is the number of data points, j is the number of
coordinates) [11], [12], [13]. times that generator i belongs to cluster cl (cm or ncm).
Thus, the pertaining of each generator to critical or non-
Yr = Vr Λ r1/ 2 (13) critical clusters will be decided when ucl > 0.5.
Xδ = Yrδ (14)
V. SIMULATION RESULTS
Once the data matrix Xδ (δ stands for rotor angle data) is The proposed approach is tested on the New England
determined, its multi-dimensional data points are grouped into system [15], whose single line diagram is shown in Fig. 3. The
two clusters using the Fuzzy C-Means (FCM) algorithm. This implementation was accomplished by using DIgSILENT
clustering algorithm allows partitioning the machines into Power Factory© [16] along with several routines written in
critical and non-critical groups. MATLAB® [17]. The case studies presented in this section
FCM is a method that uses fuzzy partitioning in order to focus on analyzing the post-fault trajectory and identifying the
allow finding groups of data where each data point set of critical machines by using the proposed algorithm. The
(observation) belongs to a cluster with a certain degree of contingencies are based on three-phase short circuits applied in
membership grade. Thus, a given data point can belong to the lines near the buses at t=0.1 s, followed by the fault
several groups with its own degree of membership between 0 clearance at t=0.22 s.
and 1, whose general formulation can be found in [14].
An additional improvement of the proposed method, as
regards the classical method, is its ability to improve the results
while more samples are analyzed. Thus, the dissimilarity
matrix is updated after each sample arrives to the control center
and new clusters are computed. This cluster upgrade allows
verifying the possibility of changing the cluster distribution
while the system dynamics evolution.
Since there is the possibility of one or more machines to not
be clear about its cluster membership (due to dynamic
closeness to critical and non-critical machines), an additional
analysis is done, in order to include the historical information
of past samples, as well as the influence of the generator’s
inertia in suggesting the critical and non-critical machines.
First, the machines are partitioned into two clusters by
means of FCM, say C1 and C2, and their partial centers of
inertia are determined as follows.
Fig. 3. New England Test System
 The analyzed case study consists on a three phase short TABLE I. MEMBERSHIP FUNCTIONS OF CUMULATIVE POSSIBILITY
OF CRITICAL MACHINE PERTAINING
circuit applied in line 2-25 near bus 2 at t=0.10 s, and the fault
clearance at t=0.22 s. G1 G2 G3 G4 G5 G6 G7 G8 G9 G10
For this specific case, the post-fault rotor machine angles, u cm 1.000 1.000 1.000 1.000 1.000 1.000 1.000 0.000 0.000 1.000
obtained from Power Factory time-domain simulation, are used u ncm 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.000 1.000 0.000
to structure the corresponding dissimilarity matrix, as explained u cm 1.000 1.000 1.000 1.000 1.000 1.000 1.000 0.000 0.000 1.000
in Section IV. Then, the corresponding principal coordinate u ncm 0.000 0.000 0.000 0.000 0.000 0.000 0.000 1.000 1.000 0.000
matrix (Yr) is also determined. u cm 1.000 0.667 0.667 0.667 0.667 0.667 0.667 0.000 0.000 1.000
After, the depicted clustering algorithm, including the u ncm 0.000 0.333 0.333 0.333 0.333 0.333 0.333 1.000 1.000 0.000
proposed formulation for including the historical information u cm 1.000 0.750 0.750 0.750 0.750 0.750 0.750 0.000 0.000 1.000
of past samples, as well as the influence of the generator’s u ncm 0.000 0.250 0.250 0.250 0.250 0.250 0.250 1.000 1.000 0.000
inertia in suggesting the critical and non-critical machines, is u cm 1.000 0.800 0.800 0.800 0.800 0.800 0.800 0.000 0.000 1.000
u ncm 0.000 0.200 0.200 0.200 0.200 0.200 0.200 1.000 1.000 0.000
applied to Yr.
u cm 1.000 0.833 0.833 0.833 0.833 0.833 0.833 0.000 0.000 1.000
TABLE I presents the results of the proposed membership u ncm 0.000 0.167 0.167 0.167 0.167 0.167 0.167 1.000 1.000 0.000
functions that allow considering the cumulative possibility of u cm 1.000 0.857 0.857 0.857 0.857 0.857 0.857 0.000 0.000 1.000
machine i belongs to the critical (cm) or non-critical (ncm) sets. u ncm 0.000 0.143 0.143 0.143 0.143 0.143 0.143 1.000 1.000 0.000
It is possible to appreciate that generators G2 – G7 present u cm 1.000 0.875 0.750 0.750 0.750 0.750 0.750 0.000 0.000 1.000
some degree of membership of belonging to critical or non- u ncm 0.000 0.125 0.250 0.250 0.250 0.250 0.250 1.000 1.000 0.000
critical machines during the signal dynamic evolution. This u cm 1.000 0.889 0.778 0.778 0.778 0.778 0.778 0.000 0.000 1.000
situation is due to the fact that, for the presented contingency, u ncm 0.000 0.111 0.222 0.222 0.222 0.222 0.222 1.000 1.000 0.000
these generators show dynamic closeness to critical and non- u cm 1.000 0.900 0.800 0.800 0.800 0.800 0.800 0.000 0.000 1.000
critical sets. Nevertheless, the proposed algorithm is capable of u ncm 0.000 0.100 0.200 0.200 0.200 0.200 0.200 1.000 1.000 0.000
correctly identifying the critical machines, since the degree of u cm 1.000 0.909 0.818 0.818 0.818 0.818 0.818 0.000 0.000 1.000
membership is always larger than 0.5. u ncm 0.000 0.091 0.182 0.182 0.182 0.182 0.182 1.000 1.000 0.000
Fig. 4 shows the results of the final critical machines u cm 1.000 0.917 0.833 0.833 0.833 0.833 0.833 0.000 0.000 1.000
u ncm 0.000 0.083 0.167 0.167 0.167 0.167 0.167 1.000 1.000 0.000
identification for this case.
u cm 1.000 0.923 0.846 0.846 0.846 0.846 0.846 0.000 0.000 1.000
u ncm 0.000 0.077 0.154 0.154 0.154 0.154 0.154 1.000 1.000 0.000
u cm 1.000 0.929 0.857 0.857 0.857 0.857 0.857 0.000 0.000 1.000
u ncm 0.000 0.071 0.143 0.143 0.143 0.143 0.143 1.000 1.000 0.000
u cm 1.000 0.933 0.867 0.867 0.867 0.867 0.867 0.000 0.000 1.000
u ncm 0.000 0.067 0.133 0.133 0.133 0.133 0.133 1.000 1.000 0.000
u cm 1.000 0.938 0.875 0.875 0.875 0.875 0.875 0.000 0.000 1.000
u ncm 0.000 0.063 0.125 0.125 0.125 0.125 0.125 1.000 1.000 0.000
u cm 1.000 0.941 0.882 0.882 0.882 0.882 0.882 0.000 0.000 1.000
u ncm 0.000 0.059 0.118 0.118 0.118 0.118 0.118 1.000 1.000 0.000
u cm 1.000 0.944 0.889 0.889 0.889 0.889 0.889 0.000 0.000 1.000
u ncm 0.000 0.056 0.111 0.111 0.111 0.111 0.111 1.000 1.000 0.000
u cm 1.000 0.947 0.895 0.895 0.895 0.895 0.895 0.000 0.000 1.000
u ncm 0.000 0.053 0.105 0.105 0.105 0.105 0.105 1.000 1.000 0.000
u cm 1.000 0.950 0.900 0.900 0.900 0.900 0.900 0.000 0.000 1.000
u ncm 0.000 0.050 0.100 0.100 0.100 0.100 0.100 1.000 1.000 0.000
Fig. 4. Example of identification of critical machines
u cm 1.000 0.952 0.905 0.905 0.905 0.905 0.905 0.000 0.000 1.000
From the results obtained through the application of the u ncm 0.000 0.048 0.095 0.095 0.095 0.095 0.095 1.000 1.000 0.000
proposed methodology, it is possible to see that the generators u cm 1.000 0.955 0.909 0.909 0.909 0.909 0.909 0.000 0.000 1.000
G8 and G9 are the critical machines. The methodology u ncm 0.000 0.045 0.091 0.091 0.091 0.091 0.091 1.000 1.000 0.000
determines the cluster of machines from the beginning, because u cm 1.000 0.957 0.913 0.913 0.913 0.913 0.913 0.000 0.000 1.000
u ncm 0.000 0.043 0.087 0.087 0.087 0.087 0.087 1.000 1.000 0.000
it is based on the trajectory of the rotor angle and not in the
u cm 1.000 0.958 0.917 0.917 0.917 0.917 0.917 0.000 0.000 1.000
"largest gap". This allows having a more robust algorithm for
u ncm 0.000 0.042 0.083 0.083 0.083 0.083 0.083 1.000 1.000 0.000
determining critical machines and using it together with other
u cm 1.000 0.960 0.920 0.920 0.920 0.920 0.920 0.000 0.000 1.000
methodologies, like SIME, in order to assess the post- u ncm 0.000 0.040 0.080 0.080 0.080 0.080 0.080 1.000 1.000 0.000
contingency system transient stability.

The same procedure is applied to several contingencies in

order to validate the proposed method. TABLE II summarizes
the results for all the analyzed contingencies.
 of pertaining of each machine to the critical and non-critical
TABLE II. CRITICAL MACHINES sets has also been included.
Simulation results highlight the good performance of the
Contingency Bus Transmission Critical Machines
No. No. Line proposed method in order to adequately determine the critical
1 2 2-3 G8,G9 and non-critical clusters of machines, overcoming the
2 2 2-25 G8,G9 drawback of the traditional approach.
3 4 4-14 G2,G3 With the rapid development of wide-area measurement
4 5 5-6 G2,G3 systems (WAMS), the proposed method shows a promising
5 6 6-11 G2,G3
future in its on-line applications on large power systems.
6 7 6-7 G2,G3
7 10 10-13 G2,G3 Further research is being carried out in order to incorporate the
8 11 10-11 G2,G3 proposed method to the assessment of transient stability in real-
9 13 13-14 G2,G3 time.
10 16 16-21 G4,G5,G6,G7,G9
11 17 16-17 G4,G5,G6,G7,G9 REFERENCES
12 21 21-22 G6,G7 [1] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose, C.
13 22 22-23 G6,G7 Canizares, N. Hatziargyriou, D. Hill, A. Stankovic, C. Taylor, T.
14 22 21-22 G6,G7
Van Cutsem, and V. Vittal, “Definition and classification of
15 23 22-23 G6,G7
power system stability IEEE/CIGRE joint task force on stability
16 24 23-24 G4,G5,G6,G7,G9
terms and definitions”, IEEE Transactions on Power Systems,
17 25 2-25 G8,G9
Vol. 19, No. 3, Aug. 2004, pp. 1387 – 1401.
18 25 25-26 G8,G9
19 26 26-27 G9 [2] M. Ribbens-Pavella, P.G. Murthy, “Transient Stability of Power
20 27 17-27 G9 Systems”. John Wiley and Sons, England, 1994.
21 28 26-28 G9 [3] Y. Zhang, L. Wenhenkel, M. Pavella, “SIME: A hybrid
approach to fast transient stability assessment and contingency
Additional comparison of the proposed data-mining-based selection”, Electrical Power & Energy Systems, Vol. 19 No. 3,
algorithm to the classical one has been also carried out. Results pp. 195-208, 1997.
have shown that, for cases 2, 10, and 11, the classical algorithm [4] D. E. Echeverria, J. L. Rueda, J. C. Cepeda, D. G. Colome, I.
have presented confusion in determining the critical machines, Erlich, “Comprehensive approach for prediction and assessment
since in these cases the signal dynamic behavior of some of power system transient stability in real-time”, Innovative
Smart Grid Technologies Europe (ISGT EUROPE), IEEE/PES,
generators is close to critical and non-critical ones. On the
2013.
other hand, the application of the proposed algorithm have
[5] W.P. Yuan, K.W. Chan, Y. Zhang, “Identification of Critical
shown excellent performance in correctly identifying the
Cluster in Transient Stability Study Using Line Potential Energy
critical machines, since it considers the cumulative possibility Method”, Power Engineering Society General Meeting, IEEE,
of pertaining of the entire signal trajectory, instead of only the 2006.
actual data point and its largest gap. [6] K.W. Chan, Q. Zhou, T.S. Chung, “Transient Stability Margin
This dynamic behavior can be observed, for case 2, in Fig. Assessment for Large Power System Using Time Domain
2. Simulation Based on Hybrid Extended Equal Area Criterion
Method”, APSCOM’2000, pp. 405-409, 2000.
VI. CONCLUSIONS
[7] P. Kundur, “Power System Stability and Control”, New York:
In this paper, a novel approach based on data-mining is McGraw-Hill, USA, 1994.
proposed for the identification of critical clusters of machines. [8] F. Milano, “Power System Modelling and Scripting”, Berlin:
This method has given a new perspective to the transient Springer, Germany, 2010.
stability assessment and critical machines’ identification, [9] M. Pavella, D. Ernst, and D. Ruiz-Vega, “Transient Stability of
because it is based on the trajectory of the rotor angle, instead Power Systems: A Unified Approach to Assessment and
of the "largest gap" of classical approach. This allows having a Control”, Norwell, Kluwer, 2000.
more robust algorithm for determining critical machines, [10] I. Kamwa, A. K. Pradham, G. Joos, and S. R. Samantaray,
whose results might be used as input of other methodologies “Fuzzy Partitioning of a Real Power System for Dynamic
like SIME for assessing transient stability. When real-time Vulnerability Assessment”, IEEE Transactions on Power
measurements are used, there is no assumption of power Systems, vol. 24, No. 3, pp. 1356 – 1365, August 2009.
system models, and only a local measurable network data is [11] J.C. Cepeda, J.L. Rueda, I. Erlich, and D.G. Colomé,
required. "Probabilistic Approach-based PMU placement for Real-time
Power System Vulnerability Assessment," in Proc. 2012 IEEE
In addition, the proposed algorithm includes additional
PES Innovative Smart Grid Technologies (ISGT) Europe
formulation to improve the results while more samples are Conference, Berlin, Germany, Oct. 2012.
analyzed. This formulation considers the historical information
[12] J.C. Cepeda, and D.G. Colomé, “Vulnerability Assessment of
of past samples, as well as the influence of the generator’s Electric Power Systems through identification and ranking of
inertia in suggesting the critical and non-critical machines. In Vulnerable Areas”, International Journal of Emerging Electric
this connection, a membership function that shows the degree Power Systems, Volume 13, Issue 1, May 2012, ISSN (Online)
1553-779X, DOI: 10.1515/1553-779X.2846.
[13] D. Peña, Análisis de Datos Multivariantes, Editorial McGraw-
Hill, España, cap. 1 – 8.
[14] K. Hammouda, and F. Karray, “A Comparative Study of Data
Clustering Techniques”, University of Waterloo, Ontario,
Canada, [Online]. Available at:
http://www.pami.uwaterloo.ca/pub/hammouda/sde625-
paper.pdf.
[15] M. Pai, “Energy Function Analysis for Power System Stability”,
Kluwer, Norwell, MA, 1989.
[16] DIgSILENT Power Factory (2011). User’s Manual Version
14.1. Edición 1, DIgSILENT GmbH, Gomaringen, Germany.
[17] MATLAB, Math Works, Inc., Natick, MA, USA.
www.mathworks.com