Академический Документы
Профессиональный Документы
Культура Документы
Received: November 21, 2011 / Accepted: Feburary 13, 2012 / Published: October 31, 2012.
Abstract: A number of contingencies simulated during dynamic security assessment do not generate unacceptable values of power
system state variables, due to their small influence on system operation. Their exclusion from the set of contingencies to be simulated in
the security assessment would achieve a significant reduction in computation time. This paper defines a critical contingencies selection
method for on-line dynamic security assessment. The selection method results from an off-line dynamical analysis, which covers
typical scenarios and also covers various related aspects like frequency, voltage, and angle analyses among others. Indexes measured
over these typical scenarios are used to train neural networks, capable of performing on-line estimation of a critical contingencies list
according to the system state.
most significant (i.e. critical) according to system required by an on-line implementation of dynamic
operation state. security assessment.
The conventional analytical techniques used in In Ref. [9], a contingency ranking method is
contingency selection and ranking are usually time presented. The contingencies are ranked according to a
consuming and therefore not always suitable for PI (performance index), which is a scalar that reflects
on-line applications. Moreover, many performance the severity degree of a contingency.
indexes based on analytical methods suffer from the In Ref. [10], an approach based on RBFN (radial
problem of misclassification and false alarm. basis function neural network) to rank the
Misclassification arises when a critical contingency is contingencies expected to cause steady state bus
classified as non-critical. A false alarm occurs when a voltage violations is presented.
non-critical contingency is classified as critical. Given In Ref. [11], a three-layer perceptron artificial neural
that, in general, the problem of contingency selection network with back propagation learning technique is
and ranking can be characterized by lack of precise designed for line flow contingency ranking. In this
mathematical model, and large volume of information paper, two new indices are defined (a severity index
to be handled, it would be possible to fulfill the and a margin index for line flow).
requirements of critical contingencies analysis when In Ref. [12], an AI (fuzzy set) based contingency
working on-line by employing AI (artificial ranking is proposed. The post contingency system
intelligence) techniques [6, 7]. variables are first expressed in fuzzy set notation, and
In recent years, there has been a growing interest in then the heuristic rules employed by the system
using AI and specifically artificial neural network operators are used to code in the form of fuzzy
based methods to screen and rank contingencies. In reasoning rules.
relation with the analytical methods, artificial neural In Ref. [13], it used a method based on a coupled
network based methods are faster, requiring a less scheme (artificial neural network and expert system)
computational burden and also they have a better for power system security enhancement, combining
accuracy for on-line contingency ranking [8]. This artificial neural network based contingency screening
paper proposes a novel critical contingencies selection with an expert system based preventive control. An
method for on-line dynamic security assessment. The extended Hopfield neural network is used for
selection method results from an off-line dynamical evaluating the change in the performance index to
analysis, which covers typical scenarios. Indexes determine the ranking of a contingency.
measured over these scenarios are used to train neural The change in the performance index is realized as
networks, capable of performing on-line estimation of the energy function of the extended Hopfield network.
a critical contingencies list according to the system The major disadvantage of the proposed approach
state. [9-13] includes only one issue regarding to dynamic
security assessment in the contingency ranking process
2. State of the Art
(for example: voltage stability, power flow limits, etc.),
Bibliography presents a variety of techniques that regardless other important transients effects, they are
can be used to carry out critical contingencies selection, required to perform a complete dynamic security
and solve the ranking problem, using AI. However, a assessment.
few experiences exists in the state of art, related to
3. Critical Contingencies Selection
implementation of these techniques performing,
quickly and effectively, the necessary tasks are Conformation of critical contingencies list is done
Critical Contingencies Ranking for Dynamic Security Assessment Using Neural Networks 1665
by comparing the severity level in which each scenarios extracted from the base case changing the
contingency affects the system. The greater the load and generation in small steps . Once the
consequences are suffered by the system, the higher the scenarios are defined, a list with all the contingencies to
severity level will be. Due to inherent power system be ranked is created . For each scenario, a power
dynamics, it becomes necessary to analyze the flow calculation is done , all the nodal voltages, line
phenomena involving a failure not only from a currents and power flows results are used to calculate
stationary viewpoint but also from a dynamic one. In the pre-contingency indexes .
addition, the contingencies severity levels as a measure These indexes represent: the system state before
of all aspects (stationary and dynamic) need to be contingency happens, and the failure characteristics.
established, when a failure in some system component More detail of these indexes is given in the next
appears. The severity level indicator approach sections. Using the obtained scenarios and the failure
presented in this paper attempts to include the list, a power flow calculation and a dynamic
following phenomena which may arise after a failure: simulation is done for each contingency following
(1) significant increases in transmission lines currents; the N-1 security criteria. Using the results of the
(2) voltage deviations which impact power quality; (3) dynamic simulations and the power flow calculations,
variations in frequency, maximum deviation and the post-contingency indexes are calculated . More
duration; (4) load disconnections that interrupt energy detail of these indexes is given in the next sections. To
supply to users. A failure in some component causes a estimate the severity level , an expert system operator
perturbation, influencing certain system variables must analyze the results (dynamic simulations, power
(undesirable failure consequences). These influences flows calculations, pre and post contingencies indexes)
must be analyzed to get the failure security level. The for each contingency in each scenario. From this
method employed to find the security level consists in analysis and his/her experience in the system operation,
taking the distance from the security state for the new the operator has to choose the most appropriate value.
operation state to the security margin function. Security This is an empiric evaluation and the severity level is
state is determined from indexes which somehow an empiric positive value that must measure the
summarize failure consequences, while the security operating point deviation from the secure zone. For this
margin function represents a hyper-surface over which reason, a severity level of 0 indicates a secure state and
the system is secure. this value increases as the security decreases. With the
pre-contingency indexes as inputs and the calculated
4. Contingency Selection Methodology
post-contingencies indexes as output, a neural network
The process of computing severity level involves is trained for post-contingencies index estimation in the
two steps, the first done in an off-line stage and the on-line stage . To train the severity level neural
other in an on-line stage. network , the pre and post contingency indexes are
employed as inputs and the expert severity level
4.1 Off-Line Stage
estimation is used as the output (see Fig. 2 for training
The off-line stage (Fig. 1) aims to train two groups of details). Once this neural network is trained, the
neural networks, one group for estimating post security margin function f (x) is extracted. This
contingency security state indexes and the other group function defines a hyper-surface where the system is
for determining the security margin function. Training totally safe [6].
information for neural network is taken from the This function is used in the on-line stage to estimate
typical system state (base case scenario) and several the security margin. To determine whether neural
1666 Critical Contingencies Ranking for Dynamic Security Assessment Using Neural Networks
1 2 4 N -1 D y n a m ic 7 P o s t-C o n tin g e n c y
In tita l S im u la tio n s In d e x C a lc u la tio n
9 P o s t-
G e n e ra to r
F a ilu re L is t C o n tin g e n c y P o s t-C o n tin g e n c y
Load and
(N -1 L is t) 5 In d e x E s tim a tio n E s tim a d e d In d e x e s
G e n e ra tio n N -1 P o w e r F lo w
N e u ra l N e tw o rk
S c e n a rio C a lc u la tio n s
T ra in in g
S e le c tio n
3 Pow er 6
F lo w P re C o n tin g e n c y In d e x C a lc u la tio n 10
C a lc u la tio n S e v e rity L e v e l S e c u rity M a rg in
N e u ra l N e tw o k F u n c tio n
T ra in in g
8 E x p e rt S e v e rity
L e v e l E s tim a tio n
networks have been correctly trained, a new test security state (x). To estimate the severity margin for
scenario is extracted from the base scenario. Over this each contingency, the distance between estimated
tests scenario all failures are simulated, indexes are security state and the surface given by security margin
computed from simulations and compared to those function is computed . Once each severity margin for
obtained from neural networks. each contingency is estimated, they are placed in
decreasing order to get the contingency ranking .
4.2 On-Line Stage
5. Pre-contingency State Indicators
On the on-line stage, the goal is to rank the
contingencies for the current system state. This task is Pre-contingency state indicators are indexes
performed in several steps (Fig. 3) but none of them is expressing the system state before failure that is
computationally intensive, some power flow expressed in Eqs. (1) and (2), and also the magnitude
calculations are done but no dynamic behavior is of the failure is expressed in Eqs. (3)-(5). These
simulated. From the current load and generation indexes are used as input data in the on-line stage, so,
scenario and the list with all the generators they must be quickly computable to minimize
contingencies to be ranked power flow calculations estimation delay. The pre-contingency indexes are
are done for the current system state and for each divided into three groups.
generator contingency . The power flow calculations
5.1 Power Flow Margin Index MI and Voltage Margin
are used to calculate the same pre-contingency indexes
Index MTI
used in the off-line stage . The post-contingency
indexes that were calculated in the off-line stage using This group of indexes evaluates the system state
dynamic simulations to train the post-contingency before a contingency. The MI index gives a measurable
index estimation neural network are now estimated for value of the apparent power flow (current flow)
each contingency in the list using the same trained capacity margin. MTI index measures the global
neural network . These indexes define the estimated voltage change margin.
Critical Contingencies Ranking for Dynamic Security Assessment Using Neural Networks 1667
NL
I ipre 2 NL
MI wli ( ) wl i
(1) 6.1 Power Flow Index SI
1 I ilim 1
pre lim The index is calculated by computing a weighted
NL : Number of lines; I i : Current on line I; I i : lim
Limit current on line I; wli : Line weight factor, average of differences between current limit flow I i
pos
between 0 and 1. and the one after contingency I i [10] which is
NN
Vi pre Vi n 2 NN calculated as the final current value obtained after the
MTI wni ( ) wn (2)
Vi lim
i
1 1
system dynamic simulation. A weight is assigned to
pre
NN : Number of nodes; Vi : Voltage on node I each transmission line in accordance with its
n
before contingency; Vi : Nominal voltage on node I; importance in the power system. Weight wli is the
Vi lim : Voltage deviation limit on node i (e.g.: 5% Vn); same value used in the pre-contingency Eq. (1) and in
wni : Node weight factor, between 0 and 1. the quasi-stationary indexes Eq. (3). This weight is
defined by the system operator. The equation used is:
5.2 Quasi-Stationary Power Flow Index MQI and NL
I ipos 2 NL
These indexes evaluate the system considering the I ipos : Current on line i after failure.
component outage but calculating the system state
6.2 Voltage Index STI
using only stationary tools (AC power flow). They are
defined as quasi-stationary indexes. The index MQI is Weighted average of differences between voltage
defined similarly to MI index, but considering the deviation before and after failure and the acceptable
quasi-stationary power flow results. For the index voltage limit [3] are computed. The weight value
MTQI, respecting to MTI index is valid the same corresponds to the importance assigned to each node by
consideration. Then: an expert operator as used in Eqs. (2) and (4). The
NL quasi NL
I
MQI wli ( wl (3) expression used is:
i
lim
)2 i
I NN
Vi pos Vi n 2 NN
STI wni ( wn (7)
1 i 1
)
quasi
Vi lim
i
I i : Current on line i after quasi-stationary failure. 1 1
NN
Vi quasi Vi n 2 NN
Vi pos : Voltage on node i after failure.
MTQI wni ( ) wn (4)
Vi lim
i
1 1
quasi 6.3 Frequency Deviation Indexes FI and FT
Vi : Voltage on node i after quasi-stationary
failure. Frequency deviation is a clear indicator of the
systems dynamic evolution after a contingency. Two
5.3 Disconnected Generation Index GI
indexes are calculated: maximum frequency deviation
This index evaluates the failures magnitude index Eq. (8) and total frequency deviation index Eq.
comparing the disconnected generation produced by (9).
the failure with total system generation. FI Fmax Fmax admissible (8)
GI Gout Gtotal Fp (5) ts
FT F( t )dt Fadmissible ts (9)
Gout : Power of failed generator; Gtotal : Total power 0
system generation; Fp : Weight factor. Fmax : Maximum frequency deviation;
Fadmissible : Maximum admissible frequency deviation.
6. Post-contingency State Indicators
These indicators define the estimated security state 6.4 Load Shedding Index PDI
during a contingency. These indicators are modeled by This index indicates the amount of load
four groups of indexes. disconnected after a failure. As the frequency deviation
1668 Critical Contingencies Ranking for Dynamic Security Assessment Using Neural Networks
some unique paths for energy to go from generators to CCOST500 1.764373 EMBA500 2.355602 CCOST500 2.256529
BBLA500 1.610828 CCOST500 1.58545 BBLA500 2.129098
loads. This topology implies other problems such as the
GMZA500 1.362748 ABAST500 1.496355 PILAR132 1.809512
conformation of islands when important radial lines
RAMA220 1.348878 BBLA500 1.483668 RAMA220 1.749722
fail and the difficulty in replacing generation lost by the ABAST500 1.348025 GMZA500 1.342004 OLAVA500 1.720783
remaining generators when one generator fails. Due to ATUCH220 1.345319 ATUCH220 1.277675 SNICO132 1.690918
these problems, Argentine power system is PILAR132 1.321216 RAMA220 1.261 ABAST500 1.674846
OLAVA500 1.299286 PILAR132 1.216443 ATUCH220 1.674553
ill-conditioned, exposed to collapse after the
SNICO132 1.276531 OLAVA500 1.212177 GMZA500 1.635527
occurrence of large failures. The initial failure list used RODRI500 1.266391 SNICO132 1.199055 RODRI500 1.632284
in this example is composed by outages in main
generators, and by outages of lines that join node index with a weighting factor. The severity level neural
GMZA500-RGRAN500 and YACIR500-RINCO500. network was trained with these values, the severity
margin function was extracted and the SM values
9.1 Off-Line Stage Neural Networks Training
shown in Table 1 were calculated using this function.
Three load and generation scenarios are proposed, It can be seen that for scenarios A and C, the list is
and failures for each single generator are simulated in similar up to the sixth contingency, this does not
the off-line stage. Proposed scenarios are: (1) Base happen with scenario B where the similarity with the
Scenario A. Generation: 5,650 MW. Load: 5,517 MW; other two is up to the first contingency. This behavior is
(2) Scenario B = A scenario + 500 MW load distributed explained by analyzing security state indexes: for the
among all load nodes; (3) Scenario C = A scenario RODRI500 generator contingency, the value of
500 MW load distributed among all load nodes. disconnected generation is similar in all of three
Table 1 shows a critical contingencies list for each scenarios 3.5% 3.3% and 3.9%, respectively. Index
proposed scenario. Instead of using an expert operator flow values are also similar, but voltage index values
to estimate the severity level of each contingency, the are not similar, in scenario A is 0.0776, in scenario C is
level was calculated adding each post-contingency 0.139689 and in scenario B is 0.377309, which is
1670 Critical Contingencies Ranking for Dynamic Security Assessment Using Neural Networks
higher than previous ones. Differences in simulated values of the severity indexes, these values
contingencies list were to be expected since scenarios are the reference, and the estimations must match them.
have different load levels. The third column gives the estimated indexes after
adding scenario D as training data. As it can be seen,
9.2 On-Line Stage Test
the third column list of contingencies is closer than the
An intermediate scenario between the training ones first column, due to the addition of scenario D as
was used as tests scenario. This tests scenario is training data. Comparing this table the second and third
described by: Scenario D = Base scenario + 250 MW columns of Table 3 with Table 2, significant
load distributed among all load nodes. Results from improvements are noticed in the contingency list
this scenario D are not used to train neural networks. estimation, however, errors prevail in some positions,
The results were employed only to obtain estimated mostly in those contingencies with similar severity
indexes values, and from them to get a severity level level values (e.g. for nodes ATUCH220, ABAST500
for each contingency. Once the indexes are estimated, and RAMAL220).
the same failures are simulated and indexes and
10. Conclusions
severity margin for each contingency are calculated.
A comparison between contingencies lists obtained The new competitive environment in that power
by estimation and by simulation is shown in Table 2. It systems operates has created a need of an on-line
can be seen that the first four contingencies are the dynamic security assessment, due to the necessity of
same, while the rest does not, the error in the predicting future operation conditions. In order to carry
contingencies list estimation is due to an existing error out a complete on-line assessment, it is necessary to
in security state indexes estimation. evaluate the power system dynamic behavior when
contingencies occur for a given operating state. This is
9.3 Final Test and On-Line Stage a time-consuming task and most contingencies
The results obtained in the estimated contingencies simulated do not result in unacceptable values of state
have a low satisfaction degree respect the simulated variables, mainly due to their small influence on
contingencies, such as shown in Table 2. For this system operation. Therefore, previous selection of
reason, a test scenario is introduced as additional data critical contingencies becomes necessary, specially to
training for the neural networks, to improve the Table 2 Simulated and estimated CC for scenario D.
estimation indexes. To verify whether improvement Scenario D (simulated) Scenario D (estimated)
exists when adding scenario D as training data, another Generator SM Generator SM
tests scenario (Scenario E) was derived from the base YACIR500 44.317445 YACIR500 44.480546
CHOCO500 3.356028 CHOCO500 3.708806
scenario. This scenario consists in adding to the base
EMBA500 2.244625 EMBA500 2.282154
scenario a 150 MW load uniformly distributed. The CCOST500 1.662381 CCOST500 2.013404
indexes and SM index were estimated for each failure BBLA500 1.536574 RODRI500 1.932605
in this scenario, and all failures were dynamically GMZA500 1.351459 SNICO132 1.928779
simulated and indexes are calculated for each ABAST500 1.300225 RAMA220 1.636928
ATUCH220 1.296962 BBLA500 1.583393
contingency.
RAMA220 1.289881 OLAVA500 1.491689
Table 3 shows a comparison between contingencies PILAR132 1.258326 ATUCH220 1.470355
lists obtained by estimation and by simulation. In the OLAVA500 1.246194 ABAST500 1.455058
first column are the estimated indexes before adding D SNICO132 1.226863 GMZA500 1.427926
RODRI500 1.223033 PILAR132 1.340695
scenario as training data. In the second column are the
Critical Contingencies Ranking for Dynamic Security Assessment Using Neural Networks 1671
get an on-line dynamic security assessment methodology, in: Proceedings of the 37th Annual North
American Power Symposium, Ames, Iowa, Oct. 23-25,
implementation. This paper presents a new approach
2005, pp. 67-73.
useful for selection and ranking of critical [3] P. Kundur, G.K. Morison, L. Wang, H. Hamadanizadeh,
contingencies. The process of selection involves On-line dynamic security assessment of power systems, in:
several steps, some of them have done in an off-line 5th International Workshop on Electric Power Control
Centres, Hviz, Hungary, June 13-16, 1999.
stage and some in an on-line stage. Five
[4] J.M.G. Alvarez, P.E. Mercado, On-line inference of the
pre-contingency state indicators and five dynamic security level of power systems using fuzzy
pos-contingency state indicators are defined. It is true techniques, IEEE Transactions on Power Systems 22 (2)
that proposed indexes do not cover all the phenomena (2007) 717-726.
[5] Y. Cheng, A. Bose, Direct ranking for voltage contingency
in the power system dynamics, buy they do not pretend
selection, IEEE Transactions on Power Systems 4 (4)
to be a security system measurement. However, they (1989) 1335-1344.
are enough for contingencies selection because they [6] J.D.M. Calley, B.A. Krause, Rapid transmission capacity
provide information about the worst dynamic effects of margin determination for dynamic security assessment
using artificial neural network, Electric Power Systems
systems failures. On the off-line stage, the pre- and
Research 34 (1) (1995) 37-45.
post-contingency indexes are calculated using dynamic [7] M. Kezunovic, I. Rikalo, D.J. Sobajic, Real-time and
simulations. A neural network is used for off-line transmission line fault classification using neural
pos-contingency index estimation on the on-line stage. networks, Intl. Journal of Engineering Intelligent Systems
4 (1) (1996) 57.
The method presented here for contingency ranking is
[8] R. Fischl, M. Kam, J.C. Chow, H.H. Yan, On the design of
an extension to similar methods used for contingency neural networks for detecting the limiting contingencies in
selection where no dynamic effects are considered. The power system operation, in: Proceedings of the 10th
extension proved to be effective in the example case Power Systems Computation Conference, Gratz, Austria,
Aug. 1990, pp. 887-894.
based only on the proposed indexes.
[9] M.A. Albuquerque, C.A. Castro, A contingency ranking
References method for voltage stability in real time operation of
power systems, in: IEEE Bologna PowerTech Conference,
[1] K. Morison, L. Wang, P. Kundur, Power system security Bologna, Italy, June 2003.
assessment, IEEE Power and Energy Magazine 2 (5) [10] T. Jain, L. Srivastava, S.N. Singh, Fast voltage
(2004) 30-39. contingency screening using radial basis function neural
[2] G.K. Stefopoulos, F. Yang, G.J. Cokkinides, A.P.S. network, IEEE Transactions on Power Systems 18 (4)
Meliopoulos, Advanced contingency selection (2003) 1359-1366.
1672 Critical Contingencies Ranking for Dynamic Security Assessment Using Neural Networks
[11] S. Ghosh, B.H. Chowdhury, Design of an artificial enhancement using a coupled ANN-ES scheme, in:
neural network for fast line flow contingency ranking, Procceding of the 4th Symp. on Expert System
Electrical Power and Energy System 18 (5) (1996) Application to Power System, Melbourne, Australia, Jan.
271-277. 4-8, 1993.
[12] Y.Y. Hsu, H.C. Kuo, Fuzzy set based contingency ranking, [14] P. Kundur, J. Paserba, V. Ajjarapu, G. Andersson, A. Bose,
IEEE Transactions on Power Systems 7 (3) (1992) C. Canizares, et al., Definition and classification of power
1189-1196. system stability, IEEE Transaction on Power Systems 19
[13] V.S. Vankayala, N.D. Rao, Power system security (2004) 1387-1401.