You are on page 1of 6

2010 IEEE International Conference on Power and Energy (PECon2010), Nov 29 - Dec 1, 2010, Kuala Lumpur, Malaysia

Analysis of Transfer Capability by Markov Chain


Monte Carlo Simulation
Magnus Perninge Lennart Söder
School of Electrical Engineering School of Electrical Engineering
Royal Institute of Technology Royal Institute of Technology
Stockholm, Sweden Stockholm, Sweden
Email: magnus.perninge@ee.kth.se

Abstract—Adequate security margins are normally applied In [4] simulations to determine the NTC over one of the
in power systems by keeping a pre-defined set of transfer critical Sections in the Swedish power system was performed.
limits through certain transmission corridors in the system. This Here, voltage stability was assumed to be the critical security
will protect the transmission lines from overheating causing a
violation of the thermal stability limits of the system. However, constraint, and a defining contingency was considered. It was
for other security limits such as, voltage stability, and lower shown that one very important factor was the distribution
voltage limits at specific nodes, the distribution of the injected of the load amongst the nodes on the consumer side of the
power amongst the nodes of the system will also be of importance. considered section. To fully be able to evaluate the risk of
In this article we propose a method for generating samples of voltage instability for a given allowed transfer we will thus
the injected power at all nodes given a set of transfers through
specified corridors of the power system. It is then showed how the have to consider the stochastic placement of the injected power
method can be used to evaluate the risk of violating the system amongst the nodes in the system and not only the transfers
stability limits induced by keeping a specific set of transfer limits. through the critical sections.
The method can be used in power system short-term planning
when setting the limits for trading and transfer between the In this article we propose a method for evaluating the
different nodes of the power system.
Index Terms—Power system security, active power transfer
probability of violating power system security constraints
limit, Markov chain Monte Carlo, voltage instability, stochastic given a set of transfers through a given set of critical sections
nodal loading. in the system. Here, the losses are approximated by the second
order Taylor expansion of the loss function in injected power
I. I NTRODUCTION around the maximum likelihood of the constrained injected
powers. The method assumes that the transfers over some
Due to the intensive use of transmission networks, power sections in the system are known and that the probability
system security [1] has become one of the most important distribution of the injected powers is known. One possible
issues in power system operation. There is a large conflict implementation of the method is if the System Operator (SO)
of interest between the market perspective, where a large of a power system wants to know what risk the system will
capacity to transfer power through the electric power grid is run of e.g. voltage instability if the transfer through a critical
required, and the security perspective, where secure opera- transmission corridor is increased to a certain level. One
tion is the main objective. To satisfy both objectives to the simplification assumed in this paper is that the fact that the
largest possible extent an adequate balance between security load is a stochastic process and not a single random variable
and capacity is preferable. To maintain secure operation, the is neglected. This will make computation easier.
flows over certain sections in the system is measured and
actions are taken when these flows exceed a predefined limit, The presentation is structured as follows. In Section II a
called Net Transfer Capacity (NTC) [2]. When determining mathematical description of the problem is given. In Sec-
an efficient NTC one has to consider fluctuations in the Total tion III we propose a choice for the point around which the
Transfer Capacity (TTC) as the system conditions vary, and Taylor expansion of the losses is to be made. Section IV is
uncertainties in system parameters. devoted to derivation of the probability distribution of the
In [3] a confidence bound for the Transmission Reliability injected power given that the transfers in the system should
Margin (TRM), a margin that is kept to be able to maintain be the set values and assuming the second order expansion
secure operation under system parameter random fluctuations of the losses. Then, in Section V, we show how to, using
and other parameter uncertainties, is calculated. This paper Markov Chain Monte Carlo (MCMC), generate outcomes
assumes that the parameters can be chosen in a way that makes from the probability distribution found in Section IV. In
them independent. Then sensitivity analysis, combined with Section VI a numerical example where the proposed method
the central limit theorem, is used to obtain a confidence bound is implemented in the IEEE 9-bus system with one critical
for the TRM. section is given.

978-1-4244-8946-6/10/$26.00 ©2010 IEEE 232


Authorized licensed use limited to: Universitas Brawijaya. Downloaded on November 23,2021 at 11:13:07 UTC from IEEE Xplore. Restrictions apply.
II. P ROBLEM FORMULATION where f is a vector of power mismatch, and λ is a vector in
injected power space, we can extract
Consider the simple power system depicted in Fig. 1.
Assume that the joint probability distribution of x = u(λ). (6)
Hence, we get the losses W (x) = W (u(λ)) as a function
3 of injected power λ. Unfortunately, u cannot easily be found,
Gen 1 so an explicit expression for the active power losses, PL , is
hard to find. However, it has been shown [5] that a second
1 order Taylor expansion of the losses is in many cases a very
close approximation of the actual losses. Write λ = λ0 + ∆λ.
A second order Taylor series expansion of PL around λ0 is
given by
2 PLSO (λ) = W + Wλ ∆λ + 21 ∆λT Wλλ ∆λ, (7)
P23
where W and its derivatives are to be evaluated at λ0 . The
4 5 vector Wλ , and the matrix Wλλ are given by
W λ = W x uλ , (8)
3 and
Wλλ = Wx uλλ + uTλ Wxx uλ (9)
To find the matrix uλ and the tensor uλλ , we differentiate (5)
Load B Load A first once to get
Fig. 1. The one-line diagram of the simple system used to explain the fx uλ + fλ = 0, (10)
problem which the article aims at solving.
which implies that
  uλ = −fx−1 fλ (11)
PD1
PD = (1) and then again to get
PD2
is known, where PDi is the active power demand in Load i. fx uλλ + uTλ fxx uλ + fλλ = 0, (12)
Assume that the transfer P12 is monitored and we wish to from which uλλ can be deduced.
know the probability distribution of
Two questions now arise:
PD | P12 = K. (2)
1) What is a good choice of λ0 ?
If losses are neglected this problem is reduced to PD | PD1 + 2) What is the resulting probability distribution of PD
K
=
SO
PD2 = K. However, when trying to keep adequate transfer PD | PD1 + PD2 + PL (PD ) = K?
limits, losses can be vital. Hence, we are searching for
III. A GOOD CHOICE FOR λ0
2
Expanding the loss function around the most probable
X
PD | PDi + PL (PD ) = K, (3)
i=1
operating point seem reasonable. Hence, if we denote by f PD
the probability density function of PD , then the solution to
where PL : R2 → [0, ∞) is the active power losses in the
system. max fPD (λ) (13)
λ∈Rn
The active power losses, south of node 2, can be written as s.t. f (x, λ) = 0, (14)
the sum of injected power, n
X
5 5 λi + W (x) = K, (15)
X X
PL = Vi Vk (Gik cos(δik ) + Bik sin(δik )). (4) i=1

i=2 k=2 will be a good choice for λ0 . Here f is the set of power
flow equations where the known parameters such as e.g.
Now, assume x = (V1 , . . . , VN , δ1 , . . . , δN )T . Hence PL
voltages at PV-buses are set to their known values, and
can be written PL = W (x), where at normal operation n
λi + W (x) = K assures that the transfer through the
P
W ∈ C ∞ (RN × [0, 2π)N → [0, ∞)). Now, through the load
i=1
flow equations, corridors have the correct value.
f (x, λ) = 0, (5)

233
Authorized licensed use limited to: Universitas Brawijaya. Downloaded on November 23,2021 at 11:13:07 UTC from IEEE Xplore. Restrictions apply.
The second order Taylor expansion PLSO of PL around λ0 Hence, if the (n − 1)-dimensional random vector Z is given
is then given by the distribution
PLSO (y) = W + Wλ (y − λ0 ) + 21 (y − λ0 )T Wλλ (y − λ0 ), (16) fZ (z) =
fPD (LK (z))
=
π(z)
, (29)
C C
where all instances of W and its derivatives is to be evaluated
where the normalization constant C is chosen such that
at λ0 . Z
IV. T HE DISTRIBUTION OF K
PD fZ (z)dz = 1, (30)
Rn−1
The random variable PD
K
is living on the (n − 1)-Manifold
d d
n then LK (Z) = PD
K
, where = denotes equality in distribution.
λ(i) + PLSO (λ) = K}.
X
ΛK = {λ ∈ Rn : (17)
i=1
V. S IMULATION
A point y on the surface ΛK will thus fulfill To estimate the system security we want to compute the
X probability of violation of the system constraints,
y (i) + W + Wλ (y − λ0 ) Z
+ 21 (y − λ0 )T Wλλ (y − λ0 ) = K. (18) pF = P(1E ) = fPDK (y)dy, (31)
E
Since where P is the probability measure of the random variable
(i)
X
λ0 + W = K, (19) PDK
, and E is the set of non-allowed operating points. One
this implies that of the obstacles faced when trying to calculate pF is the
unknown constant C. Finding the normalization constant C
A(y − λ0 ) + 21 (y − λ0 )T Wλλ (y − λ0 ) = 0 (20) by computing the integral
where
Z
(i) π(z)dz (32)
A(i) = Wλ + 1. (21) Rn−1

From (20) we see that the the hyperplane passing through λ0 will in general be complicated and a Monte Carlo method
with normal has to be used. However, for most cases the only reason
nF = A/kAk (22) for seeking the distribution of PD K
is to be able to generate
n outcomes of the injected power for a given set of transfers. If
is where PD | PDi + PLFO (PD ) = K lives, where PLFO is we chose to generate these outcomes using a Markov chain
P
i=1
a first order Taylor expansion of the active power losses. We Monte Carlo method [6] the computation of C can be omitted.
also note that nF is the normal to ΛK at λ0 . Let {ai }i=1
n−1
be In this section we will present the Markov chain Monte Carlo
orthonormal vectors in R such that nF ⊥span{a1 , . . . , an−1 }.
n method, and show how this method can be used to calculate
Hence, every vector y ∈ Rn can be written pF .

y = λ0 + Bw, (23) A. Markov chain Monte Carlo and the Metropolis-Hastings


algorithm
where
B = [nF a1 ··· an−1 ] , (24) Markov chain Monte Carlo (MCMC) is a method for
generating samples from random variables having complicated
for some w ∈ Rn . Hence, probability distributions such as PD K
. The samples Yi gen-
erated using MCMC forms an ergodic Markov chain whose
0 = ABw + wDw = kAkw (1) + wDw, (25)
asymptotic distribution is the distribution from which we try
where to generate samples. The value E[h(Y∞ )] is then estimated
D = 21 B T Wλλ B. (26) using the fact that

For every {w , . . . , w }, Equation (25) is a second degree


(2) (n)
1 X
N

polynomial in w (1) . Hence, w (1) can be written as a function E[h(Y∞ )] = lim h(Yi ). (33)
N →∞ N
of {w(2) , . . . , w(n) }, say i=1

The most common way of generating the samples Yi is by


w(1) = g(w(2) , . . . , w(n) ). (27)
the Metropolis-Hastings algorithm suggested by Metropolis et
Here, due to physical reasons, g should be taken to be the so- al. [7] and later extended by Hastings [8]. Given that Yi−1 = x
lution which is closest to the first order approximation, i.e. the we generate a proposal y for Yi from a distribution q(x, ·).
solution for which w (1) is closest to zero. A parametrization When a proposal y has been generated from q(x, ·) it is
of ΛK is given by LK : Rn−1 → ΛK defined by accepted with probability

LK (z) = λ0 + g(z)nF + [a1 ··· an−1 ] z (28) α(x, y) = min(1, r(x, y)), (34)

234
Authorized licensed use limited to: Universitas Brawijaya. Downloaded on November 23,2021 at 11:13:07 UTC from IEEE Xplore. Restrictions apply.
Gen 1
where,
π(y)q(y, x)
r(x, y) = . (35)
π(x)q(x, y)
If y is accepted then Yi = y, and if y is not accepted then 1
Yi = x. This means that unlike ordinary acceptance-rejection
(A-R) algorithms a sample is generated in each step. The
crucial feature of (35) is that we have a ratio between two -P41 4 Section A
π-values, so that the normalization constant C need not be
computed. The distribution q(x, ·) should be chosen such that
Yi can take all values for which π has a density. One way 6 5
of making this hold is to let q(x, y) be a the distribution
function of a N (x, σ)-distributed random variable. The choice Load B Load A
of q also affect the convergence rate of the distribution of Yi
to its asymptotic distribution. It can be shown that using the
Metropolis-Hastings algorithm to generate the Yi s, yields Yi
to be an ergodic Markov series with asymptotic distribution
π/C. Gen 3 Gen 2

B. Estimating pF
One way of estimating pF is as follows. We first solve
3 9 8 7 2
min
n−1
kzk (36) Load C
z∈R
s.t. LK (z) ∈ E, (37) Fig. 2. The one-line sketch of the power system used in the numerical
example. Here, node 1 is the slack node, and Section A is the section over
which can be solved using e.g. Particle Swarm Optimization which the transfer is measured to be K p.u. Hence, −P41 = K p.u.
(PSO) [9]. An example of where PSO is used to solve a
problem of this type is in [10] where a modified PSO-
algorithm is used to find the distance, in parameter space, from and
an operating point to the saddle-node bifurcation surface. Let  
1 26 9
R0 > 0 be the optimum to this problem. Now we start at some C= , (39)
20 9 17
Y0 and generate the sequence {Yi }. If kYi k < R0 , we know
that LK (Yi ) ∈
/ E and no power flow calculations will have
and PD3 = 1 p.u. The reason that the injected power in
to be performed. Since generating a new sample, Yi , given
one node is chosen to be deterministic is to increase the
Yi−1 , is much faster then evaluating whether LK (Yi ) ∈ E
pedagogical value by allowing a more visual presentation.
this procedure will save a lot of computational effort.
In Fig. 3 the bifurcation boundary Σ and the curve where
VI. N UMERICAL EXAMPLE PD | − P41 = K takes its values is plotted (both lying in the
In this section a numerical example will be given. The PD1 PD2 -plane) together with the probability density function
aim of the numerical example is to show how the method fPD |−P41 =K .
described in this paper can be applied by looking at a simple
system. The system that we have chosen for the numerical
example is the IEEE 9-bus system depicted in Fig. 2. This
system has 9 nodes where nod 1 is a slack node, node 2 and
node 3 are generator nodes, and node 4-9 are PQ-nodes. The fPD |−P41 =K
generators in nodes 2 and 3 have reactive power consumption
limits Q̄G2 = 1.4 p.u., and Q̄G3 = 1 p.u. respectively. The 5

consumption is located in nodes 5, 6, and 8. Hence, if we 4


assume that the reactive power is a deterministic function of m Σ
the active power, the load-space will be three-dimensional, 3

PD2
i.e. n = 3. 2

0
0.5 1
1
The aim of the numerical example is to estimate the 1.5
2
2.5
3
3.5 0
probability of voltage instability when the transfer through PD1 4
4.5

Section A is K = 4 p.u., (PD1 , PD2 ) is N (m, C)-distributed,


Fig. 3. The region in injected power space where −P41 = K, together with
where   the density fPD |−P41 =K and the saddle-node bifurcation surface Σ.
2
m= , (38)
2

235
Authorized licensed use limited to: Universitas Brawijaya. Downloaded on November 23,2021 at 11:13:07 UTC from IEEE Xplore. Restrictions apply.
3

A. Parametrizing ΛK
To find λ0 we chose the point where fPD |−P41 =K obtains 2

its maximal value. This gives us λ0 = [2.7208 2.5370]T . 1

The objective is thus to estimate the surface under the curve


fPD |−P41 =K in the allowed operating region. One minus this 0

number will then be used as an estimate of the probability


of voltage instability given the transfer −P41 = K. The first −1

order expansion of PL in λ around λ0 is given by, −2

Wλ = [0.1321 0.1601] . (40) −3


0 1000 2000 3000 4000 5000 6000 7000 8000 9000 1000

The normal to ΛK at λ0 is thus given by the fact for ∆λ ∈ i

T λ0 Λ K , X Fig. 4. The sequence {Yi }10000


i=0 generated using σ = 0.1.
∆λ(i) + Wλ ∆λ = 0. (41)
3

Hence,  
0.6984 2

nF = . (42)
0.7157
1

The matrix Wλλ in the second order expansion of PL at λ0


is  
0

0.1680 0.1021
Wλλ = . (43) −1
0.1021 0.1983
Now, let w = [u v]T be such that w + λ0 ∈ ΛK , then −2

   
u u −3

+ 12 [u v]Wλλ (44)
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 1000
([1 1] + Wλ ) = 0. i
v v
The vector Fig. 5. The sequence {Yi }10000
i=0 generated using σ = 1.
 
0.7157
a1 = . (45) 4
−0.6984
3

together with nF forms a orthonormal basis for R2 . If [s t]


is the [a1 nF ]-coordinates of w, then, letting 2

(46)
1
R = [a1 nF ] ,
0
we get
  −1
s
k[1 1] + Wλ kt + 12 [s t]RT Wλλ R = 0. (47)
t −2

From (47) we find hw, nF i given hw, a1 i for every point w ∈ −3


0 1000 2000 3000 4000 5000 6000 7000 8000 9000 1000
i
ΛK . The relation t = g(s) given by (47) is then used in the
map LK : R → ΛK as in (28). Fig. 6. The sequence {Yi }10000
i=0 generated using σ = 10.

B. Simulation
To start of the Markov chain in our MCMC simulation From these figures it is clear that using a too small σ will
we choose Y0 = λ0 . We then generate a proposal V for Y1 make the chain move slowly, hence the convergence is slower,
from a N (Y0 , σ)-distribution. By the symmetry of the normal and using a too large σ will result in many proposals being
distribution we have that q(x, y) = q(y, x) in this case. Hence, rejected which is not good either. Using MCMC with σ = 1
our r is simply, we got the estimate pF = 0.02. Hence, if the system operator
π(V ) fPD (LK (V )) lets the transfer through Section A reach 4 p.u. the probability
r(Y0 , V ) = = . (48) of system failure is estimated to be around 2%.
π(Y0 ) fPD (LK (Y0 ))
To decide whether to accept V for Y1 we generate an outcome VII. C ONCLUSION
of a random variable U that is uniformly distributed on [0, 1] In this paper a method for analyzing the probability of
an accept V if r(Y0 , V ) ≥ U . This procedure is then repeated violating power system security constraints given a transfer
to generate a Y2 from Y1 and so on. In Fig. 4, Fig. 5, and Fig. 6, limit is proposed. As in-signals to the proposed algorithm is
outcomes of {Yi }10000
i=0 generated using σ = 0.1, σ = 1 and the transfer into the area and probability distribution of the
σ = 10 respectively are plotted. unconstrained nodal loading is assumed. The losses are then

236
Authorized licensed use limited to: Universitas Brawijaya. Downloaded on November 23,2021 at 11:13:07 UTC from IEEE Xplore. Restrictions apply.
approximated by the second order Taylor expansion of the loss
function in injected power around the maximum likelihood
of the constrained nodal loading. To estimate the probability
of violation of the security constraints, Markov Chain Monte
Carlo Simulation is used to generate outcomes of the injected
power given the set of transfers in a Monte Carlo Simulation.
The method can be used in off-line power system security
assessment.
R EFERENCES
[1] IEEE/CIGRE Joint Task Force on Stability Terms and Definitions. Def-
inition and Classification of Power System Stability. IEEE Transactions
on Powr Systems, 19(3), 2004.
[2] Nordel. Principles for determining the transfer capacity in the Nordic
power market. Technical report, Nordel, Jan 2008.
[3] J. Zhang and I. Dobson and F.L. Alvarado. Quantifying transmission
reliability margin. Electrical Power and Energy Systems, 26, 2004.
[4] M. Rostami. Analys av späningsstabilitetsberäkningar i det svenska
elsystemet. Master’s thesis, Royal Institute of Technology (KTH), may
2009.
[5] L. Söder. Estimation of reduced electrical distribution losses depending
on dispersed small scale energy production. In Proceedings of 12th
Power Systems Computation Conference, 1996.
[6] S. Asmussen and P. W. Glynn. Stochastic Simulation. Springer, 2007.
[7] N. Metropolis, A. Rosenbluth, M.Rosenbluth, A. Teller and E. Teller.
Equations of state calculations by fast computing machines. J. Chem.
Phys., 21:1087–1091, 1953.
[8] W.K. Hastings. Monte Carlo sampling methods using Markov chains.
Biometrica, 89:731–743, 1970.
[9] R. C. Eberhart and J Kennedy. A new optimizer using particle swarm
theory. In Sixth International Symposium on Micromachine and Human
Science Nagoya, Japan, pages 39–43, 1995.
[10] B. Wu M. J. Chen and C. Chen. Determination of shortest distance to
voltage instability with particle swarm optimization algorithm. Euro.
Trans. Electr. Power, 2008.

237
Authorized licensed use limited to: Universitas Brawijaya. Downloaded on November 23,2021 at 11:13:07 UTC from IEEE Xplore. Restrictions apply.

You might also like