Sei sulla pagina 1di 8

Optimal Dispatch of Reactive Power Sources

by Using MVMOS Optimization


Jos L. Rueda, Senior Member, IEEE Istvn Erlich, Senior Member, IEEE
Institute of Electrical Power Systems Institute of Electrical Power Systems
University Duisburg-Essen University Duisburg-Essen
Duisburg, Germany Duisburg, Germany
jose.rueda@uni-due.de istvan.erlich@uni-due.de

Abstract In this paper, the swarm implementation of the Mean- so far and saved in a continually-updating archive. One
Variance Mapping optimization (MVMOS) is proposed to solve remarkable trait of the classical implementation of MVMO
the optimal reactive power dispatch problem. Apart from constitutes a single-particle approach in which the tradeoff
incorporating swarm intelligence principles, MVMOS possesses between search diversification and intensification results in fast
enhanced mapping and penalty schemes as compared to the progress rates with reduced risk of premature convergence.
classical MVMO procedure. Based on the IEEE 57- and 118- bus
systems, numerical tests and comparisons with other heuristic In this paper, the swarm implementation of MVMO, which
optimization methods are carried out. Additionally, a scheme for has been termed as MVMOS, is proposed for the solution of the
adaptive smart optimal reactive power source coordination, ORPD problem. This new approach extends the innate power
which can be used online, is suggested. of global searching of the original MVMO by starting the
search with a set of particles (i.e. swarm), each having its own
Keywords Heuristic optimization; mean-variance mapping memory and represented by the corresponding archive and
optimization; reactive power dispatch; smart grid; swarm mapping function, as well as allowing information exchange
intelligence. and dynamic reduction of the swarm size through simple rules.
I. INTRODUCTION Enhanced schemes for shape factor assignment and dynamic
penalty are also incorporated.
The optimal reactive power dispatch (ORPD) problem has
the goal of scheduling the available reactive power sources in In the following sections details of the algorithm as well as
such a way that the overall transmission losses are minimized some application examples will be provided. Following this
by satisfying a set of operational constraints [1]. ORPD has introductions, Section II presents the mathematical formulation
received a renewed interest in recent years, since optimal of the ORPD problem. In Section III, MVMO basics are briefly
integration and intelligent operation of an increasingly diverse reviewed whereas MVMOS is explained in Section IV.
mix of primary sources into the power supply system is of great Numerical results are given and discussed in Section V,
concern for a well-functioning competitive and including a performance comparison between MVMOS,
internationalized electricity market built around smart grids [2]. MVMO, and other heuristic optimization methods, and an
outline to the coupling of MVMOS with adaptive reactive
In addition to its complex underlying mathematical power source coordination as well. Finally, Section VI
framework, the ORPD problem constitutes a large scale summarizes the concluding remarks.
optimization task involving a mixture of continuous and
discrete control variables. Hence, modern heuristic II. PROBLEM STATEMENT
optimization techniques, such as genetic algorithms [3], Solving the ORPD entails determining the optimal settings
particle swarm optimization (PSO) [4], and differential of reactive power control variables (i.e. generator bus voltages,
evolution (DE) [5] are widely used for its solution. Although setting of reactive power sources/sinks and transformer tap
these techniques have some advantages over classical positions, etc.) leading to the minimum power losses while
optimization methods, their performance may entail risk of fulfilling the operational constraints. Mathematically, the
premature convergence or local stagnation when handling problem can be formulated as follows [3], [4]:
discontinuous multimodal (i.e. multiple local optima) non-
convex landscapes [6]. Minimize
Mean-variance mapping optimization (MVMO) is a recent
addition to the emerging optimization algorithms with some Ploss = g (V
k N K
k i
2
i j cos ij
+ V j2 2VV ) (1)
basic conceptual similarities to other heuristic approaches.
Several potential areas of application in research on power subject to
system optimization tasks have been reported [6]-[12]. Its
working principle is based on a special mapping function p ( v, ) p g + pd = 0 (2)
applied for mutating the offspring on the basis of mean and
variance of the set comprising of the n-best solutions attained

This work was fully supported by the University Duisburg-Essen.

978-1-4673-6002-9/13/$31.00 2013
c IEEE 29
q ( v, ) q g + q d = 0 (3) - A novel mapping function that is used for mutating genes in
the offspring based on the mean and variance of the
solution archive
v min v v max (4)
- A compact and dynamically updated solution archive that
q gmin q g q gmax (5) serves as the knowledge base for guiding the search
direction (i.e. adaptive memory). The n-best individuals that
q cmin qc qcmax (6) MVMO has found so far are saved in the archive and sorted
in a descending order of fitness.
t min t t max (7)

s s max (8)

where Ploss denotes the total active power losses of the


transmission network, gk and ij are, respectively, the line
conductance and the difference between the voltage angle
between buses i and j, and Nk is the total number of network
branches. pg and qg are the nodal active and reactive power
generation vectors whereas pd and qd are the nodal active and
reactive power demand vectors. v and stand for bus voltage
magnitude and phase angle vectors, respectively. qc and t are
the vectors of compensator reactive power output and
transformer tap positions. vmin, vmax, qgmin, qgmax, qcmin, qcmax,
tmin, and tmax denote the lower and upper bounds of bus
voltage, reactive power output of generators, compensator
reactive power and transformer tap positions. Finally, s and
smax denote actual and maximum allowed apparent power flow
vectors in the branches, respectively. p(.) and q(.) denote nodal
active and reactive power injection vectors, whose elements are
calculated using (9) and (10), respectively:

Pi ( v, ) = Vi V j Gij cosij + Bij sinij (9)


jN

Qi ( v, ) = Vi V j Gij sinij B cos ij (10)


jN

where Gij and Bij denote conductance and susceptance the link
between buses i and j, respectively. N is the total number of
buses.
III. REVISITING CLASSICAL MVMO
The flowchart of the implementation of the single-particle
MVMO is schematically sketched in Fig. 1. The procedure
begins with an initialization phase where the algorithms
parameter settings are defined and random samples for control
variables from the space of possible solutions are generated as
well. Next, an iterative loop is initiated, in which fitness
evaluation (i.e. objective function plus penalty as a function of
the degree of constraint fulfillment) is performed, the
termination criterion is checked, the solution archive (i.e.
inclusion or exclusion of candidate solutions) is updated, the
global best solution is determined (i.e. parent assignment), and Figure 1. Classical MVMO single-particle approach.
new candidate solutions are created (i.e. mutation by projection
of selected variables onto the mapping function and crossover). - A single parent-offspring pair concept is adopted. In
The salient features of the algorithm can be summarized as contrast to other methods, where the term iteration
follows: generally refers to the number of fitness evaluations (which
is proportional to the total number of individuals in the
swarm) MVMO requires only one fitness evaluation per

30 2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG)


iteration irrespective of the number of individuals saved in
the solution archive.
1 1
xi Parameter: xmean
- The restriction of the range of the search space for all x Parameter 1.0 s1 =s2= 10
optimization variables internally to [0, 1]. This is a 0.8 0.8 0.75
shape s1 = s2
precondition for using the mapping function; on top of this, 0.6 0.6
it guarantees that the generated offspring is always within 0.5
the search boundaries. However, fitness evaluation is 0.4 0 0.4
carried out in the original physical dimension. 5 0.25
10 0.2
0.2 15
Interested readers can find the basic theoretical background, 0
50
and reference values for the algorithms settings in [6]-[8] and 0 0
0 0.5 x* 1 0 0.5 x* 1
[13]. Before introducing the swarm approach, this subsection is
intended to highlight further modifications of the single- 1 1
particle implementation aiming at improved mapping x x
0.8 0.8 asymmetrical
performance and constraint handling. symmetrical s1 = 15, s2 = 5
s1 = s2 = 15
A. Enhanced mapping 0.6 0.6

The individual with the best fitness so far in the archive 0.4 0.4
(first position) is used in every iteration to generate a new asymmetrical symmetrical
descendant (i.e. assigned parent). Basically, m out of k 0.2 s1 = 5, s2 = 15 0.2 s1 = s2 = 15
dimensions of the optimization problem are strategically 0 0
selected for mutation operation via mapping function while the 0 0.5 x* 1 0 0.5 x* 1
remaining dimensions inherit the corresponding values from
Figure 2. Change of the mapping function shape under different values of
the parent. Alternative selection methods are described in [6] mean and shape factors.
and [13].
The new value of each selected dimension xi is determined Recalling (14), it is evident that the factor fs can be used to
by change the shape of the function. A small value (e.g. between
0.5 and 1.0) allows the slope of the mapping curve to increase
xi = hx + (1 h1 + h0 ) xi* h0 (11) and thus enable better exploration, whereas values above 1.0
will result in a rather flat curve and thus lead to improved
where x*i is a variable varied randomly with uniform exploitation. In general, it is recommended to start the search
distribution and the term h refers to the transformation process with a smaller fs and then increase it as the
mapping function, which is defined as optimization progresses. In several applications the random
variation of fs according to (15) has result in significant
h( x, s1 , s2 , x) = x (1 e xs1 ) + (1 xi ) e (1 x ) s2 (12)
improvements.
hx, h1 and h0 are the outputs of the mapping function, based on
different inputs given by f s = fs* (1 + rand() ) (15)
*
hx = h( x = x ), h0 = h( x = 0), h1 = h( x = 1) (13)
i
Where fs* denotes the smallest value of fs and rand() is a
si is the shape variable and is calculated as follows random number in the range [0, 1]. When the accuracy of the
si = ln(vi ) fs (14) optimization needs to be improved, the following extension
can also be added in order to allow a progressive increase of
At the start, the mean xi corresponds with the initialized fs* :
value of xi and variance vi (associated to si ) and is set to one. 2
But as the optimization progresses, they are recalculated after
every update of the archive for each selected optimization
fs* = f s*_ ini
i
+
ifinal
( fs*_ final fs*_ ini ) (16)

variable. Both input and output of the mapping function cover


The variable i represents the iteration number. If
the range [0, 1]. Note that the shape of the mapping function is
f s*_ final = f s*_ ini , the factor fs is fixed throughout the
influenced by x and the shape factors s1 and s2 . The effect of
these parameters on the form of the function is illustrated in optimization. It is recommended to set f s*_ ini = 0.9!1.0 and
Fig.2, where it is easily deduced that the search diversity can
be enhanced through proper assignment of the shape variables. f s*_ final = 1.0! 3.0 .
Note also that, with increasing value of the shape parameter,
In contrast to previous publications, the shapes factors s1
the mapping curve becomes more flat so that the space to be
searched focuses on the region near the mean value. and s2 of the variable xi are not calculated directly from (14)
but by using the following procedure:

2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG) 31


si1=si2=0 which corresponds with a straight line between zero
si1 = si2 = si
and one as the mapping function. The
T mean value in this case
if si > 0 then does not have any effect on the mapping function. Fig. 3
d = (1 + d0 ) + 2 d0 ( rand() 0.5) described procedure on the
illustrates the effect of the above-d
mapping shape.
if si > di
d i = d i d
else
di = di /d
(17)
end if
if rand() 0.5 then
si1 = si ; si2 = di
else
si1 = di ; si2 = si
end if
end if
The initial values of d i are set for all variables at the
beginning of the optimization. Experience soo far shows that
values around 1 - 5 guarantee good initial performance. At Figure 3. Variation of the mapping functio
on shape with the proposed shape
every iteration, each d i is scaled up or downn with the factor factor assignment procedure.

d . If di > si, the current di is divided by d which is always B. Constraint handling


larger than 1.0 and thus leads to reduced vallue of di. In case Every potential solution is evalluated in light of its fitness
di < si, di will be multiplied by d resultingg in increased di. measure. This involves power flow calculation to compute the
Therefore, di will always oscillate around thhe current shape objective function as well as to determine the constraint
factor si. Moreover, d is varied randomly aaround the value fulfillment, which is performed usiing the actual values in the
(1 + d 0 ) with the amplitude of d 0 adjusteed in accordance problem space. A dynamic penaalty scheme is defined to
properly consider the fulfillmentt degree along with the
with (18). dynamics of the search process. TheT fitness is calculated as
2 follows:
i
d 0 = d 0ini +
ifinal
(d0final d0ini ) (18) n
f ' = f + ncv max [ 0, gi ]
2
(19)
i =1
Experience so far has shown excellent performance by
allowing the bandwidth factor d 0 to decrease quadratically where f is the objective function value,
v ncv is the number of
within 0.40 > d 0 > 0.01 . A higher d0 entaails wider global constraint violations, gi denotes the i-th constraint, and is an
search diversification over the entire space w whereas a smaller adaptive penalty factor determined as
a follows:
one would lead to concentrated local search aiiming at accuracy = ini as long as no feasible soolution for x is obtained,
improvement. By using the described procedure the
asymmetric characteristic of the mapping funcction is also fully ifirst = iteration count when the first
f feasible solution found,
exploited by using different values for si1 annd si2 leading to i = i /2 if no feasible solutioon reached until i i /2
first max max
enhanced searching performance (i.e. robusstness) and zero
variance handling. Zero variance can occur w when all values of
2
= + i ifirst ( ) i > i
xi in the archive are identical. In this case the pprevious non-zero ini ifinal ifirst final ini first
value can be used further. However, this vvalue may result,
under circumstances, in stagnating convergennce behavior. The (20)
procedure according to (17) and (18) overcom mes this problem ini could be assigned a low vaalue (e.g. 1e3) to statically
as well. For the same reason the mean aand variance are levy low penalty to the infeasible solutions during the early
calculated for non-identical values of xi saveed in the archive searching stage, but once feasiblee solutions are found, the
only. penalty factor increases quadraticallly with the iteration number
i from ini to final (a high pre-sp pecified value, e.g. 1e6) to
The mean and variance are not calculatedd before a certain
increase the pressure with the search progress in order to drive
number of solutions are available in the archivve. Usually, they
the descendants of the best feasiible solutions towards the
end up being calculated immediately after tw wo solutions have
optimum.
been stored. However, one could also decide to do it once the
archive is filled up completely, which will resuult in more robust
initial solutions. In this stage, the search is performed with

32 2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG)


IV. THE SWARM APPROACH
1 n gbest
( )
2
The overall structure of MVMOS is shown in Fig. 4. Di = x xklbest,i
n k =1 k
(21)
Compared with classical MVMO, the swarm variant explores
the solution space more aggressively. The search process is
where n denotes the number of optimization variables. The i-th
initiated with a set of np particles, each having its own memory
defined in terms of the corresponding archive and mapping particle is discarded from the optimization process if the
function. Initially, each particle performs m steps distance Di is less than a certain user defined threshold Dmin. A
independently to collect a robust set of individual solutions. zero threshold means that all particles are considered
Then, the particles start to communicate and to exchange throughout the whole process whereas a unit threshold will
information. result in the dropping of all particles except the global best. In
this case after (m*np+np) fitness evaluations only one particle,
the gbest, remains. Intermediate threshold values entail better
adaptation to any optimization problem.
After independent evaluation, and if the particle is further
considered, its searching will be directed toward the gbest
solution by assigning x gbest , instead of x lbest,i , as parent for the
particles offspring. The remaining steps are identical with
those of the classical (single-particle) MVMO: A subset of
dimensions in the parent vector are directly inherited whereas
the remaining dimensions are selected and mutated based on
local statistical evaluation (mean and variance) of the particle
by using its own mapping function. This stage is illustrated in
Fig. 5.

Figure 5. Creation of offspring in MVMOS.

It is worth pointing out that, unlike other swarm based


optimization techniques such as PSO, MVMOS does not
strictly require several particles to proceed. Notwithstanding,
by using more than one particle in MVMOS, the global
searching capability can be considerably extended. For less
challenging optimization problems, the single-particle approach
may be sufficient. Similar to classical MVMO, the number of
iterations in MVMOS is equivalent to the number of offspring
Figure 4. Flowchart of MVMOS (The function evaluation and particle
counters are denoted by i and k, whereas m and np stand for maximum number
fitness evaluations which is in practical applications usually
of independent runs and total number of particles, respectively). much more time consuming than the optimization algorithm
itself.
However, it is not worth it to follow particles which are MVMOS requires only three additional parameters:
very close to each other since this would entail redundancy.
Hence, in the swarm implementation, the normalized distance 1) np: number of initial particles (if one is chosen, MVMOS
performs as classical MVMO)
of each particles local best solution x lbest,i to the global best
2) m: number of independent runs of particles
x gbest is calculated by: 3) Dmin: minimum distance threshold to the global best
solution

2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG) 33


V. RESULTS obtained by using the other algorith hms. Remarkably, MVMOS
Numerical experiments were performedd on a Hewlett provides the lowest minimum as in ndicated in the zoom of the
Packard Pavilion dv3 personal computer with an Intel Core last part of the convergence.
2 central processing unit (CPU), 2.2 GHz pprocessing speed, Additionally, the statistical evaaluations of the minimized
and 4GB RAM. The ORPD problem was solvved by interfacing power losses are summarized in Tab ble I, where the outstanding
routines written in Matlab for MVMOS withh the power flow performance of MVMOS is further evidenced. Besides, note
calculation functionalities of MATPOWER [[14], [15]. that the single-particle approach is also able to provide good
solutions. These issues, along with w the slight numerical
The IEEE 57- and 118- bus benchmark sysstems are used to differences associated with the otheer algorithms are reasonable
test the performance of MVMOS and compaare it with other since the small-scale optimization task
t for this test case is not
heuristic optimization algorithms. Details off system data for so challenging.
both cases can be found in [16]. Lower andd upper limits of
load bus voltages were defined at 0.95 p.uu. and 1.05 p.u,
respectively. Generator voltages at the high voltage terminal
are defined as continuous variables with lower and upper
limits set to 0.94 p.u. and 1.06 p.u. The set off discrete control
variables comprises transformer tap possitions and the
susceptance of shunt compensators. All under-load tap
changing (ULTC) transformers are assum med to have 21
discrete taps within 10% of the nominal volltage (i.e. 1% for
each position). Each transformer tap position is represented by
an integer between [-10, 10]. All shunt compensators have 11
discrete steps of different ratings (definedd by an integer
between 0 and 10).
For both test cases, the performance of MVM MOS is compared
with the classical MVMO (i.e. MVMOS exxecuted with one
particle), the comprehensive learning particle swarm optimizer
(CLPSO) [17], sliced particle swarm optim mization (SPSO)
[18], unified particle swarm optimization (UP PSO) [19], fitness
distance ratio particle swarm optimization ((FDRPSO) [20],
dynamic multi-swarm particle swarm optimizzer with harmony Figure 6. Average convergence performaance IEEE 57 bus test system.
search (DMS-PSO-HS) [21], differential evollution (DE) [22],
and an adaptive differential evolution algorithhm with variable B. IEEE -118 bus system
population size (JADE-vPS) [23]. The paraameters of these In this test case, the search spaace of this problem has 77
algorithms were tuned following the guidelinees provided in the dimensions. The continuous variab bles consist of 54 generator
corresponding references. To ensure fair com mparisons, every bus voltages. There are 23 discrete variables:
v 9 transformer taps
algorithm used the same dynamic penalty scheme (and was and 14 shunt compensators. All shun nt compensators are defined
independently run for 30 trials. Furthermorre, every search by susceptances with the lower lim mit of 0.0 p.u. and the upper
process is terminated upon completion off a pre-specified limit of 0.2 p.u. Such an optim mization task has a large
number of iterations (power flow calculationss), that is, 30000 dimensionality and is more challen nging. The population size
and 100000 for the 57 and 118 bus system ms, respectively. for MVMOS was set to 5 whereas itt was set to 100 for all other
Since the most time-consuming part of thee analysis is the algorithms.
repeated power flow calculation, the computaation time for all The average convergence of active power losses is depicted
algorithms does not differ significantlyy. Rather, the in Fig. 7 whereas the resulting statiistics are given in Table II.
comparison provided in this paper focuses on the convergence The superiority of MVMOS, in term ms of average convergence
rate and the quality of the final results. (the minimum is reached on average after 70000 function
A. IEEE -57 bus system evaluations) and lowest statisticcs, can be more clearly
The search space has 27 dimensions corrresponding to 7 appreciated in this case. Furthermo ore, it is worth mentioning
generators, 17 ULTC transformers and shunnt reactive power that only one particle remaineed after 78000 function
compensators connected to buses 18, 25 and 53, whose evaluations, thus, highlighting the enhanced search capability
susceptance limits are [0,0.2], [0,0.18] annd [0,0.18]. The achieved by starting the optimizattion with several particles.
population size for MVMOS was set to 5 wheereas it was set to Moreover, this demonstrates thaat running MVMOS with
50 for all other algorithms. For all algorithhms, the average several particles leads to the global solution without stagnation
convergence behavior of the objective function (i.e. (as observed for other optimization algorithms).
minimization of active power losses) is shhown in Fig. 6.
Although both variants of MVMO seem to bee the fastest ones,
the results obtained with them are on averagge close to those

34 2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG)


C. Adaptive reactive power control heme intended for real-time
it can be easily embedded in any sch
Since the framework of MVMOS is ggeneric and the control purposes.
algorithm exhibits a robust and fast convergennce performance,
TABLE I. STATISTCS OF ACTIVE POWER LOSSES IEEE 57 BUS TEST SYSTEM
Algorithms
Ploss (MW)
MVMOS MVMO CL
LPSO SPSO UPSO FDRPSO DMS-PSO-HS DE
D JADE-vPS
Minimum 24.8211 24.7075 244.8358 24.8412 24.8724 24.8433 24.8372 24
4.8365 24.8336
Maximum 25.0736 25.3918 255.1816 25.1506 25.1472 25.2269 25.2178 25
5.0508 25.0244
Mean 24.8898 24.9468 244.9745 24.9298 24.9771 24.9415 24.9135 24
4.9307 24.9118
Std. 0.0547 0.1055 00.0980 0.0927 0.0706 0.0864 0.0793 0.0702
0 0.0573

TABLE II. STATISTICS OF ACTIVE POWER LOSSES IEEE 118 BUS SYSTEM
Algorithms
Ploss (MW)
MVMOS MVMO CL
LPSO SPSO UPSO FDRPSO DMS-PSO-HS DE
D JADE-vPS
Minimum 117.0802 117.0074 1200.2117 121.8049 123.1174 119.1387 123.4717 118
8.7199 118.1047
Maximum 118.1662 125.1501 1322.0461 125.4654 130.2011 123.5461 128.5504 121.1128 120.2177
Mean 117.4251 119.3353 1222.2499 123.6784 125.6709 121.6536 125.0562 119
9.7737 118.9533
Std. 0.2285 1.9386 22.0533 0.9145 1.7663 1.0501 1.2033 0.6289
0 0.5321

Figure 8. Adaptive reactive power con


ntrol scheme including online
optimization.
Figure 7. Average convergence performance IEEE 1118 bus test system.

The coupling between real-time data aacquisition, state VI. CONCLU


USIONS
estimation and MVMOS is sketched in Fig. 88, using a small- The swarm implementation of th he so-called mean-variance
size grid (comprising controllable conventionaal and renewable mapping optimization was employeed in this paper to solve the
generating systems, transformers and compenssators) created for optimal reactive power dispatcch problem. Using two
illustrative purposes. Basically, the acquired reeal-time snapshot benchmark systems, the perform mance of MVMOS was
of the system is transferred to the state eestimator, which examined and compared with otther heuristic optimization
provides the data required to solve the OR RPD problem via
tools. Numerical results showed that MVMOS possesses a
MVMOS-based optimizer. In this wayy, the optimal
settings/reactive power reference for the availlable controllable remarkably better convergence sp peed and lower statistical
devices are determined in order to optimally adapt the overall measures than the other algorith hms, advantages that are
system response to continuously changiing steady-state attributed to its underlying searcching process framework.
requirements. The proposed scheme can be fuurther extended to Simply put, it allows the start of o the search with several
include forecasts of short-term variations of demand and particles, each representing an MV VMO entity. After a certain
renewable-based energy supply as well as re-aadjusting costs to number of independent runs all entiities are guided towards the
perform the optimization in a predictive mannner. For large-size global best. If individuals are gettin
ng close to the global best,
interconnected systems, the ORPD could be aalso decomposed they are stopped. In this way, th he number of particles is
into interrelated sub-problems resemblingg a multi-area
reduced progressively until only the global best survives.
decentralized principle.
Therefore, the algorithm merits furtther attention, especially to

2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG) 35


handle hard-to-solve non-convex, multimodal, mixed integer [11] G.K. Venayagamoorthy, K. Rohrig, and I. Erlich, "One Step Ahead:
Short-Term Wind Power Forecasting and Intelligent Predictive Control
problems. As such, a scheme, based on MVMOS, is finally Based on Data Analytics," IEEE Power and Energy Magazine, vol.10,
suggested to perform real-time adaptive reactive power no.5, pp.70-78, Sept. 2012.
control. [12] R. Pringles, and J.L. Rueda, "Optimal Transmission Expansion Planning
using Mean-Variance Mapping Optimization" Sixth IEEE/PES
REFERENCES Transmission and Distribution: Latin America Conference and
Exposition, Sept. 2012.
[13] I. Erlich, G. K. Venayagamoorthy, and W. Nakawiro, A mean-variance
[1] H. Liu, and J. Jiang, Optimal reactive power dispatch based on dynamic optimization algorithm, 2010 IEEE Congress on Evolutionary
readjusting cost, 2011 IEEE Power Engineering and Automation Computation, pp.1-6, July 2010.
Conference, vol. 2, pp. 508-511, Sept. 2011.
[14] R.D. Zimmerman, C.E. Murillo-Snchez, and R.J. Thomas,
[2] J.G. Vlachogiannis, and K.Y. Lee, Optimal Operation of Smart Grids in "MATPOWER: Steady-State Operations, Planning, and Analysis Tools
an Open Energy Market Environment, 18th World Congress of the for Power Systems Research and Education," IEEE Transactions on
International Federation of Automatic Control, vol. 78, pp. 3701- 3703, Power Systems, vol. 26, no. 1, pp. 12-19, Feb. 2011.
Aug.-Sept. 2011.
[15] D. Zimmerman, MATPOWER, PSERC. [Online]. Software Available
[3] A. Saraswat, and A. Saini, Optimal reactive power dispatch by an at: http://www.pserc.cornell.edu/matpower
improved real coded genetic algorithm with PCA mutation, 2nd
International Conference on Sustainable Energy and Intelligent Systems, [16] "Power Systems Test Case Archive," University of Washington,
pp. 310-315, July 2011. Department of Electrical Engineering, Seattle, USA. [Online]
http://www.ee.washington.edu/research/pstca/.
[4] N.K. Sharma, B. D. Suresh, and S.C. Choube , Application of particle
swarm optimization technique for reactive power optimization, 2012 [17] J.J. Liang, A.K. Qin, P.N. Suganthan, and S.Baskar, "Comprehensive
International Conference on Advances in Engineering, Science and learning particle swarm optimizer for global optimization of multimodal
Management, pp. 88-93, March 2012. functions," IEEE Transactions on Evolutionary Computation, vol. 10,
no. 3, pp. 281- 295, June 2006.
[5] C.-M. Huang, S.-J. Chen, Y.-C. Huang, and H.-T. Yang, Comparative
study of evolutionary computation methods for active-reactive power [18] H. Garg, S.S. Pattnaik, S. Devi, K.M. Bakwad, B.K. Panigrahi, and S.K.
dispatch, IET Comparative study of evolutionary computation methods Das, "Sliced Particle Swarm Optimization (SPSO): A computationally
for active-reactive power dispatch, vol.6, no.7, pp. 636- 645, July 2012. efficient optimization technique," in Proc. 2009 World Congress on
Nature & Biologically Inspired Computing, pp. 61-66, Coimbatore,
[6] W. Nakawiro, I. Erlich, and J.L. Rueda, A novel optimization algorithm India, Dec. 2009.
for optimal reactive power dispatch: A comparative study, 4th
International Conference on Electric Utility Deregulation and [19] K. E. Parsopoulos and M. N. Vrahatis, "Unified Particle Swarm
Restructuring and Power Technologies, pp. 1555-1561, July 2011. Optimization for Solving Constrained Engineering Optimization
Problems," Springer-Verlag, pp. 582- 591, 2005.
[7] I. Erlich, F. Shewarega, C. Feltes, F. Koch and J. Fortmann,
"Determination of Dynamic Wind Farm Equivalents using Heuristic [20] E. Muneender, D. Kumar, " Optimal rescheduling of real and reactive
Optimization," 2012 IEEE PES General Meeting, San Diego, USA, July powers of generators for zonal congestion management based on FDR
2012. PSO," in Proc. 2009 IEEE Transmission & Distribution Conference &
Exposition: Asia and Pacific, pp. 1-6, Seoul, South Korea, Oct. 2009.
[8] J.C. Cepeda, J.L. Rueda, and I. Erlich, "Identification of Dynamic
Equivalents based on Heuristic Optimization for Smart Grid [21] J S. Z. Zhao, P. N. Suganthan, Quan-Ke Pan, M. Fatih Tasgetiren, "
Applications" 2012 IEEE world congress on computational intelligence, Dynamic Multi-Swarm Particle Swarm Optimizer with Harmony
June 2012. Search," Expert Systems with Applications, vol. 38, no. 4, pp. 3735
3742, April 2011.
[9] J.L. Rueda, J.C. Cepeda, and I. Erlich, "Estimation of Location and
Coordinated Tuning of PSS based on Mean-Variance Mapping [22] K. Price, R. Storn, and J. Lampinen, Differential Evolution A Practical
Optimization," 2012 IEEE PES General Meeting, July 2012. Approach to Global Optimization, Springer Verlag, 2005.
[10] P. Chakravarty and G.K.Venayagamoorthy, "Development of optimal [23] W. Nakawiro, and I. Erlich, "A new adaptive differential evolution
controllers for a DFIG based wind farm in a smart grid under variable algorithm for voltage stability constrained optimal power flow," in Proc.
wind speed conditions," 2011 IEEE International Electric Machines & 17th Power Systems Computation Conference, Stockholm, Sweden,
Drives Conference, pp. 723-728, May 2011. Aug. 2011.

36 2013 IEEE Computational Intelligence Applications in Smart Grid (CIASG)

Potrebbero piacerti anche