Documentos de Académico
Documentos de Profesional
Documentos de Cultura
a r t i c l e
i n f o
Article history:
Received 11 September 2015
Received in revised form 8 December 2015
Accepted 22 December 2015
Available online 14 January 2016
Keywords:
Distribution system
Voltage and reactive power control
Stochastic search algorithm
Dynamic programming
a b s t r a c t
This paper considers voltage and reactive power control models that take into account the operational
costs of switching devices and the correlation between loads and voltage proles. In addition to maintaining acceptable voltage at all points along the distribution feeder, the aim is to determine the proper
settings of shunt capacitors and transformer load tap changers in order to minimize the energy loss or
the energy consumption over a selected planning horizon. A random search algorithm is proposed for
optimizing these models. The algorithm searches the optimal control schedule by randomly sampling
from a sequence of probability distributions over the space of all possible settings of tap changers and
shunt capacitors. The algorithm is shown to be globally convergent and its effectiveness is illustrated by
comparing its performance with that of three other commonly used procedures on a 69-bus distribution
system and a 24-bus system with distributed generation.
2015 Elsevier B.V. All rights reserved.
1. Introduction
The voltage and reactive power control (VVC) problem is an
important category of optimization problems in smart grid, especially in power distribution systems. One of the objectives of VVC
is to adjust the voltage prole and the reactive power supply to
increase the energy efciency. Usually, two efciency measures are
used: the energy loss and the total energy consumption. The voltage prole and the reactive power supply are controlled through
adjusting tap positions of transformer load tap changers (LTCs) and
on/off states of shunt capacitors (SCs), so optimal control schedules
of LTCs and SCs need to be computed in VVC problems.
The VVC problem is usually formulated as a mixed-integer nonlinear programming problem (MINLP), where tap positions of LTCs
and on/off states of SCs are modeled as discrete variables while
other variables (e.g., voltages and power ows) are continuous.
Many algorithms have been proposed for solving the problem.
Dynamic programming (DP) approaches were proposed in Refs.
[79,15,19]. A standard DP algorithm was used in Ref. [7] for solving a problem with four capacitors. However, as the number of
capacitors increases, the complexity of the algorithm increases
exponentially due to the well-known curse of dimensionality.
To address this issue, heuristic procedures were proposed in Refs.
[15,19] to try to reduce the sizes of state and action spaces, whereas
Corresponding author.
E-mail address: jqhu@ams.sunysb.edu (J. Hu).
http://dx.doi.org/10.1016/j.epsr.2015.12.025
0378-7796/ 2015 Elsevier B.V. All rights reserved.
Nomenclature
Nsc
Nltc
xsc (i, t)
xsc
xltc (i, t)
xltc
x
xt
L
Nnode
PL(xt )
V
P
Q
I
V
V
329
T
t=1
Floss (xt , t) =
L
(1)
PLt,l (xt ) +
l=1
+ + sc
Nltc
i=1
Nsc
i=1
330
2
T Nltc
t=1
i=1
T Nsc
is the total switching cost of LTCs, and sc t=1 i=1 |xsc (i, t)
xsc (i, t 1)| is the total switching cost of SCs.
The problem of minimizing the total energy consumption with
switching costs can be formulated as:
min
T
Ftotal (xt , t) =
S
i
Vi
Nnode
l=1
(7)
i=1
xltc (i, t) xltc (i, t 1)
+ ltc
Nltc
i=1
Nsc
sc
x (i, t) xsc (i, t 1) ,
(3)
P(i, t) =
i=1
(4)
V (i, t)
S0 (i, t) Z % cos(Z ) +
V02
V (i, t)
S0 (i, t) I % cos(I )
V0
+ S0 (i, t) P % cos(P ),
Q (i, t) =
S(i, t)
PLt,l (xt ) +
+ sc
(6)
(2)
t=1
L
[V ] = Vsub [V ],
V (i, t)
V02
S0 (i, t) Z % sin(Z ) +
V (i, t)
S0 (i, t) I % sin(I )
V0
+ S0 (i, t) P % sin(P ),
where V(i, t) is the voltage level of node i at time t and V0 is the nominal voltage, for example, 120 V. S0 (i, t) is the electricity load at V0
for node i at time t and can be estimated through load forecasting.
P(i, t) and Q(i, t) are the respective real and reactive load of node
i at time t. The ZIP model assumes that the electricity load is the
sum of three parts: a constant impedance component, a constant
current component, and a constant power component. Nk 0 .
is the percentage of the constant impedance load, I% is the percentage of the constant current load, and P% is the percentage of the
constant power load. Z , I and P are phase angles of these three
components. For simplicity, assume that Z = I = P = S , where S is
the phase angle of S0 (i, j). Then it follows that
P(i, t) =
3. Power ow computation
V (i, t)
P0 (i, t) Z % +
V02
V (i, t)
P0 (i, t) I % + P0 (i, t) P % ,
V0
(8)
(5)
where [V] is the voltage drop vector of length Nnode with its ith
element being the difference between the voltage levels at the
substation and the ith node, [DFL] is a matrix that transforms the
current injection vector into the voltage drop vector, and [I] is the
current injection vector with its ith element being the current that
the ith node injects into the system, which can be calculated as Si /Vi .
[BIBC] is an L L matrix representing the impedance of branch lines,
and [BIBC][I] is a vector that represents voltage drops between the
two ends of all branch lines. [BCBV][BIBC][I] adds the voltage drop
of each branch line along the paths between the substation and a
node to get the voltage drop of each node. The voltage level of each
node can be represented as
Q (i, t) =
V (i, t)
V02
Q0 (i, t) Z % +
V (i, t)
Q0 (i, t) I % + Q0 (i, t) P % ,
V0
(9)
where P0 (i, j) and Q0 (i, j) are the respective real and reactive parts
of S0 (i, j). In this work, the load forecast is assumed to be voltage
dependent. In traditional electric grids, the load forecast is usually based on the historical load data at the feeder level and the
weather forecast [6]. Measurements of voltage levels and loads at
the customer level are usually not available. Thus, the load forecast
is usually independent of voltage proles. However, as the installation of smart meters increases, the historical load and voltage data
can be obtained at the customer level, which makes the voltage
dependent load forecast feasible in smart grids.
3.3. Power ow computation with the ZIP load model
When the ZIP load model is involved, the power ow computation method in Section 3.1 is no longer applicable, because Si is not
a constant but a variable depending on voltage proles. However,
P(i, t) + jQ(i, t) can be used to replace Si in Eq. (7). P(i, t) and Q(i, t) can
be calculated according to Eqs. (8) and (9). Then, the same recursive
algorithm in Section 3.1 can be used to compute power ows.
Note that for systems with distributed generation units, the
computation of the power ow can also be easily handled using the
same approach by treating the outputs of distributed generators as
negative constant power loads.
qsc
(i, t)
k+1
x Xk
h(x) =
T
t=1
(max(V (i, t) V , 0)
i=1 t=1
(10)
qltc
(i, j, t)
k+1
qsc
t)
k (i,
xsc (i,t)
1 qsc
(i, t)
k
1xsc (i,t)
i=1
Nltc
qltc
(i, xltc (i, t), t)
k
i=1
(x; qsc
, qltc
)=
k
k
T
t=1
ltc
(x;
qsc
, qltc
) = (1 k )(x; qsc
, qltc
) + k (x; qsc
0 , q0 ).
k
k
k
k
ltc
k
f (x)I{xltc (i,t)=j}
x Xk
f (x)
x xk
f (x)
)qsc
(i, t) (11)
+ (1 sc
k
k
+ (1 ltc
)qltc
(i, j, t),where
k
k
x xk
x Xk ,
1.
Assume
sc 2
and
( ) < ;
2 k
ltck=0
k=0
Nsc
x xk
Nnode T
F(xt , t) + v
f (x)I{xsc (i, t) = 1}
sc
k
331
Then
< ;
that
(1) sc
> 0k,
k
(2) ltc
> 0k,
k
(3) Tk 0 as k ;
qsc
(i, t) I{xsc (i, t) = 1}, i, t,
k
qltc
(i, j, t)
k
I{x
sc
(i, t) = j}, i, j, t,
sc
k=0 k
ltc =
k=0 k
=
and
(4) Nk k as k .
(12)
(13)
nite horizon MDP (X, A, R, P, T), where the state space X is the set
of all possible settings of LTCs and SCs at time t, the action space
A = {a := (altc , asc )} is the set of all feasible operations of LTCs and SCs,
R(x, a) = F(x, t) is the one-stage cost function, and T is a pre-dened
planning horizon.
The system evolves as follows. Let xt be the state of the system at
time t and at be the action applied at time t. If xt = x X and at = a A,
then at time t + 1 the system will transition to a deterministic state
xt+1 specied by action a with probability one, and a cost of R(x, a)
is incurred.
The standard DP method with running complexity O(|X|2 |A|T)
can be applied for solving the MDP. Note that an action a is feasible
332
only if the settings of LTCs and SCs specied by a are feasible, for
example, the new state xt+1 the system transitions to by applying
action a at time t should satisfy the voltage constraints.
The computational complexity of the dynamic programming
algorithm may become prohibitive when the system under study
is large. However, the algorithm guarantees the optimality of
the resulting solutions and could be a very reliable approach for
solving VVC problems when the numbers of SCs and LTCs are
relatively small.
5.2. Simulated annealing
For the purpose of numerical comparison, the simulated annealing (SA) algorithm proposed in Ref. [16] is implemented, but with
different objective functions and power ow equations. Let H(x)
be the objective function. The SA algorithm for VVC is described
below.
1. Initialize a feasible setting of LTCs and SCs, e.g., an initial feasible
solution x0 . Specify an annealing schedule {Tk }.
2. Randomly perturb a value in xk to get a new solution x .
3. Solve the power ow equations at x . If one or more constraints
are violated, go to step 2; otherwise, compute the objective function H(x ) and continue to step 4.
4. If H(x ) H(xk ), set xk+1 = x ; otherwise, set xk+1 = x with prob
ability e[H(xk )H(x )]/Tk and set xk+1 = xk with probability 1
)]/T
[H(x
)H(x
k
k
e
.
5. If a stop condition is satised, stop; otherwise, let k = k + 1 and go
back to step 2.
Note that step 2 is executed repeatedly until a feasible solution x satisfying all voltage constraints is generated. Therefore,
the original objective function instead of the augmented objective function with penalty terms is used in the optimization
process. In other words, the function H at step 3 is taken
T
to be H(x ) =
F (x , t) when minimizing the energy loss
t=1 loss t
H(x )
T
=
and
consumption.
F
(x , t)
t=1 total t
2.0 GB RAM and Windows 7 OS. The following set of parameter values is used in all experiments: smoothing parameters
0.51
0.6
sc
= 1/(k + 100)
and ltc
= 1/(k + 100) , sample size Nk =
k
k
max(50, k), annealing schedule Tk = c/(k + 1)0.5 , where c is the difference between the minimum and the median of all objective
function values obtained at step 3 of the ASA algorithm. Note that
the above setting satises the relevant conditions in Theorem 1
for convergence. In addition, the penalty coefcient for violation
of voltage constraints is set to v = 1. The algorithm stops either
when the number of iterations reaches 10,000 or when the following conditions are satised: min(1 qsc,k (i, t), qsc,k (i, t)) < 0.001, i,
t, and min(1 qltc,k (i, j, t), qsc,k (i, j, t)) < 0.001, i, j, t, indicating that
elements of the two stochastic matrices are either getting very close
to 0 or very close to 1.
The DP algorithm is implemented using MATLAB on the same
plat-form as the ASA algorithm. In particular, the standard backward induction algorithm is implemented.
In actual implementation of the SA algorithm, the following parameter values are used: the annealing schedule
Tk+1 = Tk 0.9995, and T1 = 1000. The search process ends when
Tk < 0.001. At step 2 of the SA algorithm, a control device, e.g., LTC
or SC, is randomly selected, and its status is perturbed to generate
a new solution.
For the genetic algorithm, the following parameter values are
used: the crossover probability is set to 0.8, mutation probability is set to 0.01, and the size of population is set to 50. As in
ASA, the maximum number of iterations is set to 10,000. The
algorithm is stopped either when the maximum number of iterations is reached or when no further improvement in the tness
value is obtained for 1000 consecutive iterations. The penalty for
the violation of voltage constraints is the same as that used in
ASA.
333
Fig. 1. (a) Diagram of the PG&E 69-bus system; (b) normalized load proles.
Table 1
Energy losses with voltage-independent load and no switching cost.
Table 3
Energy losses with ZIP model and no switching cost.
Comparison algorithms
ASA
DP
SA
GA
Comparison algorithms
ASA
DP
SA
GA
3003.2
1638
2
16
3003.2
8207
2
16
3012.7
2395
2
54
3028.5
1764
10
69
3003.5
1541
5
19
3003.4
7311
5
17
3012.4
2106
7
51
3013.2
1462
15
62
Table 4
Energy losses with ZIP model and non-zero switching costs.
Comparison algorithms
ASA
DP
SA
GA
3010.6
1475
3
11
3010.0
7599
3
11
3032.4
2337
3
31
3045.3
1788
11
67
Table 5
Total energy consumption with ZIP model and no switching costs.
Comparison algorithms
ASA
DP
SA
GA
Comparison algorithms
ASA
DP
SA
GA
3009.7
1577
2
10
3009.1
9296
2
10
3020.5
2628
2
22
3083.9
1666
16
64
159295.3
2074
3
15
159295.1
7209
3
11
159304.0
2130
3
29
159331.7
1918
11
61
334
0.9
0.9
0.8
0.8
0.7
0.7
Normalized Power
Normalized Power
Active Power
Reactive Power
0.6
0.5
0.4
0.6
0.5
0.4
0.3
0.3
0.2
0.2
0.1
0.1
10
15
20
25
10
15
Time(h)
Time(h)
(a)
(b)
20
25
Fig. 3. (a) Load curves of the 24-bus system; (b) capacity factor curve of photovoltaic generators.
with buses 15, 19, 24, respectively. The load curves for active and
reactive powers are shown in Fig. 3a and the load capacity curve of
photovoltaic generators is given in Fig. 3b.
Tests were performed on the system with the load ZIP model
and non-zero switching costs, where the base power and voltage
of the system and the parameters of the ZIP model and switching
costs of LTC and SCs are taken be the same as those described in
Section 6.2.
Tables 7 and 8 show the performance of the four comparison
algorithms for the respective objectives of minimizing the energy
Table 7
Energy losses in the 24-bus system with ZIP model and switching costs.
Comparison algorithms
ASA
DP
SA
GA
345.7
654
3
9
344.9
945
3
7
346.6
674
3
11
354.9
1260
12
43
Table 8
Energy consumption in the 24-bus system with ZIP model and switching costs.
Comparison algorithms
ASA
DP
SA
GA
Comparison algorithms
ASA
DP
SA
GA
159299.4
1651
3
5
159299.1
8917
3
5
159313.8
2193
3
17
159342.5
2128
7
43
104018.6
402
3
0
104018.6
1002
3
0
104018.6
585
3
0
104033.8
992
21
43
335
The proof of Theorem 1 follows the proof given in Ref. [10], with
modications appropriate to the context of solving VVC problems.
Let Fk be the -eld generated by the set of solutions sampled up
to iteration k 1, i.e., Fk = {X1 , X2 , . . ., Xk1 }. The conditional prob
ability P( | ) and expectation E[ | ] are with respect to (x;
qsc
, qltc
)
k
k
at the kth iteration. Let S denote the solution space of x (note that
x = (xsc , xltc )).
For
notational
convenience,
dene
Yksc =
xS
j}, Zk =
xS
Y ksc = Nk1
Ykltc =
xS
e(h(x)/Tk+1 ) .
x Xk
Y kltc = Nk1
x Xk
Z k = Nk1
x Xk
336
Let gk+1 (x) = e(h(x)/Tk+1 ) / x S e(h(x)/Tk+1 ) be a Boltzmann probability mass function on the set of feasible solutions. Let g k+1 (x) =
1 (x; qsc , qltc )/
1 (x; qsc , qltc ), where
e(h(x)/Tk+1 )
e(h(x)/Tk+1 )
xS
k
k
k
k
Xk is the set of samples generated at iteration k. Thus g k+1 (x) can
be viewed as an empirical Boltzmann probability mass function
that approximates gk+1 (x).The next result shows that the sequence
of idealized Boltzmann distribution {gk } converges to a limiting
distribution that concentrates only on the optimal solution x*.
Lemma 1.
If Tk 0 as k , then as k ,
(A.1)
(A.2)
sc
sc
(i, t) = 1}|]
|I{x (i, t) = 1}
eh(x)/Tk
x se
h(x)/Tk
x=
/
h(x)/Tk
|I{xsc (i, t) = 1}
1+
x=
/ xe
x=
/ xe
(h(x)h(x))/Tk
x=
/ x
x=
/ x
Lemma 2 shows that Y ksc /Z k and Y kltc /Z k are asymptotically unbiased estimators of Y sc /Z k and Y ltc /Z k .
k
Y ksc
|Fk
Z k
Proof.
If Nk k as k , then w.p.1
Yksc
Zk
Y kltc
Z k
|Fk
Ykltc
Zk
sc
sc
Y k
Y k Zk Z k
Yksc
E
|Fk
Z |Fk Zk = E
Z Z
k k
E[|Zk Z k |Fk ]
Zk
Y ksc
Z k
2 1/2
1
E Zk Z k |Fk
H older
sinequality
Zk
1
Zk Nk
1
Nk
xS
1/2
1 (x; qsc , qltc )
e(2h(x)/Tk+1 )
k
k
xS
=
since
1/2
ltc 1
0 (x; qsc
gk+1 (x)
0 , q0 )
xS
1/2
1 (x; qsc
, qltc
)gk+1 (x)
0
0
Nk 0
2
E[k ] < ; (iv)
k=1
P(lim inf n > 0, k=1 Uk < ) = 0.
sc 2
k
< and (iv) can be proved
k=0
(h(x)h(x))/Tk
e(h(x)h(x))/Tk .
Lemma 2.
eh(x)/Tk
x se
sc
x=
/ x
References
[1] A.G. Bakirtzis, P.N. Biskas, C.E. Zoumas, V. Petridis, Optimal power ow by
enhanced genetic algorithm, IEEE Trans. Power Syst. 17 (May (2)) (2002)
229236.
[2] M.E. Baran, F.F. Wu, Optimal capacitor placement on radial distribution systems, IEEE Trans. Power Deliv. 4 (January (1)) (1989) 725734.
[3] F. Batrinu, E. Carpaneto, G. Chicco, M. De Donno, R. Napoli, R. Porumb, P. Postolache, C. Toader, New nested evolutionary programming approach for voltage
control optimization with distributed generation, in: Proceedings of the 12th
IEEE Mediterranean Electrotechnical Conference, 2004, pp. 10071010.
[4] F. Capitanescu, I. Bilibin, E. Ramos, A comprehensive centralized approach for
voltage constraints management in active distribution grid, IEEE Trans. Power
Syst. 24 (2014) 933942.
[5] S.N. Evans, N.C. Weber, On the almost sure convergence of a general stochastic
approximation procedure, Bull. Aust. Math. Soc. 34 (1986) 335342.
[6] E.A. Feinberg, D. Genethliou, in: J.H. Chow, F.F. Wu, J.J. Momoh (Eds.), Load
Forecasting, Applied Mathematics for Restructured Electric Power Systems:
Optimization, Control, and Computational Intelligence, Springer, 2005, pp.
269285.
[7] Y.-Y. Hsu, H.-C. Kuo, Dispatch of capacitors on distribution sys-tem using
dynamic programming, IEE Proc. Transm. Distrib. 140 (6) (1993 Nov)
433438.
[8] Y.-Y. Hsu, F.-C. Lu, A combined articial neural network-fuzzy dynamic
programming approach to reactive power/voltage control in a distribution
substation, IEEE Trans. Power Syst. 13 (November (4)) (1998) 12651271.
[9] Y.-Y. Hsu, C.-C. Yang, A hybrid articial neural network-dynamic programming
approach for feeder capacitor scheduling, IEEE Trans. Power Syst. 9 (May (2))
(1994) 10691075.
[10] J. Hu, H.S. Chang, An Approximate Stochastic Annealing algorithm for nite
horizon Markov decision processes, 2010 49th IEEE Conference on Decision
and Control (CDC) (December 2010) 53385343.
[11] Z. Hu, X. Wang, H. Chen, G.A. Tayloy, Generation, Volt/Var control in distribution
systems using a time-interval based approach, IEE Proc. Transm. Distrib. 150
(September (5)) (2003) 548554.
[12] C. Ionescu, C. Bulac, F. Capitanescu, L. Wehenkel, Multi-period power loss optimization with limited number of switching actions for enhanced continuous
power supply, in: The 16th IEEE International Conference on Harmonics and
Quality of Power, May 2014, pp. 3438.
[13] D. Kirshner, Implementation of conservation voltage reduction at Commonwealth Edison, IEEE Trans. Power Syst. 4 (November (4)) (1990) 11781182.
[14] D.M. Lauria, Conservation voltage reduction (CVR) at Northeast utilities, IEEE
Trans. Power Deliv. 2 (October (4)) (1987) 11861191.
[15] R.-H. Liang, C.-K. Cheng, Dispatch of main transformer ULTC and capacitors
in a distribution system, IEEE Trans. Power Deliv. 16 (October (4)) (2001)
625630.
[16] R.-H. Liang, Y.-S. Wang, Fuzzy-based reactive power and voltage control in a
distribution system, IEEE Trans. Power Syst. 18 (April (2)) (2003) 610618.
[17] M.B. Liu, C.A. Canizares, W. Huang, Reactive power and voltage control in distribution systems with limited switching operations, IEEE Trans. Power Syst.
24 (May (2)) (2009) 889899.
[18] Y. Liu, P. Zhang, X. Qiu, Optimal vol/var control in distribution systems, Int. J.
Electr. Power Energy Syst. 24 (4) (2002) 271276.
337
[23] T. Senjyu, Y. Miyazato, A. Yona, N. Urasaki, T. Funabashi, Optimal distribution voltage control and coordination with distributed generation, IEEE Trans.
Power Deliv. 23 (2008) 12361242.
[24] R. Shalwala, J. Bleijs, Voltage control scheme using Fuzzy Logic for residential
area networks with PV generators in Saudi Arabia, IEEE PES Innovative Smart
Grid Technologies (ISGT), 2011, pp. 16.
[25] J.-H. Teng, A direct approach for distribution system load ow solutions, IEEE
Trans. Power Deliv. 18 (July (3)) (2003) 882887.
[26] H.L. Willis, Characteristics of Distribution Loads, Electrical Transmission and
Distribution Reference Book, ABB Power T&D Company, Raleigh, NC, 1997.
[27] Y. Zhang, Z. Ren, Optimal reactive power dispatch considering costs of adjusting
the control devices, IEEE Trans. Power Syst. (2005) 13491356.