You are on page 1of 6

OPTIMAL LOAD SHEDDING STRATEGY BASED ON

PARTICLE SWARM OPTIMIZATION

Fei HE1,2 Yihong WANG1,2 Ka Wing CHAN1 Yutong ZHANG1 Shengwei MEI2
1
Department of Electrical Engineering, The Hong Kong Polytechnic University
2
State Key Lab of Power Systems, Department of Electrical Engineering, Tsinghua University

ABSTRACT [3] proposes a problem of load cut minimization


using the Interior Point algorithm. [4] formulates an
This paper proposed a new optimization model to OPF which load shedding strategy is based on the
minimize power loss and the load curtailments minimization of the difference between the sum of
necessary to restore the equilibrium of operating the generations and the amount of connected loads.
point. The solution algorithm is based on the particle [5] proposes a model used to simulate the cascading
swarm optimization (PSO) method in which the load failure. In this model, the load near stressed line will
shedding would be considered as the penalty in the be shed if the conventional OPF cannot converge.
optimization cost function. The aim function would
be optimized for minimum power loss under normal Most of the existent methods are based on gradient,
operating conditions and minimum load shedding and have limitation in the location of the load
during emergency conditions. In contrary to the shedding and the amount of the load shedding. The
other load shedding strategies, there is no limitation PSS algorithm breaks a new path in solve the load
in where the load could be shedding and how much shedding problem.
could be shedding. Furthermore, this algorithm has
the advantages of PSO method. That is this In this paper, a load shedding strategy based on PSO
algorithm doesn’t require the gradient information, is proposed. It regulates the generator and the load
and has better global search ability. The simultaneously; optimizes the power loss and load
effectiveness of this model is verified by the shedding with the same model. The content of this
simulation results concerning the IEEE 30-bus paper is arranged as follows: the PSO algorithm is
system. first introduced briefly, and then the new strategy is
proposed. The verification of the effectiveness of
this new method on the IEEE 30-bus system is then
1. INTRODUCTION presented and followed by discussions and
conclusions.
The main objective of power system is to supply
energy with quality, continuity and economy
requisites. In order to get more economic balance,
lots of equipments often operate close to their 2. PARTICLE SWARM OPTIMIZATION
limitations. As a result, some large-scale blackouts
Particle swarm optimization (PSO) is a novel
have happened in recent years, and scholars have
optimization method proposed by Kennedy and
paid more attention to this recently.
Eberhart[6]-[8]. This algorithm is motivated by
social behavior of organisms, such as bird flocking.
Once blackout happens, lots of load will lost, so in
It is one of evolutionary computation techniques
some emergencies, some load should be shedding to
essentially. In PSO, a number of particles constitute
avoid blackout. An algorithm which could help the
a swarm, and each particle is a solution of the
operators to determine which load and how many of
optimization problem. Each particle flies in the
it could be shedding to stabilize the system is needed.
search space towards the global best position
according to its own experience and experience of
Different methods have been proposed for load
neighboring particles, making use of the best
shedding. [1] considers a boarding based on
position encountered by itself and its neighbors.
sensitivity that attributes different priorities to the
load and limit the total cut. [2] proposes a method
to re-dispatch the generator besides load shedding. Considers a swarm consists of N particles. At
The formulations penalizes as the load shedding as generation t, each particle i has its position xi ,t , and
the deviation of dispatch generators. This nonlinear its flight velocity vi ,t . For each particle, a fit value f
problem is then solved by using linear programming. based on its position xi ,t could be calculated, and
this value quantifies the quality of the position of inequality constraints. The common format of
this particle. Generally, this fit value is the function optimization problem is
value of the aim function of the optimization min f ( x)
problem. Supposes the optimization problem is a
st. gi ( x) = 0 i=1,2,…N eq (3)
minimization problem, so the less the value is, the
better the position is. For a particle, the best h j ( x) ≤ 0 j=1,2,…N ueq
previous position of itself is represented as pbesti ,t .
The above problem with constraint could be
The best previous position of all the particles is
transformed to following problem without constraint.
represented as gbestt . In order words, gbestt is
the best of pbesti ,t . So the position and velocity of N eq N ueq

particle i in the next generation can be calculates as min f ( x ) + wg ∑ g i ( x ) + wh ∑ max(0, h j ( x))


i =1 j =1

vi ,t +1 = w ⋅ vi ,t + ϕ1 ⋅ r1 ⋅ ( pbesti ,t − xi ,t ) (4)
where wg and wh are penalty factors which are
+ ϕ2 ⋅ r2 ⋅ ( gbest t − xi ,t ) (1)
chosen large number. Because the penalty factors
xi ,t +1 = xi ,t + vi ,t +1 are so large, the control variable must vary to the
point where no constraints are unsatisfied.
where w is inertia weight; ϕ 1 and ϕ 2 are
acceleration constants; r1 and r2 are uniform random
number in the range of [0, 1]. 3. LOAD SHEDDING STRATEGY BASED ON
PSO ALGORITHM
After generating the new generation as above, the
3.1 OPTIMAL POWER FLOW (OPF)
pbestit +1 and gbest t +1 could be calculated:
OPF is a typical optimization problem for which the
t +1
⎧⎪ pbestit if f (xit +1 )>pbestit PSO method is very suitable. Normally, the control
pbest i = ⎨ t +1 (2)
⎪⎩ xi if f (xit +1 )<pbestit variable x in an OPF formulation contains active
power and reactive power (or voltage) of generator.
gbest t +1 = best ( pbest1t +1 , pbest2t +1…pbest Nt +1 ) Suppose that the dispatcher wants the system to
operate with less power loss, the aim function of
Fig.1 shows the flowchart of the basic PSO method. OPF could be the power loss.

min f ( P, Q) = ∑ Pgi − ∑ Plj (5)


i∈G j∈L

where G is the generator set, and L is the load set;


Pgi is the active power output of generator i, and
Plj is active load of load j; P and Q are generator
output (active and reactive power), they are
controllable variable.

There are many constraints in OPF. Suppose the


system has n buses and m transmission lines. And
the most important four constraints are:

1. The power flow equation


n
Pi = Vi ∑ V j (Gij cos θij + Bij sin θij ) (6)
i =1
n
Qi = Vi ∑ V j (Gij sin θij − Bij cos θij ) (7)
Fig.1 The flowchart of PSO i =1

Generally, penalty function is used if there are some where Pi is the active power injected in bus i; Qi is
constraints in the optimization problem. Suppose the reactive power injected in bus i; Vi is the
there are N eq equality constraints and N ueq magnitude of voltage of bus i; Gij , Bij are
conductance and susceptance of transmission line selected, one variable of it is chosen randomly to
from bus i to bus j; θ ij = θ i − θ j is phase-angle modify its value. For example, the variable j of the
difference between bus i and bus j. particle i is selected to be mutated. The value of
xij is modified as follows:
2. Generator output constraint;
r = rand()− 0.5
Pgimin ≤ Pgi ≤ Pgimax (8)
⎪⎧ xi + r × ( xmax − xi ) if(r ≥ 0) (13)
j j j

Q min
≤ Qgi ≤ Q max
(9) xij = ⎨ j
⎪⎩ xi + r × ( xi − xmin ) if(r < 0)
gi gi j j

where Pgi and Qgi are active power and reactive where rand() returns a uniform random number in
power of the generator i respectively; Pgimax and the range of [0,1].
Qgimax are upper limits; Pgimin and Qgimin are lower
limits. As a result, the proposed OPF based on PSO method
is solved as shown in the flowchart in Fig.2.
3. Voltage constraint

Vi min ≤ Vi ≤ Vi max (10)

where Vi max and Vi min are upper limit and lower


limit of Vi .

4. Transmission capacity constraint.

− Fl max ≤ Fl ≤ Fl max (11)

where Fl is power flow of line l , and Fl max is the


capacity of line l .

3.2 IMPLEMENTATION OF PSO FOR OPF

With the help of penalty function, the OPF problem


could be transformed to a problem without
constraint. Fortunately, the equality constraints are
imposed implicitly by the power flow calculation
and hence needn’t be included into the aim function.
Fig 2 The OPF based on PSO
Hence the following optimization problem is
obtained.
3.3 PSO BASED LOAD SHEDDING
min F ( P, Q ) = ∑ Pgi − ∑ Plj STRATEGY
i∈G j∈L
N ueq
(12)
As mentioned in previous sections, some load could
+ wh ∑ max(0, h j ( x)) be shed in emergency condition to protect the other
j =1
load. But it is difficulty for dispatcher to determine
where inequality constraints contain generator which load and how many could be shedding.
output constraint, voltage constraint and Because load shedding bring up large economic lost,
transmission capacity constraint. dispatcher must try his best to protect majority load.
Therefore, an algorithm of load shedding could be a
When the particle swarm comes to a suboptimal great help to the dispatcher.
point, the particle maybe moves near this point until
It is noticed that the penalty function in PSO could
the velocity becomes zero, and then the algorithm
be considered as the load shedding in a power
will not be able to find the optimal point but a
system which is the penalty for maintaining the safe
suboptimal point. As a remedy, a mutation as found operation of the power system, and the load
in Genetic Algorithm can be introduced to the PSO. shedding strategy can be optimized with the PSO.
In this paper, a simple mutation schedule is adopted
as follows. After updating the particle swarm, 20% The mathematical model of this new strategy is:
particles are chosen to be mutated. For every particle
min f ( P, Q) = ∑ Pgi − ∑ Plj' + ∑ wlj ( Plj − Plj' ) part is the penalty of violation of constraints; and the
i∈G j∈L j∈L third part is the economic cost of load shedding. The
n inequality constraint now only contains voltage
st. Pi = Vi ∑ V j (Gij cos θij + Bij sin θij ) constraint and transmission capacity constraint.
i =1
n
Qi = Vi ∑ V j (Gij sin θij − Bij cos θij ) It should also be noted that load shedding is much
i =1 (14) more expensive than power loss, so the factor wlj
P min
≤ Pgi ≤ Pgimax must be much larger than 1; at the same time, the
gi
constraint must be submitted, so the penalty factor
Qgimin ≤ Qgi ≤ Qgimax wh must be much large than wlj . This means
Vi min ≤ Vi ≤ Vi max 1 << wlj << wh . In this paper, the settings adopted
are wlj =1000 and wh =100000.
− Fl max ≤ Fl ≤ Fl max
0 ≤ Plj' ≤ Plj The setting of penalty factor as above is to make a
masking effect. The particle firstly flies to position
The control variable P and Q not only contains the where no constraint is violated, because the penalty
output of generator but also the active load. Plj is to constraint violation is highest. Secondly, the
demand power at bus j, but only Plj' is supplied, so particle searches the position where no load is
the amount of load ( Plj − Plj' ) is shed, the factor shedding. And at last, the particle tries to decrease
wlj is the economical lost coefficient for the load j. the power loss.
It must be pointed that the reactive load always
variable with the active load. In this paper, the
reactive load will decrease with active load with the 4. CASE STUDIES
power factor remaining unchanged.
This new load shedding strategy is applied to IEEE
30-bus system to verify its effectiveness. The Fig.3
While the equality constraints could be eliminated as
shows the connection diagram of IEEE 30-bus
explained in previous section, some of the inequality system.
constraints could be handled easier than being added
to the aim function as a penalty function. It is also
noted that generator outputs and the loads are
control variables, and they could be bounded after
updating the position of particles.

Suppose x is the control variable vector consisting of


P and Q, etc., and it’s dimension is nx . For particle i,
let xi = ( xi1 , xi2 , …, xinx ) , and the upper limit and
j
lower limit for xij are xmax j
and xmin respectively.
After updating xi according to xi ,t +1 = xi ,t + vi ,t +1 ,
xi becomes the following:
⎧ xmin
j
if xij < xmin
j
⎫ Fig.3 IEEE 30 buses-system
⎪ j ⎪
xij = ⎨ xmax if xij > xmax
j
⎬ (15) Table 1 shows the amount of every load in the IEEE
⎪ j ⎪ 30-bus system.
⎩ xi others ⎭
Table 1 Load settings of the IEEE 30-bus system
So the problem (12) could be transformed to
P Q P Q
min F ( x) = ∑ Pgi − ∑ Plj
Bus Bus
(MW) (MVAr) (MW) (MVAr)
i∈G j∈L 2 21.7 12.7 17 9 5.8
N ueq
3 2.4 1.2 18 3.2 0.9
4 7.6 1.6 19 9.5 3.4
+ wh ∑ max(0, h j ( x)) (16) 7 22.8 10.9 20 2.2 0.7
j =1 8 30 30 21 17.5 11.2
+ ∑ wlj ( Plj − Plj' ) 10 5.8 2 23 3.2 1.6
12 11.2 7.5 24 8.7 6.7
j∈L
14 6.2 1.6 26 3.5 2.3
It can be seen that the aim function consists of three 15 8.2 2.5 29 2.4 0.9
16 3.5 1.8 30 10.6 1.9
parts. The first part is the power loss; the second
Three different cases were studied. They are normal While the outputs of every generator under this
operating condition, single-line fault condition, and single fault condition are shown in Table 3, Fig.5
severe fault condition. gives the convergence characteristics.

4.1 NORMAL OPERATION CONDITION

In normal operation condition, no load is shedding,


and the power loss is 1.9634MW. The outputs of
every generator are shown in Table 2.
Table 2 Generator outputs under normal operating condition
Generator Bus P (MW) Q (MVAr)
1 7.5032 -20
2 53.3442 12.1268
22 31.3837 22.6370
27 41.3806 26.3202
23 17.5521 7.7025
13 40.0000 36.9050
Fig.5 Convergence curve under single-line fault condition
Fig.4 shows the convergence characteristics in semi-
logarithmic scale. 4.3 SEVERE FAULT

Now, a severe fault is imposed to this system by


tripping the transmission line from bus 6 to bus 8 as
well as tripping three generators connected to bus 22,
27 and 13.

The result of this condition is that 18.9595 MW,


1.7922 MW, and 4.6641 MW load will be shed at
bus 8, 26 and 30, respectively with power loss of
5.4320 MW.

The outputs of every generator under this severe


fault condition are shown in Table 4.
Fig.4 Convergence curve under normal operating condition
Table 4 Generator outputs under severe fault condition

4.2 SINGLE-LINE FAULT Generator Bus P (MW) Q (MVAr)


1 69.3881 4.2385
2 79.9961 60.0000
Now an artificial fault would be introduced to cause 23 19.6983 29.7622
load shedding. Since most of the energy supply to
bus 8 is transmitted by the transmission line from Fig.6 gives the convergence characteristics under
bus 6 to 8, line 6-8 is an important line which is this severe fault condition.
fully loaded in normal operation condition.

Suppose that a faulty is applied in line 6-8 and


caused it to be tripped. The result of this condition is
that 7.3726MW load will be shedding at bus 8, and
the power loss is 2.2734MW. It shall be noted that
the amount of power loss is of little importance as
compared to load shedding.
Table 3 Generator outputs under single-line fault condition
Generator Bus P (MW) Q (MVAr)
1 10.5965 -6.8001
2 45.8882 11.8837
22 36.0384 19.2129
27 35.2383 25.0658
23 16.3393 7.1972
Fig.6 Convergence curve under severe fault condition
13 40.0000 39.0586
5. DISCUSSIONS [3] S. Granville, F. C. Mello, and A.C. G. MELLO,
“Application of interior point methods to
In this new strategy, both the power loss and the power flow unsolvability,” IEEE Trans. Power
load shedding would be considered and coordinated Syst.,vol. 11, no. 2, pp. 1096–1103, May 1996.
together during the generation optimization. So this
model not only could be used in emergency [4] M. A. Mostafa, M. E. EL-Hawary, G. A.
condition but also normal operation condition. N.Mbamalu, M. M. Mansour, K. M. EL-Nagar,
and A. N. EL-Arabaty, “Steady-state load
On the other hand, there is no limitation in where the shedding schemes: A performance comparison,”
load could be shedding and how much could be Elect. Power Syst. Res. 38, pp. ,105–112, 1996
shedding. The location and the amount are
calculated by this algorithm without limitations. [5] Nedic D P, Dobson I, Kirschen D S, et. al..
Criticality in a cascading failure blackout model.
The gradient information is not used in this strategy Proceedings of the 15th power systems
because the core part to solve the optimization Computation Conference, Liege Belgium, Aug
problem is PSO method. But the disadvantage of 2005
PSO comes into this new strategy too. That is
calculating speed. The PSO method calculates [6] Kennedy, J., and Eberhart, R.: ‘Particle swarm
slower than method based on gradient. Fortunately, optimization’. Proc. IEEE Int. Conf. Neural
PSO method suits well in parallel computing Netw., 1995, vol. 4, pp. 1942–1948.
environment for fast computation. But further
research is needed for the implement of parallel [7] Kennedy, J., and Eberhart, R.: ‘Swarm
computing. intelligence’ (Morgan Kaufmann Publishers,
2001).

6. CONCLUSIONS [8] Eberhart, R., and Kennedy, J.: ‘A new optimizer


using particle swarm theory’. Proc. 6th Int.
A load shedding method based on PSO method is Symp. on Micro Machine and Human Science,
proposed in this paper. The load shedding is Nagoya, IEEE Service Center, October 1995,
considered as a penalty part of the aim function. The pp. 39–43.
masking effect caused by the setting of penalty
function makes this model effective in both normal
operation condition and emergency condition. The
effectiveness of this new method is verified by the
simulation results of the IEEE 30-bus system.
Further research on the use of parallel processing to
speed up the overall execution time will be needed
for practical use of the proposed approach on large-
scale power systems.

7. ACKNOWLEDGMENTS

The authors gratefully acknowledge the support of


the Hong Kong Polytechnic University under
Project 4-ZV28.

8. REFERENCES

[1] D. K. Subramanian, “Optimum load-shedding


through programming techniques,” IEEE Trans.
Power App. Syst., vol. PAS-90, pp. 89–95,
1971

[2] S. M. Chan and F. C. Schweppe, “A generation


reallocation and load shedding algorithm,”
IEEE Trans. Power App. Syst., vol. PAS-90, pp.
26–34, 1979.

You might also like