Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract— The formulation and computation of an optimal hydrogen, and others could be harvested more effectively.
load shedding time algorithm is presented. This new concept Smaller facilities could be developed faster and with less
was developed as an enhancement to the application of optimal regulatory intervention depending on the technology. Such ini-
load shedding for corrective control to support islanded power
systems. The methodology combines nonlinear mathematical pro- tiatives could spark local interest into cleaner, environmentally
gramming and discretized differential-algebraic power systems friendly energy. Issues like location, interconnection, stability,
equations to estimate the optimal amount of load to be shed as protection, and security amongst others are being actively
well as the best time to shed it. Several simulated scenarios are evaluated [1]–[6]. Some of the work on these areas is yielding
studied and results presented. promising results.
Index Terms— Distributed Generation, load shedding, opti- At the heart of the technical controversy lies the issue
mization. of islanding. When generation equipment is located within
traditionally radial distribution systems, unintentional islands
N OTATIONS may result. These could have fatal consequences to linesmen
x state variables vector and others who might not be aware of the danger. However,
y algebraic variables vector under certain conditions, the same techniques used to detect
u control variables vector and eliminate unintentional islands might end up as potential
nx number of state variables sources of instability [7]. Under the current guidelines (IEEE
ny number of algebraic variables P1547 [8]), generators connected to the power system are
nT number of time steps required to detect an islanded situation (through voltage or
Hk discretized transition equation frequency excursions or other means [9], [10]) and disconnect.
C total cost of current iteration As a consequence of this, most islanding results in a loss of
C1 cost associated with the state variables supply. While this is clearly appropriate in the current situation
C2 cost associated with the algebraic variables where DG is isolated and penetration is low, it is really not
C3 cost associated with the control input tenable in a possible future situation in which the overall
Q weights matrix power system relies on a high DG penetration. In this scenario
L augmented cost function reliability considerations start to become dominant, and there
γ lagrangian multiplier, state constraints will be substantial motivation to examine strategies that are
β lagrangian multiplier, network constraints based on sustaining islanded operation to gain one of the major
PLSi load shedding fraction at bus i benefits from distributed generation. This contribution seeks
t1 load shedding time, measured after to examine a new methodology for optimal load shedding in
formation of island anticipation of this future scenario. However, it is recognized
that there are also issues related to system ownership and
I. I NTRODUCTION the point of common coupling which go beyond the strictly
nificant enhancement to the optimization technique developed method. Further improvements have come in the form of
in [14]. The new procedure yields the optimal load shedding increased detail in the power system model. The approach
time as well as the load shedding amount. It also uses a global reported in [20] takes into consideration machine governor
approach to drive a disturbed system towards a known stable droop and the frequency and voltage characteristics of loads.
operating point. New control variables and goal functions have The latest methods consider load flow restrictions as well as
also been implemented in this formulation. All algorithms system dynamics before and after a load shedding operation
assume centralized load shedding as well as a communication has taken place. The desired effect is to capture the dynamic
infrastructure to be available [15], [16]. Also, all observers response of each load shedding operation and to detect the
and switching operations are assumed to be ideal. best load shedding option to minimize an objective function
based on system quantities as well as other variables. Dynamic
methods are based on the discretization of the system’s state
II. P REVIOUS WORK
trajectories. After discretization, the dynamic state equations
The use of load shedding as a mitigation tool has been are treated as algebraic constraints in the optimization prob-
constantly evolving and different approaches have been for- lem. These trajectories are squared and integrated to estimate
mulated. Relaying schemes like under-frequency (UF) and the “cost” of an iteration and to obtain the next load shedding
rate of change of frequency (ROCOF) are some examples values that should further reduce the cost function. This is
of the mechanisms implemented to trigger a load shedding done until convergence is achieved.
event. Typical UF schemes trigger load shedding based on Application of a load shedding action as fast as physically
pre-configured frequency levels. Two or more shedding stages possible [21] and at arbitrary, predetermined time steps [22]
may exist depending on the utility philosophy. When the have been reported. The optimal load shedding time, to the
system frequency reaches a specified threshold value, a time best of the authors’ knowledge, has never been suggested
delay is inserted prior to the shedding action to try to avoid or used before. The current trend is to apply all corrective
over shedding and assist the coordination of the next stage. measures as soon as possible or delayed only for the sake
This technique is not without inherent complications. Too of event discrimination. However, simulation results show
few frequency levels could lead to over-shedding, but, on the that increased damping and enhanced recovery trajectories are
other hand, time delays between stages could add up and may observed when the corrective action is applied at the optimal
not allow for enough load to be shed in time to reestablish time.
nominal frequency. The implementation of ROCOF techniques
mitigates some of these problems. It permits better estimation III. PARTITIONED E XPLICIT (PE) M ETHOD
of the load-generation unbalance paving the way for schemes
that could adapt the relay settings to a specific event. Cutoff In the Partitioned Explicit (PE) method the set of algebraic
frequencies as well as the amount of load to be shed have typ- equations (2) are decoupled from the transition function H, and
ically been determined based on an operator’s experience and all system algebraic equations are now treated as a separate
simplified power system models. Newer adaptive techniques group of constraints in the optimization problem.
have come to light to fulfill more complicated and challenging Consider a power system represented in the compact form,
modern demands on power systems. These techniques are also ẋ = f (x, y) (1)
constantly changing to accommodate newer technologies and
0 = g(x, y, u) (2)
operating philosophies [17], [18].
The optimal load shedding problem formulation has also Defining H as the transition equation, in its continuous form
been evolving. An early version stated the problem as a it must satisfy,
minimum load shedding problem [19]. In this approach, the H = ẋ − f (x, y) = 0 (3)
goal was to find the minimum amount of load to shed while
satisfying load flow equations and static constraints like line Applying the trapezoidal rule to (3) yields the discretized
flows, voltage, angular limits and shedding constraints. The transition equation, Hk ,
formulation has several advantages over conventional load ∆t
Hk = xk+1 − xk − f (xk+1 , y k+1 ) + f (xk , y k ) (4)
flow. One of its strongest points is that it provides an opti- 2
mization frame for distributing the slack between generation The cost function is given by:
and load throughout the available nodes. This is particularly
useful, especially after disturbances or equipment outages, C(x, y, u) = C1 (x) + C2 (y) + C3 (u) (5)
when corrective action, other than rescheduling, might be
where
necessary. Transmission limits could take the form of transfer
C1 (x) = (x − x∗ )T Q∆t(x − x∗ )
limits or angle differences.
Load shedding constraints provide a suitable framework C2 (y) = (y − y ∗ )T Q∆t(y − y ∗ )
for restricting the maximum amount of load to be shed at
2
a specific site. Critical loads like hospitals or prime customers C3 (u) = PLS (tf − t1 )
can benefit from this capability. The weight imposed on each The optimization problem is:
load shedding bus could also be adjusted to reflect a specific
customer hierarchy, tariff, or any other customer differentiation minimize C(x, y, u)
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. X, NO. XX, MONTH 200X 3
subject to Start
Hk (x, y) = 0
g(x, y, u) = 0 u=0
Using Lagrangian multipliers, the augmented cost function 0 = g(x 0 , y 0 )
is given by: 0 = f(x 0 , y 0 )
L = C1 (x) + C2 (y) + C3 (u) + γ T Hk (x, y) + β T g(x, y, u)
First Order Necessary Conditions (FONC) : u 0 = [t1 PLS ]
dL dC1 dH dg
= + γT + βT =0 (6)
dx dx dx dx
0 = g(x(t c ), y(t +c ))
dL dC2 dH dg
= + γT + βT =0 (7)
dy dy dy dy
Continue
dL dC3 dg Simuation
= + βT =0 (8)
du du du No t = t1
dL
= Hk (x, y) = 0 (9)
dγ Yes
∂L
dL u k +1
= u +α
k
0 = g(x(t1 ), y(t1+ ))
= g(x, y, u) = 0 (10) ∂u
dβ
∂L ∂L
A. PE Algorithm 0= =
The mathematical problem at hand has several aspects. ∂x ∂y
First, the overall problem is best described as a constrained, Find β
nonlinear mathematical programming problem. The goal is to
minimize system voltage and frequency oscillations following
∂L ∂C3 ∂g
a disturbance with an optimal corrective load shedding action. = + βT
This problem requires an iterative solution. Each iteration of ∂u ∂u ∂u
No
the optimization problem controls all subsequent stages of the
algorithm. Second, the nonlinearities of the system trigger the
∂L
use of discretization techniques to estimate the dynamics of ≤ε
the system. These discretized trajectories are later used, each ∂u
time-step, as equality constraints for the overall optimization
Yes
routine to estimate the cost of the current iteration. Third,
the time domain simulations require the simultaneous solution
of state equations and algebraic network interface equations. Return
Transmission grid as well as load characteristics and system
voltages are all part of the algebraic set of variables that Fig. 1. PE method conceptual diagram.
describe the network.
All the elements previously described are found within the
algorithms developed to solve the optimization problem. The then disturbed and the condition of the system following the
well-known lambda search gradient method [23] has been formation of the island is saved and subsequently used as
implemented and acts as the main engine of the algorithm. the initial condition for the time domain simulations of the
response to every input. The input uk contains two quantities,
B. Proposed solution the load shedding time t1 and the vector PLS . The simulation
The goal of the algorithm is to iteratively find the input that progresses until t = t1 at which time the load shedding occurs.
minimizes (5). The cost computation requires the time domain Once the load has been shed, the simulation continues to its
simulation of the islanded section of the grid from the time end.
immediately after clearing the fault until a final time. This final Once the simulation has concluded, the lambda search
time must be long enough to capture all the dynamics but also, gradient method calls for the calculation of the Lagrangian
as short as possible to reduce the order of the system. multipliers, particularly β. The linear system given by (6) and
Fig. 1 shows a conceptual flowchart of the PE method. (7) must be solved. Given the size of the problem it is very
Before iterating to find the optimal input, several actions advisable to exploit sparsity. Some of the partial derivatives
must take place. A stable initial condition [xo yo ] is extracted involved are shown below.
from a steady state solution to (1) and (2). The system is Partial derivatives with respect to x are given by:
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. X, NO. XX, MONTH 200X 4
dg1
0 ... 0 A single-line diagram for a test power system is shown
dx1
in Fig. 2. The scenario simulated with this system is repre-
0 dg2
. . . 0 sentative of a simple, grid-connected, DG system. Before the
dx2
dg disturbance, the utility and the DG system are tied together.
= (13)
dx .. .. .. The DG system is assumed to be importing some power
. . . from the utility. A three-phase fault occurs in the tie line
and after the fault is cleared a DG island is formed. Prior
dg
0 0 . . . dxnT to the disturbance, the island was importing power and, after
nT
it, excess load exists.
Similarly with respect to y: During the disturbance the system voltage drops, load
dg1 decreases and generators start accelerating. Once the fault is
dy 0 ... 0
1 cleared, an island is formed, its voltage starts to recover and
excess load exists. The main idea behind the optimal load
0 dg2
... 0
dy2 shedding time is to make use of this excess load to quickly
dg
= (14) slow down the generators, and optimally shed the excess load
dy .. .. ..
. . . once the island approaches synchronous speed.
Several one-dimensional and multi-dimensional examples
dgnT have been completed. Simplified conditions facilitate the use
0 0 ...
dynT
of two and three-dimensional plots to illustrate the various
After evaluating (12) and (13) the resulting matrices are not stages of the problem. Their outlining assumptions and pre-
necessarily square, the number of dynamic states, nx , may liminary results are presented below.
IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. X, NO. XX, MONTH 200X 5
0.9
A. Reduced dimension visualization
0.8
One-dimensional examples, despite their lack of practical
application, are useful to enhance visualization of the problem.
0.9
0.84
0.85
0.83
0
0.8
0.82
0.75
0.81
Cost
0.7 Cost
0.8
0
0.65
0.79
0.6 1
0.78 2 3
0.55
0.77
0.5
231
0.76
0.45
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.75
Fractional Load Shedding 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Load Sheding Time,t1(s)
presented in Fig. 5.
Fig. 6. Single bus optimized load shedding time and amount.
1.5
1
momentum to decelerate beyond synchronous speed and the
cost
0.5 smaller, “follower”, machine at bus #2 has reached syn-
0
0 2 4 6 8 10 12 14 16 18 20
chronous speed and it is just about ready to start accelerating
1 again. The algorithm detects that shedding prior to this point
will reduce the braking torque and prolong the oscillation.
t1
0.5
Furthermore, shedding after this point would not significantly
0
1
0 2 4 6 8 10 12 14 16 18 20 alter the trajectory of the dominant machine, sending it to an
under-frequency condition likely to prolong the oscillations.
1
PLS
0.5
Therefore, it was found that 0.19 seconds was the optimal
0
0 2 4 6 8 10 12 14 16 18 20
time in this instance.
1
0.5
The methodology developed represents a significant en-
0
0 2 4 6 8 10
Iteration
12 14 16 18 20 hancement to the disciplines of corrective control and islanding
for distributed power systems. Compared to early optimal
Fig. 7. Two busses optimized load shedding time and amount. load shedding approaches, the algorithm incorporates the most
significant missing feature, dynamic trajectories of the system.
machine was 82 seconds. With respect to newer methods, the optimal load shedding time
translates into more efficient and effective use of a shedding
1.02 scheme.
1.01
The optimization formulation is both flexible and robust. It
Speed(pu)
optimized
1
can handle a variety of conditions implemented either through
0.99
the goal function or the power system model. With modifica-
tions, the algorithms could also handle a variety of scenarios
0.98
−1 0 1 2 3 4 5 6 7
like multiple islands or an optimal load shedding schedule,
where more than one shedding stage could be determined.
The incorporation of semiconductor inverters (required for
1
asynchronous generators and most renewable energy sources)
0.8
creates a very different set of circumstances and, is a priority.
Voltage(pu)
0.6 optimized
Given the close ties between the dominant machine and the
0.4
optimal time, it is likely that heavy penetration of very fast
0.2
devices will reduce the optimal shedding time to a point at
0
−1 0 1 2
Time(s)
3 4 5 6 7 which it becomes ineffective.