Electric Power Systems Research 126 (2015) 21–28

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Predictive underfrequency load shedding scheme for islanded power

systems with renewable generation
Urban Rudez ∗ , Rafael Mihalic 1
University of Ljubljana, Faculty of Electrical Engineering, Trzaska 25, 1000, Ljubljana, Slovenia

a r t i c l e i n f o a b s t r a c t

Article history: Underfrequency load shedding is one of the most important protection mechanisms in a power system.
Received 4 December 2014 However, in the majority of power systems it has remained unchanged for decades, despite the advances
Received in revised form 23 April 2015 in computer and communication technologies. Wide Area Monitoring System proved to be very useful
Accepted 28 April 2015
and reliable in the last few years and this is why much effort is given into development of Wide Area
Available online 16 May 2015
Monitoring, Protection and Control concepts. Adaptive underfrequency load shedding is certainly a very
suitable candidate for the task. In order to accelerate changes in this area, a proposal for an adaptive
Keywords:
scheme has been developed, which is a good option from both the technical and economic points of view.
Adaptive control
Frequency response
In contrast to majority of adaptive schemes, it handles problems of active-power deficit estimation and
Power system islanding its variations during the transient due to voltage dependent loads differently, i.e., by on-line forecasting
Stochastic approach the operating point trajectory in a phase space. The scheme has been tested on previously proven (by
Underfrequency load shedding comparison to WAMS measurements) dynamic model of a part of a Slovenian power system and the
System protection schemes results indicate a large improvement compared to the traditional scheme.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction even though a certain deviation exists) is with satisfactory accuracy

Electric power system (EPS) protection is playing a vital role
in ensuring the safe operation of vast and complex EPSs. It is of
vital importance to provide all the EPS’s consumers with reliable
and high-quality electrical energy, despite the disturbances that
constantly occur in the EPS. Hence, a power-system blackout should
be avoided at all costs, both from economic and technical points of
view.
The EPS’s frequency is often considered as a parameter that
never deviates much from its nominal value (e.g., first-swing stabil-
ity analysis). However, one has to keep in mind that the frequency
of the sinusoidal voltage, produced on the generators’ terminals,
determines the term that is known as EPS’s frequency.
Consequently, after/during larger disturbances in the EPS, these
generators might oscillate between one another, and this is
reflected in greater deviations of the local frequencies from the
global one [1]. The only indicator of the global frequency in such
cases is the frequency of the Centre of Inertia (COI). This was
acknowledged and confirmed in many cases [1–4]. However, a
system-wide frequency (referred to in the paper as COI frequency,
∗ Corresponding author. Tel.: +386 1 4768 418.
E-mail addresses: urban.rudez@fe.uni-lj.si (U. Rudez), rafael.mihalic@fe.uni-lj.si
(R. Mihalic).
1
Tel.: +386 1 4768 438.
estimated by simply averaging several (those available) avail-
able PMU frequencies. This can be explained by imagining the
high-voltage network busses somewhere near the electrical centre
between generating units, each of them connected to a high-voltage
network by a step-up transformer whose impedance far exceeds
transmission line impedances.
A sudden unexpected islanding of some part of an EPS is
undoubtedly considered as a major disturbance. There is a high
possibility that the formation of a power-system island is accom-
panied by a generation/consumption power imbalance (generally
an imbalance of the active as well as the reactive power). The island
condition might appear with different power-imbalances, system
might become a subject to either deficit or surplus of power. If
the active-power deficit is in question, the COI frequency in the
island starts to decrease. Due to the limitations introduced by syn-
chronous generators, which in ENTSO-E interconnection should
not operate at frequencies below 47.5 Hz [5], underfrequency load-
shedding (UFLS) protection might be required to regain the power
balance by disconnecting certain amounts of the power-system
load. An overview of many different types of UFLS protection
approaches is systematically provided in [6].
In the majority of EPSs around the globe, UFLS has remained
unchanged for decades, despite the advances in computer and
communication technologies in recent years. Even so, many pro-
posals can be found in the literature for constructing an adaptive,

http://dx.doi.org/10.1016/j.epsr.2015.04.017
0378-7796/© 2015 Elsevier B.V. All rights reserved.
22 U. Rudez, R. Mihalic / Electric Power Systems Research 126 (2015) 21–28
efficient, secure and simple UFLS scheme that uses the advantage of
the state-of-the-art technology. The Wide Area Monitoring System
(WAMS) proved to be very useful, especially in off-line analysis,
and this is why much effort is given into the development of Wide
Area Monitoring, Protection and Control (WAMPAC) concepts [7].
Adaptive UFLS is one of the most suitable candidates for the task.
The majority of adaptive UFLS schemes are based on the system
frequency-response model [8], which enables a calculation of the
active-power deficit in the system by measuring the initial value of
the frequency’s first time derivative (FTD) [3,9–13]. However, the
power-system load is, in general, affected by the voltage deviations
and, consequently, the calculation of a correct value for the active-
power deficit is rather difficult to obtain [14]. The importance of
considering voltage changes in the formation of UFLS scheme was
also noted in [15] where a suggestion was made how to com-
bine both the voltage and frequency stability issues, rather than
treat them separately. Similarly, but on the local level, a combi-
national approach to UFLS is tested in [16]. Nevertheless, in [17]
a mechanism is presented that is able to successfully identify the
parameters of the system frequency-response model on-line. This
gives a basis for a good estimation of the active-power deficit. How-
ever, an unanswered question remains as to what is the best way to
distribute the calculated deficit into several shedding steps [11,18].
A completely different approach to UFLS is given in [4], which
laid the basis for the so-called predictive UFLS schemes. The fre-
quency’s second time derivative (STD) is used for forecasting the
response of all the regulatory mechanisms that influence the EPS
frequency. By doing so, the frequency trajectory can be on-line fore-
seen in advance which enables the decision-making according to
the severity of the situation. Two issues arise from this approach,
and these are the use of an iterative approximation procedure and
the use of the frequency STD, which might run into difficulties
during the actual implementation of the scheme. Consequently,
a more implementation-friendly idea of using a frequency versus
frequency FTD locus diagram from [19] and [20] is used (referred
to in [19] as a phase plane) in this paper. The iterative procedure is
replaced by a simple analytical calculation, while the results remain
at a very high level.
Instead of using the swing equation and measurement of initial
value of frequency FTD (as it is usually the case in research papers on
this topic), a stochastic analysis is made off-line, prior to applying
the UFLS scheme in the system. By doing so, the probability density
curves of system load (consumption) as well as generation can be
considered. This means that the whole range of different operating
conditions are taken into account and most importantly renew-
able generation (such as wind and solar power plants) can easily
be implemented in the model. As a result, one obtains the probabil-
ity function of active power deficit in the system. However, for the
proposed mechanism, only the maximum deficit value among all
possible conditions is required. Details of how this is accomplished
are given in the paper.
In Section 2 the background of the presented methodology is
described. First, an important, but not vital procedure of an off-line
stochastic analysis of active-power imbalance is presented. Next,
the phase-plane trajectory of a frequency versus frequency FTD is
studied under condition of the active-power deficit. In the same
section the procedure for future frequency situation forecasting is
presented. In the last subsection, the way of selecting the amount
of disconnected load per step is also presented. In Section 3, the
study case and the overview of the testing results are given, and
finally, the conclusions are drawn.
2. Methods
In order to understand the presented UFLS methodology, first
some characteristics of the islanded EPS suffering an active-power
Fig. 1. The probability density function of load active-power at substation Anhovo.
deficit have to be presented and explained. Firstly, the presented
approach takes the advantage of an off-line stochastic analysis of
active-power imbalance in the observed EPS (actual data from real-
istic Slovenian test-case is used, forming a ground-base for results
presented in Section 3.2). Such analysis is expected to be period-
ically performed on a monthly to yearly basis, as according to its
results the parameterization of UFLS scheme is performed. Second,
it is important to understand the typical shape of the curve that cor-
responds to the COI in the frequency versus frequency FTD phase
plane during underfrequency conditions. Purely theoretical IEEE
9-bus test system was used for the purpose of clear and unambigu-
ous explanation of the method’s core philosophy. Tests on the other
hand (shown in Fig. 10) were performed on a (as much as possible)
realistic model of 110 kV part of a Slovenian power system (results
are). The model was constructed during a two year research project
for the Slovenian TSO that required a validation of model simula-
tion results by comparing it to WAMS measurements for numerous
events taking place in the last few years.
2.1. Off-line stochastic analysis of active-power imbalance
The major issue of adaptive UFLS schemes with active power
deficit calculation is the uncertainty of the calculated amount.
Namely, research in the past [9] has shown that this calculation
does not take into account the load voltage dependency, which
might have a substantial influence on the result and might be a root
cause of the system collapse due to inadequate amount of shedding.
This is why it is reasonable to establish a different approach. As the
amount of installed renewable generation is getting larger in the
last years, an approach is needed that takes the stochastic nature
of generation sources into account. In order to make the procedure
as general as possible, the stochastic nature of the consumption is
considered as well.
For the purpose of this paper past measurements of active-
power consumption on several 110 kV substations in the
north-western part of the Slovenian EPS were obtained (total num-
ber of nine consumption busses). One among of the most diverse
(substation Anhovo) is shown in Fig. 1. The obtained probability
density function (PDF) has several local maxima, the most probable
being around 2 MW.
On the other hand, the past measurement data for the gen-
eration capacities were not available. The analysed part of the
Slovenian EPS, which is relatively often required to operate in the
islanded mode due to poor connections, includes ten hydro gen-
eration units. However, in the future plans several wind turbines
are included as well. This is why it was decided that testing should
be done with the assumption of two different PDF distributions for
the generating capacities in order to show that the selection of the
 U. Rudez, R. Mihalic / Electric Power Systems Research 126 (2015) 21–28 23

Fig. 2. The probability density function of total active-power deficit in the part of a Slovenian power system—generation modelled with Weibull (left) and Uniform (right)
distribution.

distribution does not make much difference. First, Weibull distri-
bution and second, uniform distribution is considered for all the
generating capacities in the network.
The calculation of active-power imbalance was performed sev-
eral 1000 times (10,000 to be exact) and it took approximately
45 min, which is not problematic as this is done off-line and on
yearly rather than monthly or even daily basis. Each time a ran-
dom function was employed, which produces a random number
between 0 and 1with equal probability for all real numbers on this
interval to be chosen. However, by combining a random function
with the cumulative PDF of generation units and loads, this is no
longer the case, i.e., load and generation values with higher prob-
ability are more likely to be chosen more often. And so with each
call of the random function, the deterministic generation and load
values are obtained—which are the input for the deterministic load-
flow calculations. This procedure is known as Monte Carlo [21] and
is repeated until the desired number of Monte Carlo calculations is
reached. Finally, by plotting the histogram of the results, graphs on
Fig. 2 are derived, which show (solid thick line) the PDF of active-
power flow on the single 110 kV transmission line, connecting that
part of the network with the rest of the Slovenian EPS (active-power
losses are considered as a part of the consumption as well). It is clear
from results that PDF is almost equal to the Gaussian distribution
(dashed thin line) in both cases.
This PDF curve can be used to recognize the maximum distur-
bance (active-power deficit) that might occur in this part of the
system. For the case under question this equals to approximately
100 MW (see Fig. 2 limit at 100 MW).
source, acting as the rest of the system by supplying the extra power
that the generators are not able to produce themselves. Conse-
quently, the frequency starts to decrease with the initial frequency
FTD which is gradually improved due to already mentioned regu-
latory mechanisms. The point of initial frequency decay is located
on the vertical axis of the phase plane in Fig. 3, as it indicates the
moment just slightly after the deficit occurrence when the fre-
quency is still at its nominal value, while the frequency FTD already
obtains a non-zero value (Fig. 3—point B). In order to regain the bal-
ance between the active-power generation and consumption, the
primary frequency control increases the generated active power,
whereas the voltage control and loads react to voltage changes in
the system (which also affects the imbalance of the active power
[14]). When the balance of the active power is finally obtained,
the frequency is at its lowest value (Fig. 3—point C). When a new
steady-state is achieved (Fig. 3—point D) with increased active-
power production due to primary frequency-control activation, the
frequency is lower than its nominal value (due to a permanent
droop [23]).
2.3. Forecasting the curve in a phase-plane
It was shown in [24] that the COI trajectory in the phase-
plane retains its typical shape of a spiral regardless of the
governor parameters and the EPS conditions. This makes a solid

2.2. Phase-plane trajectory

As already mentioned, the frequency FTD is often used in var-

ious UFLS scheme proposals as a mean of enabling the adaptive
operation of UFLS schemes. In [19] and [20] a specific use of the fre-
quency FTD is employed by means of a frequency versus frequency
FTD phase plane. As in general there are various regulatory mech-
anisms in an EPS that influence the EPS frequency (the primary
frequency control, the voltage control, the self-regulating effect of
the power-system load), the mentioned phase-plane diagram has
the typical shape of a spiral regardless of the situation. This can
be seen from Fig. 3, which corresponds to the islanding of an IEEE
9-bus test-system ([22]).
In the steady-state operation, frequency equals its nominal
value of 50 Hz and the frequency FTD (dfCOI /dt) equals zero
(Fig. 3—point A). The transition into island operation with an active- Fig. 3. A typical frequency versus frequency FTD phase-plane curve during under-
power deficit can be simulated by disconnecting the infinite-power frequency conditions.
24 U. Rudez, R. Mihalic / Electric Power Systems Research 126 (2015) 21–28
Fig. 4. General definition of an ellipse.
ground-base for the proposed mechanism. First it should be deter-
mined which part of the spiral is relevant for UFLS purposes. As the
UFLS should be active before the frequency drops below the criti-
cal value (47.5 Hz in ENTSO-E), only part of the curve in the third
quadrant is relevant. This part of the spiral can be suitably repre-
sented by a section of an ellipse with a sufficient level of accuracy.
The general ellipse equation can, according to Fig. 4, be written as:
x2 y2
+ 2 =1
a2 b
2
where, a is called the major radius and b is the minor radius
2
2 (5)
of the ellipse. a and b are one half of the major and minor diam-
eters, respectively. A part of the ellipse in the third quadrant (i.e.,
between T4 and T3 ), suitable for our purposes, is depicted with a
thicker line. An analogy of the ellipse in Fig. 4 with the spiral in
Fig. 3 should be explained for better understanding. By considering
the ellipse in terms of the presented curve in the phase plane (the
horizontal X axis is considered as the frequency and the vertical y
axis as the frequency FTD), T4 represents the moment slightly after
the active-power deficit’s appearance, with the initial frequency
FTD (df/dtt = 0+ ) and the frequency still being at its nominal value.
As the initial frequency FTD can be measured, T4 can be considered
as the first known point on the ellipse:
   the fact that the first few forecasts are the least accurate, regardless
df 
T4 (0, −b) = T4 (xT 4 , yT 4 ) ⇒ T4 0, 
dt t=0+
If we further assume that both the frequency and its FTD are
continuously measured during the frequency fall, another point (a
constantly moving one) on the ellipse is also known, marked as TC
(xTC , yTC ):
 df (t)

TC (xT C , yT C ) ⇒ TC f (t),
dt
where, f(t) represents the current measured value of the COI fre-
quency and df dt
(t) is the current measured value of the COI frequency
FTD. One must keep in mind that this point is travelling along
the trajectory (spiral) towards T3 in the phase-plane with respect
to time t. Nevertheless, as two points on the ellipse are known,
this makes it possible to calculate the major radius a, which in
terms of the phase-plane represents the lowest value that would
be achieved by the frequency if no further actions are undertaken
(e.g., load shedding). This can be calculated by using the following
expression:
 load shedding is completed, the operating’s point path in the phase-
xT2C × yT24 − yT2C × xT24
a=
yT24 − yT2C
Fig. 5. An example of an islanding dynamic simulation with forecasting of the lowest
frequency value.
or in terms of a curve in the phase-plane:
(1)  

 f (t)2 df (t)/dt 
a =  
t=0+
df (t)/dt  − df (t)/dt
t=0+
As an example, the results of a dynamic simulation are presented
in Fig. 5, obtained by islanding the IEEE 9-bus test system with an
active-power deficit of 189 MW, which corresponds to 60% of the
total system load. In the upper graph, the actual COI frequency is
depicted with the black curve and the calculated major radius a (the
lowest frequency value forecast) with the grey curve. It is clear that
the calculation according to Eq. (5) does not begin at the moment of
deficit’s appearance. Instead, it is slightly delayed (approximately
400 ms) in order for the points T4 and TC to be sufficiently far
apart from each other. Otherwise, the inaccuracy of the calculation
increases to an unacceptable level, which can be partly seen from
(2) of the 400 ms time delay. However, this inaccuracy does not have
much influence on the efficiency of the whole UFLS methodology as
the whole load shedding process is distributed among as many load
shedding steps as necessary, according to the seriousness of the sit-
uation (explained in the following sections). In the second graph,
the actual COI frequency FTD is depicted with the black curve and
its maximum (initial) value is shown with the grey curve.
(3)
2.4. Load shedding effect on the phase-plane trajectory
A few modifications have to be made to the forecasting method-
ology in order to apply it to the EPS where the load-shedding takes
place. This is due to the fact that a sudden disconnection of certain
loads in the EPS manifests itself as a step change in phase-plane
trajectory. This means that the presented spiral in Fig. 3 is strongly
influenced by the load shedding. The situation is presented in Fig. 6.
After the load shedding takes place (20% of the total island load
in Fig. 6) the frequency FTD suddenly decreases in its absolute value,
whereas the frequency itself remains unchanged (a sudden transi-
tion of operating point closer to the abscissa). Therefore, after the
plane shifts to another trajectory, which can also be described as
(4) an ellipse, even though with other values for the major and minor
radii. This means that the Eq. (4) after the load shedding should
 U. Rudez, R. Mihalic / Electric Power Systems Research 126 (2015) 21–28
Fig. 6. Load-shedding effect on the phase-plane trajectory.
consider the new point T4 ’, which differs from T4 in the value of
frequency FTD. Namely, T4 corresponds to the ellipse before the
shedding took place, whereas point T4 ’ is defined by a moment
after the step change in the frequency FTD. Consequently, the calcu-
lated major-radius value of a (the lowest frequency value forecast)
before the load shedding is marked as a1 and after the load shed-
ding as a2 . This is graphically represented in Figs. 6 and 7, where
the calculation of a2 after the load shedding follows the expression:

 2
be discussed:
 f (t) df4 /dt 2 − df (t)/dt 2 (f4 )2
a2 = 
(6) • The load should be disconnected as soon as possible after the
2 2
df4 /dt − df (t)/dt
The increasing penetration of converter-based renewable
energy sources into the generation mix increases system sensitiv-
ity to power mismatches (both generation outages as well as load
shedding). Consequential faster frequency changes df/dt appears
not due to large power mismatch but low inertia instead. Setting
of proposed UFLS for extreme shares of converter-based renew-
able energy sources should be done with care, especially avoiding
disconnections of large amounts of load in a single step. This is
why it is important to limit individual shedding steps (see Section
2.5). Also, off-line Monte Carlo calculations (Section 2.1) should be
repeated for any large changes on the generation side, i.e., in case of
installing/uninstalling any larger generation unit or several smaller
Fig. 7. Load-shedding effect on the lowest frequency value forecast.
25
generation units whose common capacity is comparable to that of a
large unit. This is necessary as the load-flow situation in that part of
the network is changed significantly and also is the possible active-
power imbalance. It is of no importance to the procedure whether
installed/uninstalled units are renewable or conventional, as long
as their PDF is modelled appropriately.
With regards to renewables’ frequency protection, since 2009
a directive [25] is in place for all generation plants connected
to the medium voltage grid, no limitations are given for under-
nominal frequency operation down to 47.5 Hz. This makes the
threshold 47.5 Hz the underfrequency protection limit for all gener-
ating plants, including renewables. Such rules are very reasonable,
as possible conflicts with UFLS setting are avoided.
It can be concluded from Fig. 7 that the minimum frequency
forecast gives a valuable information about the future frequency
drop. According to the forecast it is possible to perform efficient
load shedding—described in detail in the following subsection. The
high level of efficiency is set by considering the result from off-line
analysis.
2.5. Determining per-step load-shedding amount
In order to determine the optimal amount of load that should
be disconnected in the islanding process after active-power deficit
occurrence, first the target performance of a UFLS scheme needs to
island’s formation in order to decrease the frequency FTD before
the frequency drops to a hazardous level. In practice, a threshold
of 49.0 Hz is suggested [5].
• The total amount of disconnected load should be as small as pos-
sible, i.e., the number of loads with a disconnected power supply
should be minimized.
• It is preferable to shed required amount of load in a single step.
However, this feature is difficult to accomplish in reality due to
unreliable calculation of required deficit of any UFLS scheme.
Therefore, multiple-stage shedding is necessary.
• A frequency drop below the lowest acceptable limit (47.5 Hz in
ENTSO-E) should be prevented in any given situation the EPS
might be in. Namely, this is threshold for the underfrequency-
protection activation of the generation facilities [5]. If the
frequency is kept above this threshold, the active-power imbal-
ance can be excluded as a reason for the EPS blackout. In order
to increase the flexibility of the approach, this limit can be set to
any desired level, e.g., to accommodate any additional frequency
limitations introduced by e.g., certain auxiliary equipment [26].
According to above statements, the frequency-manoeuvre space
for the UFLS scheme encompasses the frequency band between
49.0 Hz and 47.5 Hz. Of course, the frequency should be brought
back to the nominal value as soon as possible. However, UFLS is
an automatic procedure that has to react promptly in order to stop
the frequency drop in time and therefore it is not sensible to plan
an UFLS scheme to do something that an Automatic Generation
Control (AGC) is supposed to do anyway later on (i.e., increase
the production of conventional units’ active power in the island to
bring the frequency back to nominal value and release the primary
frequency control).
In the proposed methodology, the load shedding itself is divided
into several shedding steps, whose total number and size varies
according to the current state of the system. However, taking
into account the approximate measurement gathering, commu-
nication and circuit-breaker time requirements [27,28], 250 ms
should be enough time between two subsequent load-shedding
steps. The actual shedding can be triggered only by the shedding
26 U. Rudez, R. Mihalic / Electric Power Systems Research 126 (2015) 21–28
Fig. 8. A linear function for the determination of each individual load shed.
impulse (generated in control centre with a constant frequency
1/0.250 s = 4 Hz) being accompanied by the minimum frequency
prediction lower than 47.5 Hz. Each of these two conditions alone
is not sufficient for disconnection taking place. Instead, meeting
both conditions simultaneously is required. It is also reasonable
to agree that the larger load-shedding steps should be performed
prior to smaller ones [10]. For this reason, in the proposed method-
ology, more loads are chosen for disconnection when there is a
lower value of a calculated. Namely, a lower calculated a corre-
sponds to a more serious situation, as lower values of the achieved
minimum frequency are expected (i.e., a higher deficit in the EPS).
Next, the amount of disconnected load in each load-shedding step
should not be set too high, as the forecast accuracy is limited and
over-shedding should be avoided. At this point, it is very important
to note that the maximal amount of active-power deficit in the sys-
tem under question is approximately 100 MW (the value calculated
by an off-line stochastic analysis – see Section 2.1). Considering the
fact that the maximal amount of active-power consumption in this
part of the network equals 418 MW, corresponding percent value
100
of the deficit equals approximately 418 ≈ 25% of the maximum sys-
tem load. For lower values of the system load, the percent value of
the same MW deficit will be of course higher. This means that in
order to avoid over-shedding 25% of the total system load should be
selected as the maximum value of load that can be disconnected in
a single step. However, such amount is scheduled for tripping only
in recognizing most severe frequency conditions. If higher percent-
age was selected, a danger of over-shedding might occur at system
load less than the maximal 418 MW. Fig. 8 shows a linear concept
that is being used within the presented methodology (with a thick
solid black curve) for the determination of the disconnected load
in each individual shedding step.
A linear concept was selected for reasons of simplicity and is
therefore preferred compared to the other possibilities (e.g., expo-
nential) which are much more complicated and do not contribute
significantly to the scheme’s efficiency (this was also the case in [4]).
In cases when certain additional frequency limitations are required
to accommodate the needs of e.g., certain auxiliary equipment, the
slope of the curve in Fig. 8 can be suitably altered, as is indicated
by the two dashed thick black curves.
3. Simulations, results and discussion
3.1. Study case
In this subsection a study case is explained in detail, from which
the advantages of the presented methodology can be recognized. In
the upper graph in Fig. 9 the frequency is depicted for two applied
UFLS schemes: the scheme, presented in this paper with the solid
black curve and the traditional scheme from Slovenian EPS with
Fig. 9. Study case: active-power deficit of 60 MW in the islanded part of a Slovenian
power system.
the dashed black curve (details about the scheme provided in [14]).
In addition, the calculated forecast value of the minimum achieved
frequency is also depicted with a solid grey curve. The second graph
shows the disconnected amount of load in each load-shedding step,
again for both UFLS schemes with the same colours. The third graph
depicts the frequency FTD for both schemes, where the line types
and colours used are the same.
The simulation is studied in detail in the continuation. Some
122 ms after the start of the simulation the network under study
transits into the island operation. Consequently, the COI frequency
starts to decrease with a frequency FTD of −1.926 Hz/s, due to an
active-power deficit of 60 MW or 49.78%. The first forecast cal-
culation of the minimum achieved frequency is obtained after
additional 22 ms and is clearly not accurate (equal to a = 49.11 Hz).
However, as the frequency is decreasing the forecast value drops
below 47.5 Hz at t = 598 ms. As at that moment the frequency is
still above 49.0 Hz, the load shedding does not start until t = 848 ms,
when the obtained forecast equals a = 47.27 Hz. According to Fig. 8
this corresponds to disconnecting 6% of the total system load. After
250 ms the forecast is still below 47.5 Hz and is again equal to
a = 47.28 Hz. As a consequence, again 6% of the load is discon-
nected. After an additional 250 ms the forecast value is a = 47.40 Hz
and 5.2% of the load is disconnected. After this last shed the cal-
culated forecast remains above 47.5 Hz and consequently load
shedding is no longer required. One should note that the frequency
is still decaying; however, the minimum value reached is 47.60 Hz,
exactly 2.647 s after the beginning of the simulation. The total load-
shedding amount is therefore the sum of all three shedding steps:
17.2% of the total system load or 20.73 MW.
Next, the results obtained by using the traditional scheme from
[14] are analysed. The first shedding step of 10% is initiated when
the frequency passes the threshold of 49.0 Hz (712 ms after the
beginning of the simulation). The following two steps have the
same predefined amount of disconnected load (15%); however, they
are performed at the thresholds 48.8 (after additional 126 ms) and
48.4 (after additional 538 ms). Together, this corresponds to 40%
of the total system load, which is 48.21 MW. From the frequency
 U. Rudez, R. Mihalic / Electric Power Systems Research 126 (2015) 21–28
versus time plot it is obvious that the last shed is unnecessary as
the frequency has almost recovered without it. However, the tra-
ditional scheme does not include any mechanism for recognizing
such a situation. Namely, the only criterion for load shedding is
the frequency passing a certain threshold. Therefore, it is common
with traditional schemes that the last shed causes a steep increase
in the frequency, which might even introduce the generator over-
frequency protection activation.
One should also note that by decreasing the number of discon-
nected loads, the under-frequency operation time increases. This
is an unavoidable consequence of selecting the minimization of
disconnected loads as a criteria for the evaluation of the scheme
operation. However, after the automatic procedure of UFLS stops
the frequency from falling below the unacceptable limit, the AGC
can release the primary frequency control and by that bring the fre-
quency to the nominal value. From [5] it can be understood that a
short-term operation somewhere between 47.5 Hz and 50 Hz does
not represent a hazard for the equipment. Also, one should keep
in mind that critical operating conditions give TSO’s special rights
and especially obligations to retain the system integrity at all costs.
So the lower quality of supply in the form of longer-lasting low
frequency conditions is undoubtedly part of a collateral damage.
According to authors view on this matter, the aim of UFLS is not to
stabilize the frequency anyway but its first time derivative instead.
3.2. Results summary
This subsection contains a summary of several simulation
results, obtained by a dynamic simulation of islanding for the north-
western part of the Slovenian EPS, which is rather often subject
to islanding. The used dynamic model includes all the relevant
features that are needed to ensure the appropriate conditions for
testing the UFLS protection: voltage dependent loads, governors
and excitation controllers, transmission lines with ohmic losses
considered and eight-order models of synchronous generators. To
make the model even more representative, its simulations results
were compared to WAMS measurements and show a good match
[29]. The efficiency of the presented methodology is compared
to the traditional UFLS scheme, which is currently used in the
Slovenian EPS [14]. The traditional scheme represents the current
situation, which should be improved by the implementation of a
new UFLS scheme, supported by WAMPAC technology. The sum-
mary of the results is given in the graph—Fig. 10. On the horizontal
axis a simulated active-power deficit of nine individual simulations
is given in MW. On the vertical axis the total load-shedding amount
(again, in MW) for every simulated case is given, respectively for
each of the UFLS schemes applied. With the dashed black curve the
results of the traditional scheme are depicted and with the solid
black curve the results of applying the proposed methodology are
shown. One should note that in all the presented results from Fig. 10
the frequency is kept above 47.5 Hz. For example, the results for the
traditional scheme are not provided for the deficit values of 90 and
100 MW. This means that the frequency dropped below the limit
of 47.5 Hz and it is assumed that in such case the blackout was
unavoidable.
It can be clearly seen from the graph that the traditional scheme
absolutely yields worse results, as in all the simulated cases far
more load is shed than would keep the island in operation. In addi-
tion, a stepwise curve is obtained by using the traditional scheme,
whereas in the case of presented scheme the curve is almost linear
with respect to the simulated deficit. This is a consequence of hav-
ing fixed, predefined load-shedding amounts in all four shedding
steps of the traditional scheme. The fixed amount of the maximum
load shedding of the traditional scheme means that in cases with
simulated deficits above 80 MW, a frequency drop below 47.5 Hz is
not prevented (therefore, as already written, the last result in the
27
Fig. 10. Summary of the results obtained by dynamic simulations of a north-western
part of a Slovenian power system.
case of a traditional scheme is given for a deficit value of 80 MW).
Consequently, traditional schemes are efficient only for a certain
interval of possible deficits.
This is not the case with the presented approach as the total
load-shedding amount is determined by the seriousness of the dis-
turbance. In all the simulated cases this amount is far less than
by applying the traditional approach (as much as 57% less load
is disconnected in the case of a simulated deficit of 60 MW, see
Fig. 10).
4. Conclusions
In the majority of EPSs around the globe the implemented UFLS
protection has not kept up with the development of communica-
tions and computer technologies. WAMPAC is gradually becoming
a real possibility and this is why adaptive proposals should be made
that support a centralized protection. Also, economic interests are
becoming a more and more important driving force in all areas
of modern society. Ensuring the highest level of a secure supply
of electrical energy is one such example. There are much too many
situations where traditional UFLS schemes react inappropriately by
disconnecting more loads than would be necessary. This is usually
linked to high financial costs that could be avoided by implement-
ing an adaptive UFLS methodology.
In this paper an adaptive UFLS approach is presented that sub-
stantially lowers the total load-shedding amount. In addition, an
EPS blackout is prevented in any situation, compared to traditional
schemes which are efficient only in a certain interval of possible
deficits. This approach requires some communication between the
dispatch centre and the underfrequency relays, which in many sys-
tems is already installed and in operation (e.g., as a part of WAMS
system). Also, due to an increasing amount of installed renewable
generation, the presented approach takes the stochastic nature of
generation sources into account. This makes it possible to consider
the intermittent generating capacities such as the solar and wind
power plants.
Using the frequency versus the frequency FTD phase-plane
trajectory of the Centre of Inertia, the forecasting of the future
minimum frequency value is performed by applying a simple alge-
braic equation. Simple enough to make it a strong possibility for the
actual implementation. Based on this forecast, load shedding fol-
lowing a linear concept is possible, which makes use of the whole
available frequency-manoeuvre space. This is achieved by perform-
ing an off-line stochastic analysis of the active-power deficit in the
network under question.
28 U. Rudez, R. Mihalic / Electric Power Systems Research 126 (2015) 21–28
The efficiency of the proposed methodology was verified on
the dynamic model of a north-western part of the Slovenian EPS,
which is relatively often subject to islanding operation, and shows
great improvement compared to the traditional scheme. The total
load-shedding amount in all the simulated cases is far less than
by applying the traditional approach (as much as 57% less load is
disconnected). Re-tuning of traditional UFLS might improve sys-
tem frequency response in certain cases, but will make responses
in several other situations worse. This is due to non-adaptability
of traditional UFLS. On the other hand, this is not the case with
presented predictive UFLS, which exhibits high level of adaptabil-
ity and therefore operates well in the whole range of situations. Of
course, appropriate periodically modified parameterization is still
necessary for a successful operation. The future work on this subject
will include testing of the procedure on Real Time Digital Simulator
(RTDS) applying actual PMU and relay equipment. Hardware-in-
the-loop simulations are proved to be the most effective type of
control and protection mechanism testing at this point in time as
it is the closest thing to testing in real life.
Acknowledgements
This work was supported by a Slovenian Research Agency as
a part of the research program Electric Power Systems, P2-356.
Authors also acknowledge the help from a Slovenian distribution
company Elektro Primorska for providing the needed data.
References
[1] U. Rudez, R. Mihalic, Monitoring the first frequency derivative to improve adap-
tive underfrequency load-shedding schemes, IEEE Trans. Power Syst. 26 (2011)
839–846, http://dx.doi.org/10.1109/TPWRS.2010.2059715
[2] A. Mokari-Bolhasan, H. Seyedi, B. Mohammadi-ivatloo, S. Abapour, S.
Ghasemzadeh, Modified centralized ROCOF based load shedding scheme in
an islanded distribution network, Int. J. Electr. Power Energy Syst. 62 (2014)
806–815, http://dx.doi.org/10.1016/j.ijepes.2014.05.029
[3] A. Ketabi, M. Hajiakbari Fini, An underfrequency load shedding scheme for
islanded microgrids, Int. J. Electr. Power Energy Syst. 62 (2014) 599–607, http://
dx.doi.org/10.1016/j.ijepes.2014.05.018
[4] U. Rudez, R. Mihalic, A novel approach to underfrequency load shedding, Electr.
Power Syst. Res. 81 (2011) 636–643, http://dx.doi.org/10.1016/j.epsr.2010.10.
020
[5] ENTSO-E, Continental Europe operation handbook, appendix 1: load-frequency
control and performance, (n.d.). https://www.entsoe.eu/publications/system-
operations-reports/operation-handbook/Pages/default.aspx, in press
(accessed 23.02.15).
[6] J.A. Laghari, H. Mokhlis, A.H.A. Bakar, H. Mohamad, Application of com-
putational intelligence techniques for load shedding in power systems: a
review, Energy Convers. Manag. 75 (2013) 130–140, http://dx.doi.org/10.1016/
j.enconman.2013.06.010
[7] V. Terzija, G. Valverde, D. Cai, P. Regulski, V. Madani, J. Fitch, et al., Wide-area
monitoring, protection, and control of future electric power networks, Proc.
IEEE 99 (2011) 80–93, http://dx.doi.org/10.1109/JPROC.2010.2060450
[8] P.M. Anderson, M. Mirheydar, A low-order system frequency response model,
IEEE Trans. Power Syst. 5 (1990) 720–729, http://dx.doi.org/10.1109/59.65898
[9] V.V. Terzija, Adaptive underfrequency load shedding based on the magnitude
of the disturbance estimation, IEEE Trans. Power Syst. 21 (2006) 1260–1266,
http://dx.doi.org/10.1109/TPWRS.2006.879315
[10] P.M. Anderson, M. Mirheydar, An adaptive method for setting underfrequency
load shedding relays, IEEE Trans. Power Syst. 7 (1992) 647–655, http://dx.doi.
org/10.1109/59.141770
[11] S.-J. Huang, C.-C. Huang, An adaptive load shedding method with time-based
design for isolated power systems, Int. J. Electr. Power Energy Syst. 22 (2000)
51–58, http://dx.doi.org/10.1016/S0142-0615(99)00035-6
[12] H. You, V. Vittal, Z. Yang, Self-healing in power systems: an approach using
islanding and rate of frequency decline-based load shedding, IEEE Trans. Power
Syst. 18 (2003) 174–181, http://dx.doi.org/10.1109/TPWRS.2002.807111
[13] M.S. Pasand, H. Seyedi, New centralized adaptive under frequency
load shedding algorithms, in: Proceedings of the 2007 Large Engi-
neering Systems Conference On Power Engineering, 2007, pp. 44–48.
doi:10.1109/LESCPE.2007.4437350.
[14] U. Rudez, R. Mihalic, Analysis of underfrequency load shedding using a fre-
quency gradient, IEEE Trans. Power Deliv. 26 (2011) 565–575, http://dx.doi.
org/10.1109/TPWRD.2009.2036356
[15] J. Tang, J. Liu, F. Ponci, A. Monti, Adaptive load shedding based on combined
frequency and voltage stability assessment using synchrophasor measure-
ments, IEEE Trans. Power Syst. 28 (2013) 2035–2047, http://dx.doi.org/10.
1109/TPWRS.2013.2241794
[16] B. Hoseinzadeh, F.M. Faria da Silva, C.L. Bak, Adaptive tuning of frequency
thresholds using voltage drop data in decentralized load shedding, IEEE Trans.
Power Syst. PP (2014) 1–8, http://dx.doi.org/10.1109/TPWRS.2014.2351015
[17] Y. Tofis, L. Hadjidemetriou, E. Kyriakides, An intelligent load shed-
ding mechanism for maintaining frequency stability, in: Proceedings
of the PowerTech POWERTECH 2013 IEEE Grenoble, 2013, pp. 1–5.
doi:10.1109/PTC.2013.6652296.
[18] B. Delfino, S. Massucco, A. Morini, P. Scalera, F. Silvestro, Implementation
and comparison of different under frequency load-shedding schemes, in: Pro-
ceedings of the Power Engineering Society Summer Meet 2001, (1) 2001, pp.
307–312. doi:10.1109/PESS.2001.970031.
[19] V.N. Chuvychin, N.S. Gurov, S.S. Venkata, R.E. Brown, An adaptive approach to
load shedding and spinning reserve control during underfrequency conditions,
IEEE Trans. Power Syst. 11 (1996) 1805–1810, http://dx.doi.org/10.1109/59.
544646
[20] L.J. Shih, W.J. Lee, J.C. Gu, Y.H. Moon, Application of df/dt in power system
protection and its implementation in microcontroller based intelligent load
shedding relay, in: Proceedings of the Industrial and Commercial Power Sys-
tems for, Technical Conference, 1991. Conference Rec. Paper Presented at 1991
Annual Meeting, 1991, pp. 11–17. doi:10.1109/ICPS.1991.153060.
[21] G.J. Anders, Probability Concepts in Electric Power Systems, Wiley, 1990.
[22] Anderson et al., Anderson, Power System Control and Stability, 2nd Edition,
2nd ed. edition, Wiley-IEEE Press, Piscataway, N.J, n.d.
[23] J. Machowski, J.R. Bumby, Power System Dynamics and Stability, John Wiley &
Sons, 1997.
[24] U. Rudez, R. Mihalic, Underfrequency load shedding with forecasting of the
power system’s frequency response, Int. Rev. Electr. Eng. 7 (2012) 3969–3978.
[25] German Association of Energy and Water Industries—BDEW, Technische
Richtlinie Erzeugungsanlagen am Mittelspannungsnetz, (n.d.). http://www.
bdew.de/internet.nsf/id/A2A0475F2FAE8F44C12578300047C92F/$file/BDEW
RL EA-am-MS-Netz Juni 2008 end.pdf, in press (accessed 01.10.14).
[26] Y. Tofis, Y. Yiasemi, E. Kyriakides, A plug and play, approximation-based, selec-
tive load shedding mechanism for the future electrical grid, in: E. Luiijf, P. Hartel
(Eds.), Critical Information Infrastructure Security, Springer International
Publishing, 2013, pp. 74–83, http://link.springer.com/chapter/10.1007/978-3-
319-03964-0 7 (accessed 0.09.14).
[27] ABB, Live tank circuit breakers, buyer’s guide, http://www08.abb.com/
global/scot/scot245.nsf/veritydisplay/433e1e613e67d06ec1257d0400307817/
$file/Buyers%20Guide%20HV%20Live%20Tank%20Circuit%20Breakers%20Ed%
206en.pdf, 2014 (accessed 08.09.14).
[28] IEEE standard for synchrophasor data transfer for power systems,
IEEE Std C371182-2011 Revis. IEEE Std C37118-2005. (2011) 1–53.
doi:10.1109/IEEESTD.2011.6111222.
[29] D. Kopse, U. Rudez, R. Mihalic, Applying a wide-area measurement system to
validate the dynamic model of a part of European power system, Electr. Power
Syst. Res. 119 (2015) 1–10, http://dx.doi.org/10.1016/j.epsr.2014.08.024.