Professional Documents
Culture Documents
fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
1
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
2
verification of its normality assumption using actual data for For grid vulnerability studies, two characteristics of
generation and load and proposed uncertainty model in [21]; uncertainty coming from different sources are important; its
b) Development of a mixed OPF-stochastic model to analyze size and dynamics. The probability density function (PDF) of
the impact of uncertainties introduced by renewables and a random variable is a common statistic used to evaluate the
simulate grid cascading failure evolution; c) Incorporation of probability of different values of uncertainty as shown in Fig.
thermal stability model and island detection algorithm for line 1. It is usually of interest to fit an approximate PDF to the
outage simulation and power rebalance to simulate operators empirical distribution when characterizing forecasting error
corrective actions, and d) Inclusion of time-space correlation [22].
of line flows uncertainty in the modeling of line tripping pro- The dynamics of uncertainty is related to their variation
cess. over time and contains useful information about their nature.
The rest of the paper is organized as follows. Section II Based on proposed model by [20], frequency components of
presents the model that describes the uncertainty of injection injected power in buses which will translate into frequency
power from generators and smart grid loads. Section III components of line flows have a direct impact on grid
analyzes the statistical distribution of dynamic line flows and vulnerability to random disturbances. Therefore, to identify
verifies their normality assumption in a power grid network the dynamic behavior of uncertainty coming from each source,
which are derived from input power and grid network their occupied bandwidth (OBW) is calculated using the
equations. Section IV proposes a mixed OPF-Stochastic model power spectrum of the uncertainty signal ().
of cascading failure based on [12], [20]. Section V provides The power spectrum () of a time series () describes
numerical simulation results on the IEEE 300-bus system and the distribution of frequency components composing that
Section VI gives conclusion and discusses future works. signal as shown in Fig. 1. The power spectral density (PSD) of
a signal refers to the spectral energy distribution per unit of
II. WIND UNCERTAINTY MODELING time. The spectrum of physical processes often contains
Uncertainties coming from different sources such as essential information about the nature of them. One particular
renewable generation and loads show different characteristics information that can be useful for our purpose is the OBW of
in terms of magnitude and frequency of occurrence. We the signal. It can clearly represent the dynamics of signal that
proposed an uncertainty model in [21] that represents the in our case shows the dynamics of uncertainty for different
injection power from each component (i.e. generator output sources. The mathematical representation of PSD can be
power and load demand power) with two terms as shown in expressed by (2):
(1).
() = |()|2 = () = () (2)
() = () + ()
where () is the time-varying mean of the power signal or where () is the Fourier transform of the signal and () is
in other words it is what we expect to have for each the autocorrelation function which describes the correlation
component ahead of time and () which is a zero mean between values of the process at different times, as a function
signal, representing the uncertainty that may come from of the two times or of the time lag. In this study, the 99%
forecast error or mismatch in output power for conventional OBW is considered as a metric to show the dynamics of the
generators. Note that in this study, the output power of uncertainty. The OBW is the bandwidth containing 99% of the
generators including conventional and wind generation and total integrated power in the spectrum. Using real data for
demand power from loads are modeled with the above generation including wind and conventional generators and
representation. Fig. 1 shows this representation for output loads from Electricity Reliability Council of Texas (ERCOT)
power signal of a wind generator as an example. power system with 4 second resolution and forecasting using
Autoregressive Moving Average (ARMA) approach, it is
found that wind generation injects the highest amount of
uncertainty into the grid in terms of its bandwidth and
magnitude [21]. Table I shows the 99% OBW, Root Mean
Squared Error (RMSE), and Mean Absolute Percentage Error
(MAPE) for each uncertainty source.
TABLE I
UNCERTAINTY STATISTICS FOR DIFFERENT SOURCES
Uncertainty OBW (Hz) RMSE MAPE (%)
Load 61.84 0.20 4.16
Co-generation 72.45 0.94 6.22
Fossil fuel generation 63.03 1.05 5.91
Wind generation 117.24 8.09 10.32
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
3
III. NORMALITY VERIFICATION FOR LINE FLOWS that the lines with larger coefficients will not frequently get
In our line outage model presented in the next section, the tripped and dont have much impact on our stochastic model.
line overload distance is calculated based on the Gaussian
assumption for line flow uncertainty as proposed by the
authors in [20]. Authors in [20] assume Gaussian or normal
distribution for the line flows without accounting for wind
generation uncertainty in the network. Therefore, it is of
interest to verify this assumption for the line flows of a grid
with high penetration of wind energy with the uncertainty
model proposed earlier. This enables us to calculate the
overload distance for the line flows which will be later used in
the line tripping model presented in section IV. A. The IEEE
300 bus system with several added wind generators is selected
to implement the proposed uncertainty model and investigate
the normality assumption for the line flows. This system is a
synthesized network from New England power system and has
a topology with 300 buses and 411 transmission lines. Fig. 2. The IEEE 300 bus system line flow uncertainty normality test results.
The initial observations on the line flows uncertainty
TABLE II
distributions reveals a close proximity to normal distribution. SUMMARY OF NORMALITY VERIFICATION FOR THE IEEE
There is numerous approach for normality test in the literature 300 BUS SYSTEM LINE FLOWS
each suitable for different needs. Kurtosis and Skewness Total Generation Load
coefficients are used widely for normality test for large Connection
Count connected connected
samples as many other tests such as Kolmogorov-Smirnov Normal lines 349 49 (14%) 199 (57%) 101 (29%)
test, Jarque-Bera test, and Shapiro-Wilk test almost always Out of normal
62 37 (59%) 20 (33%) 5 (8%)
bound lines
reject the null hypothesis of normal distribution for large Total 411 86 (21%) 238 (58%) 87 (21%)
samples. This is because when the number of samples gets
larger, even the smallest deviation from perfect normality will IV. CASCADING FAILURE SIMULATION
lead to a significant result. Kurtosis is a measure of whether
the data are heavy-tailed or light-tailed relative to normal This study aims to evaluate the grid vulnerability to black-
distribution. Skewness is a measure of symmetry, or more outs from uncertainty viewpoint. We propose a mixed OPF-
precisely, the lack of symmetry. A perfect normal distribution stochastic approach that simulates the uncertainty in the line
would have both kurtosis and skewness coefficients equal to flow process while simultaneously calculates the flow re-
zero. However, in [19] and [20] the 2 range for these dispatch based on the DC power flow. Note that unlike the
measures are introduced as the acceptable range in order to OPA model proposed by Carreras et al. in [12], in our study
prove normal univariate distribution. We will use this range in the OPF is only used to determine the initial dispatch of the
our verification step. conventional generators based on forecast profiles for loads
The absolute value for both measures are calculated for and wind generation and then the flow re-dispatch is calculat-
flows of all 411 lines in the IEEE 300 bus system and then ed based on a DC power flow and power balance algorithm for
sorted in ascending order. Fig. 2 shows the results for the line the newly formed islands. This helps simulate the CF process
flows uncertainty kurtosis and skewness. Kurtosis coefficients more accurately since in real time there is not enough time for
are sorted in ascending order and the corresponding skewness OPF analyses during escalation phase of the failures. The
coefficients are superimposed in the figure to consider both thermal stability model for overhead transmission lines is
measures simultaneously. It is found that 95% of the lines combined with stochastic model presented by [20] to simulate
have kurtosis and skewness coefficients less than 5 with 85% the most probable path of line outages under high penetration
(349 lines) within acceptable range of 2. This means that of wind uncertainty. Furthermore, an automatic power balance
majority of lines satisfy the normality criteria with a trivial algorithm combined with an online island detection algorithm
deviation from a perfect normal distribution. Table II are introduced to simulate the corrective actions taken by the
summarizes the results of normality test for grid lines based on system operators to alleviate operating contingencies. The
kurtosis measure. It is also found that 59% of the lines that are flow chart of the simulation procedure is depicted in Fig. 3.
outside of the acceptable bound (with coefficient larger than 2) A. DC power flow and line overload modeling
are connected directly to generation buses as shown in Table Determining the steady-state operating conditions of the
II. Our statistical analysis on line capacities for the actual grid power grid requires solving the full power flow equations that
data from Federal Energy Regulatory Commission (FERC) provide information about the voltage magnitudes and phases
suggests that these lines tend to have larger loading margin and the active and reactive power flows through each trans-
compared to other lines. Hence, their overload probability is mission line. Unfortunately, solving repeatedly the full nonlin-
relatively low. It implies that we can use Gaussian assumption ear power flow equations becomes computationally prohibi-
for the line flows in our line outage model considering the fact tive due to numerous solutions during the evolution of CF. In
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
4
Calculate for the overloaded Detect new islands formed by time t and = ( (), ( )) . In the simulation
line (eq. 14) the tripped line (island detec-
tion algorithm) we utilize the unbiased estimates as
1
((), ()) = +/2 ( )
( ) for the
1 =/2
No sample j of flow uncertainty. Note that sliding window method
> ? Trip overloaded line after
seconds (Fig. 4) for covariance calculation selects a predetermined number (N)
of observations of line flow errors and then puts them in the
Yes
observation matrix. Each element of covariance matrix ( ) is
Fig. 3. Flow chart of the simulation procedure of cascading failure based on calculated based on the observation matrix. The variance for
proposed mixed OPF-stochastic model. flow process of each line can then be calculated by taking
The power flow equations for a power grid with n nodes square root of each diagonal element in covariance matrix
and m links can be expressed as:
with () = , (, 0).
() = ()()
With a Gaussian assumption for the distribution of () in
() = ( ())()
[20], the overloading probability () = {| ()| > }
where () represents the vector of injected real power, () can be calculated using Q-function as below:
the nodal voltage angles, and () the flows on the lines. The
() ( )
matrix () is defined as
(5) ()
() = ( ()) where =
()
is the normalized overload distance of
2
where () = 1/ is the line admittance; ( ()) the lth line and Q-function as () = 2 /(2).
represents a diagonal matrix with entries of { (), = In an actual power grid, the overload status of the line
1,2, , }. (, ) is the line-node incidence matrix, flows can be expected to have both temporal and spatial
arbitrarily oriented and defined as: , = 1; , = -1, if the th correlation with each other. Most lines stay safely below their
line is from node to node and , = 0, , . Note that capacities under normal operating conditions; while during the
we assume that all the bus voltages in the grid network are escalation phase of CFs, some of the lines may become over-
closed to the rated voltage levels. loaded and get tripped within a very short time. Using the co-
The standard version of the DC OPF introduced by [26] is variance matrix (, 0), we define Mahalanobis overload
solved to find the initial dispatch of the generators. The distance [27] as
problem formulation is based on total cost minimization and is
as follows: = ( ()) (, 0)( ())
min ( ) which can be viewed as the overload distance of all lines in
,
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
5
the network as a whole. Here (. ) represents pseudo inverse overloaded lines may not be overloaded anymore. Therefore,
by taking the reciprocal of each non-zero singular values on an update of line states after every line trip in the relay
the diagonal, leaving the zeros in place, and transposing the tripping time is necessary. The time to trip of each overloaded
resulting matrix. On the other hand, if the line flows are transmission line is determined according to (14) and then
independent of each other, the covariance matrix (, 0) will after every line outage, the time to trip for other overloaded
become diagonal and the network Mahalanobis overload lines are updated. Note that updating stage for trip time
distance be equal to the Euclidian overload distance, considers the overload duration for each relay from the first
= 2 . Under normal operating conditions we have overloading instant. In other words, our relay model is with
. We will analyze both overload distances for grid memory, since the overloaded lines are already heated up due
lines during CFs under increased uncertainty coming from to excess power flowing through them and the new time to trip
wind generation in section V. accounts for the gained heat. This concept is illustrated in Fig.
Finally, using the normalized overload distance ( ) and 4.
overloading probability ( ) for each line we can calculate the Fig. 4 (a) shows the power flow of three different
mean overload time for flow process () as follows [20]: transmission lines each overloaded initially. The time to trip
for each relay is shown in Fig. 4 (b). According to tripping
2 /2
2 mechanism explained earlier, after the first trip ( = ) all
= other relays need to update their timers. With memory effect
assumption for relay operation, the new time to trip for relay C
where is the equivalent bandwidth of the flow process for is 1 . Note that this updated time to trip is smaller than that
the lth line and can be calculated using the spectral power
of memoryless operation (1 ). Also note that for relay B, the
density (SPD) of the flow process discussed earlier [21].
overload flow at = is larger than initial overload flow at
B. Tripping mechanism and relay model = 0, hence the new time to trip is smaller while the overload
The trip time of thermal overload relays is determined flow for relay C at = is smaller than initial overload flow
based on the maximum allowable current flowing in the at = 0 which means larger time to trip. However, for both
conductor without causing thermal instability. Generally, the cases with memory effect for relay operation, the new time to
overload relays for HV transmission lines have time- trip equals the time to trip calculated by (14) minus the total
dependent tripping characteristic, which is determined using time duration from the first overload instant till the updating
the well-known dynamic thermal balance between heat gains time.
and losses in the conductor [28]. The maximum or hot spot
temperature determines the time to trip for thermal relays and
considering initial operation current and applying necessary
changes, the time to trip can be calculated using [29]:
2 2
= . ln ( )
2 2 (a)
where is overloaded line flow (p.u.), is initial operating
flow (p.u.), is the line flow threshold, and is thermal
time constant which is related to conductor type and
environmental parameters such as wind speed and ambient
temperature [30]. In this study, it is assumed that all
transmission lines use typical HAWK (477 kcmil) ACSR
conductor with =450 sec.
In this study, for the tripping mechanism, both relay time to
trip and overloading probability are considered simultaneously
to select the most probable line trip during the escalation
phase of CF. At every time step, first the time to trip for each
overloaded line is calculated, then using normalized overload
(b)
distance ( ) and overloading probability ( ) the mean
overload time ( ) is determined. If relay time to trip is larger
than the mean overload time, the trip timer is set to zero,
otherwise the trip timer is set to the relay time to trip. This
tripping mechanism enables us to model the stochastic process
of CF and identify the most probable path for its propagation.
After every line trip, the topology of the grid changes and
Fig. 4. Memory effect in relay time to trip calculation: a) line flows for 3
so as the flow distribution across the grid network. Therefore, different lines, b) relay time to trip for the corresponding lines.
some new lines may become overloaded and some of the
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
6
C. Island detection and power balance ) and generators location. Assuming the resistance for all
Successive line tripping during escalation phase of CF lines in the network equal to 1 and using the Kirchhoffs
usually causes the formation of several islands in the power current and voltage laws, the equations describing system
network. The electrical frequency of the system is driven behavior are solved. To detect buses belonging to each island,
based on the power balance according to the well-known generators are activated (assuming output current of 1 A) one
electro-mechanical equation [31] at a time. After identifying all present separate islands in the
grid, their power balance is maintained by shedding actions.
= . 2 For detailed island detection algorithm please refer to [31].
Collect generation and load
where is the total produced power, is the total settings and bus type vector
consumed power, is the global inertia, and is the electrical for new and mother islands
energized and algorithm will cut the total load of the island. No
= 2?
Otherwise, a comparison is made between the total load and =
Set = . Yes
the total generation and maximum available generation to
balance the power accordingly to minimize the total load loss.
Note that generation or load adjustment is distributed meaning = 2?
Yes
End
that balancing adjustment is applied as percentage to every
No
generator or load.
An automatic island detection algorithm inspired by the Fig. 5. Flow chart of the power balance algorithm for newly formed islands in
the power grid.
approach proposed in [31] is used after each trip to identify
newly formed islands. Clusters of generator(s) and load(s) that
are not connected to the grid are called island(s). The V. RESULTS & DISCUSSIONS
algorithm consists of three steps; connectivity check, critical Two different scenarios are considered to study the impact
events identification, and island identification. Connectivity of wind uncertainty on grid vulnerability and for each of them,
check determines how many islands are present in the power some of the conventional generators in the original IEEE 300
system, and their structure. Critical event detection identifies bus system are replaced with wind farms. In the first scenario
which breakers should create an island if opened. And the 11 conventional generators are replaced with wind generators
final step, island detection, identifies the buses belonging to at buses 80,88,125,128,156,199,222,256,258,262,295. The
each possible island and calculates their load balance. The load and generation data are received from Electric Reliability
actual dispatch of the network is not required for the algorithm Council of Texas (ERCOT) with 4-second sampling rate. The
since it only depends on gird topology (grid incidence matrix high sampling rate for data allows us to capture high-
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
7
frequency dynamics of different sources. These data are for uncertainty of wind power and is the initial uncertainty.
numerous wind generators, load demand, and different types To see the impact of larger forecasting errors on grid
of conventional generations. In this study, fossil fuel genera- vulnerability to overload CFs, is increased from 1 to 5 with
tors and co-generation are used as conventional generation and 0.25 steps to find the uncertainty level in which the first CF
load data are scaled according to the original settings of the occurs. All other settings of the system remains the same
IEEE 300 bus system. Uncertainty modeling of loads and gen- during first scenario. Table IV shows the results for increased
erations are based on the model proposed in [21]. The Auto- wind uncertainty level.
regressive Moving Average (ARMA) forecasting technique For between (1-2.25) there is no tripped line thus no CF
for 30 minutes horizon is employed to model the initial uncer- happens for this uncertainty range. Moving beyond =2.25
tainty signal coming from wind generation and electrical loads multiple line overloads are observed that leads to a series of
based on the actual data. The increased uncertainty level mod- CFs that forms multiple islands and isolated buses. Automatic
els the use of different forecasting techniques and horizons power balance on each island causes the load to be dropped to
with different accuracy and characteristics and illustratively a certain level that can be supplied by generators inside the
shows how the accuracy of the forecasting method affects the island. Successive line trips continue until all line flows drop
results. Simulations of the CF scenarios are performed in the safely below line thresholds. Also, the first and second tripped
MATLAB environment and MATPOWER is used for OPF lines and their respective times are given to identify the most
and PF calculations [26]. Wind installation settings are shown vulnerable lines in the network for a given wind uncertainty.
in Table III for the two studied scenarios. Note that wind gen- is increased further to see the impacts of even larger
eration capacity is selected the same as the IEEE-300 bus sys- uncertainty levels on severity of CF. Fig. 6 shows the total
tem original setting for conventional generators. number of tripped lines and total load shedding percent for
TABLE III different wind uncertainty levels. As the increases, the more
WIND GENERATION LOCATIONS FOR SIMULATION lines get tripped during CF which in turn leads into formation
SCENARIOS of more islands and larger load shedding as shown in Fig. 6.
Penetration TABLE IV
Wind generation location (bus number)
level CASCADING FAILURE RESULTS FOR THE FIRST SCENARIO:
First ={80,88,125,128,156,199,222,256,258, WIND UNCERTAINTY LEVEL
= 0.036
scenario 262,295}
Total # of Total First Second
= 0.036 = LS a
trip formed LS a tripped tripped
= 0.05 ={ ,69} count islands (MW)
(%)
line b line b
= 0.09 ={ ,69,131,169} 2.00 0 0 0 0 - -
Second = 0.105 ={ ,69,131,169,254} 2.25 0 0 0 0 - -
scenario = 0.125 ={ ,69,131,169,254,260} 2.50 68 18 7941 32 365 @ 54.7 205 @ 55
= 0.150 ={ ,69,131,169,254,260,215} 2.75 70 19 8344 33.6 99 @ 44.8 205 @ 44.9
= 0.173 ={ ,69,131,169,254,260,215,248} 3.00 71 19 8496 34.2 117 @ 44.8 207 @ 44.9
3.25 72 20 8966 36.1 365 @ 44.6 205 @ 45
= 0.223 ={ ,69,131,169,254,260,215,248,122,255}
3.50 74 21 9135 36.8 117 @ 30.4 205 @ 30.5
3.75 75 21 9397 37.8 115 @ 29.2 205 @ 29.3
The initial operating equilibrium and conditions 4.00 80 22 9749 39.2 365 @ 29.2 205 @ 29.4
((0), (0), (0), (0)) are taken or derived from the power 4.25 83 23 10005 40.3 365 @ 28.2 205 @ 28.3
flow solution. The equivalent bandwidth of flow process for 4.50 83 23 10092 40.6 205 @ 24.9 117 @ 25.1
4.75 83 24 10495 42.3 365 @ 15.8 205 @ 16.1
each line under the initial uncertainty level is then calculated 5.00 86 25 10665 43 99 @ 15.6 205 @ 15.7
and stored to use later on stochastic tripping mechanism. Since a
Load Shedding
the original setting of the IEEE 300 bus system doesnt b
Line number @ time (min)
provide enough information on line capacities, they are set as
= {|(0)|, 2.0(. . )} with =1.20. Here we take
(0) as the rational flow distribution under normal operating
conditions and assume that the line capacity allows a load
increase up to 20% [20]. Note that we select a near congestion
operating conditions for the grid to better see the impact of
increased uncertainty from wind generation on multiple line
overloads leading to CFs. The minimum of line capacity is set
to be 2.0 p.u. so that the vibration in the lines which usually
carry small flows will not cause frequent line trips.
A. Wind uncertainty level and grid vulnerability
The first scenario is considered to study the impacts of
forecasting relative error which comes into the picture in the Fig. 6. The total number of line trips and total load shedding versus uncertain-
ty level ().
form of uncertainty from wind generation. For this scenario,
the uncertainty signal magnitude for wind generator is The evolution process of CFs for different wind uncertainty
level is shown in Fig. 7. All the curves are comparable to
increased by factor = where is the new actual failures recorded in history and reported in [19]. Each
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
8
evolution curve consists of two phases, the escalation phase in lines and load shedding for each penetration level . It is
which the line trip rate is as high as 12 lines per minute, and found that the higher the wind penetration ratio, the more line
the damping phase with line trip rate of approximately one trip and load shedding occurs in the network. In other words,
line per minute. Also, it is found that as wind uncertainty level given the same settings for all other generations and loads of
increases, the first trip happens earlier than lower wind the network, under congestion conditions in the network,
uncertainty level which indicates that the minimum safety installing more wind farms increases the risk of blackout due
time of the entire network decreases under the same operating to cascading overload failures.
conditions. For example, the black bold line shows the TABLE V
cumulative number of line trips for uncertainty level increased CASCADING FAILURE RESULTS FOR THE SECOND SCENARIO:
by the factor 5. As compared with uncertainty level increased WIND PENETRATION LEVEL.
by factor 2.5 (green line with square marker), the former Total # of Total First Second
results into higher number of tripped lines due to high level of LS a
trip formed LS a
(%)
tripped tripped
wind uncertainty. Also, high uncertainty level causes count islands (MW) line b line b
contingency in multiple lines earlier compared to lower 0.036 0 0 0 0 - -
0.05 0 0 0 0 - -
uncertainty levels. For example, the earliest cascading process 0.09 68 20 8149 33 137 @ 2 101 @ 3.8
is associated with the highest uncertainty level, =5, as 0.105 72 21 10156 41.1 137 @ 2 274 @ 3.8
indicated in Table IV and happens after 15 minutes of 0.125 75 23 10701 43.3 137 @ 2 83 @ 2.7
0.15 81 24 10890 44 137 @ 2 83 @ 2.5
beginning of simulation, which implies that as more 0.173 94 27 11650 47.2 137 @ 2 83 @ 2.4
uncertainty is injected to the grid, its survival time gets 0.223 98 30 13958 56.5 137 @ 2 274 @ 2.3
shorter. a
Load Shedding
b
Line number @ time (min)
B. Wind penetration level and grid vulnerability
It is also observed that for all CFs beyond =0.09, line 137
The second scenario aims to investigate the impacts of is the first line getting tripped and it happens almost at the
increased penetration level of wind energy on grid same time for all above 0.09. This could be explained
vulnerability to cascading overload failures. For this scenario, considering the location of the new wind farm installation. A
,
wind penetration ratio is defined as = , where certain wind farm added to the network at a particular location
,
injects additional uncertainty to one of the backbone
, is the total wind generator capacity and , is transmission lines in the network leading to further line trips
the total grid generation capacity. By replacing more and propagation of CFs. However, the second line trip is
conventional generators from original setting of the network different for each penetration level which determines the
with wind generators in addition to those already installed, cascading path and eventually the total number of trips and
is increased to see the impacts of higher wind penetration on load shedding. This is particularly interesting for planning
grid vulnerability. Note that small to medium generators are purposes, since this will detect the most vulnerable lines of the
selected to be replaced with additional wind farms to have network.
smaller steps for . All other settings of the system remains
the same.
Fig. 8. The total number of line trips and total load shedding versus wind
penetration level ().
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
9
Fig. 10. Mahalanobis and Euclidian overload distance during normal and CF
in the IEEE 300 bus system.
VI. CONCLUSION
In this study, we analyzed the impacts of wind generation
on grid vulnerability to cascading overload failures in terms of
Fig. 9. The evolution process of CFs for different wind penetration levels. its penetration and uncertainty level. We first verified the
normality assumption for line flow uncertainty based on actual
C. Overload distance analysis during cascading failure generation and load data and then using a mixed DC OPF-
As mentioned earlier in section IV, the overload distance of stochastic model and thermal stability model for overload
line flows can be a good indicator of the behavior of the power lines simulated the cascading failure in power grids with high
grid during normal operation and CF. The Mahalanobis penetration of wind generation. In addition, the automatic
overload distance ( ) shows the overload distance of the power balance and island detection algorithm are incorporated
whole network and considers the correlation between line to consider corrective actions during propagation of failures.
flows using flow covariance matrix, while the Euclidian Two scenarios are considered to study the impact of wind
overload distance ( ) assumes no correlation among the line generation uncertainty on CFs. First, it is found that increased
flows and considers them as independent random variables. uncertainty injected from wind generation could cause
Fig. 10 shows both and during normal operation of the cascading failures in the grid and the higher the injected
system (without increase in wind uncertainty) and CF uncertainty the more severe the situation in terms of the total
resulting from increased wind uncertainty by the factor =4.5. number of tripped lines and load shedding. Second, our
From Fig. 10 we can see that Mahalanobis overload dis- analyses show that given the current operating condition of the
tance ( ) is always greater than Euclidian overload distance grid, increasing wind penetration to a certain level may result
( ) either for normal operation or during the CF. This veri- into cascading overload failures and higher penetration makes
fies the fact that in an actual power network, the line flows the grid more vulnerable to failures. In addition, overload
have a strong correlation with each other that causes a more distance of the network as a measure of grid safety to CF is
robust and reliable operation compared to independent line analyzed for normal and contingency operating conditions.
flows. The first line trip happens at =25 min. From =0 to Simulation results suggest that appropriate management of
first tripping instant is smaller for the case leading to CF. uncertainties via energy storage or advanced forecasting
This is because of increased wind uncertainty compared to techniques is necessary in order to achieve sustained growth of
normal operation. In other words, increased uncertainty in line renewable generation in current grid operation. In the future,
flows results into smaller overload distance and consequently we may consider more detailed and accurate models in our
higher chance for CF in the power system. Another interesting study of interdependence of cascading failures, renewable
finding is that both and show an increasing trend as generation and smart grid loads, such as full AC power flow in
line tripping spreads throughout the grid. This is because of the grid modeling and Geographical Information System (GIS)
the load shedding actions to maintain power balance for newly that indicates the potential location of renewable integration
and additional types of contingencies other than overload
formed islands in the network. As more generation units are
separated from formed islands, larger portions of electrical failures.
loads get curtailed which means reduction in line flows and
according to (8) the overall distance of line flows from their REFERENCES
threshold increases. This also explains why all CFs tend to [1] W. Weisheng et al., On the Road to Wind Power: Chinas Experience at
Managing Disturbances with High Penetrations of Wind Generation,
stop after several line trips if appropriate load shedding mech- IEEE Power Energy Mag., vol. 14, no. 6, pp. 2434, Nov. 2016.
anisms were employed in the operation and control of the grid. [2] P. Henneaux, P.-E. Labeau, and J.-C. Maun, Blackout Probabilistic Risk
Assessment and Thermal Effects: Impacts of Changes in Generation,
IEEE Trans. Power Syst., vol. 28, no. 4, pp. 47224731, Nov. 2013.
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSTE.2017.2718518, IEEE
Transactions on Sustainable Energy
10
[3] Jianan Mu, Hongbin Sun, Qinglai Guo, Wenchuan Wu, Fengda Xu, and for Power Systems Research and Education, IEEE Trans. Power Syst.,
Boming Zhang, Design of an online intelligent alarming system for vol. 26, no. 1, pp. 1219, Feb. 2011.
cascading failures of group of wind farms, in 2013 IEEE Power & [27] MAHALANOBIS and P. C., On the generalized distance in statistics,
Energy Society General Meeting, 2013, pp. 15. Proc. Natl. Inst. Sci., vol. 2, pp. 4955, 1936.
[4] Xi Ye, Ying Qiao, and Zongxiang Lu, Cascading tripping out of [28] J. Carneiro and L. Ferrarini, Analysis and design of overload protections
numerous wind turbines in China: Fault evolution analysis and simulation for HV lines with a probabilistic approach, in 10th International
study, in 2012 IEEE Power and Energy Society General Meeting, 2012, Conference on Probabilistic Methods Applied to Power Systems PMAPS ,
pp. 111. 2008.
[5] H. Khazaei and Y. Zhao, Ex-post Stable and Fair Payoff Allocation for [29] S. E. Zocholl and G. Benmouyal, On the Protection of Thermal
Renewable Energy Aggregation, in 2017 IEEE Power & Energy Society Processes, IEEE Trans. Power Deliv., vol. 20, no. 2, pp. 12401246,
Innovative Smart Grid Technologies Conference (ISGT), 2017, pp. 1-5. Apr. 2005.
[6] Y. Zhao and H. Khazaei, An incentive compatible profit allocation [30] R. Thrash, Overhead Conductor Manual 2nd Edition. Southwire
mechanism for renewable energy aggregation, in 2016 IEEE Power and Company, 2007.
Energy Society General Meeting (PESGM), 2016, pp. 15. [31] P. Pinceti and M. Vanti, An Algorithm for the Automatic Detection of
[7] A. Scala, S. Pahwa, and C. Scoglio, Cascade failures and distributed Islanded Areas Inside an Active Network, IEEE Trans. Smart Grid, vol.
generation in power grids, Int. J. Crit. Infrastructures, vol. 11, no. 1, pp. 6, no. 6, pp. 30203028, Nov. 2015.
2735, 2015.
[8] A. E. Motter and Y.-C. Lai, Cascade-based attacks on complex
networks, Phys. Rev. E, vol. 66, no. 6, p. 65102, Dec. 2002. Mir Hadi Athari (sM14) received the B.S. degree
[9] L. Zhao, K. Park, and Y.-C. Lai, Attack vulnerability of scale-free from University of Tabriz in 2012 and M.S. degree
networks due to cascading breakdown, Phys. Rev. E, vol. 70, no. 3, p. from Amirkabir University of Technology (Tehran
35101, Sep. 2004. Polytechnic) in 2014, both in electrical engineering.
[10] I. Dobson, B. A. Carreras, and D. E. Newman, A loading dependent He is currently pursuing his Ph.D. degree in the
model of probabilistic cascading failure, Probab. Eng. Informational Department of Electrical and Computer Engineering
Sci., vol. 19, no. 1, pp. 1532, Jan. 2005. at Virginia Commonwealth University, Richmond,
[11] I. Dobson, J. Kim, and K. R. Wierzbicki, Testing Branching Process VA. His research interests include electric power
Estimators of Cascading Failure with Data from a Simulation of system operation and control, vulnerability analysis,
Transmission Line Outages, Risk Anal., vol. 30, no. 4, pp. 650662, Apr. renewable integration, and smart grids.
2010.
[12] B. A. Carreras, V. E. Lynch, I. Dobson, and D. E. Newman, Critical
Zhifang Wang (M02, SM12) received the B.S. and
points and transitions in an electric power transmission model for
the M.S. degree in 1995 and 1998 respectively from
cascading failure blackouts, Chaos An Interdiscip. J. Nonlinear Sci., vol.
EE Tsinghua University Beijing, China and the Ph.D.
12, no. 4, pp. 985994, Dec. 2002.
degree in 2005 from ECE Cornel University.
[13] Hui Ren, I. Dobson, and B. A. Carreras, Long-Term Effect of the n-1
She is an Assistant Professor in the Department
Criterion on Cascading Line Outages in an Evolving Power Transmission
of Electrical and Computer Engineering at Virginia
Grid, IEEE Trans. Power Syst., vol. 23, no. 3, pp. 12171225, Aug.
Commonwealth University, Richmond, Virginia,
2008.
since August 2012. Her research interests include
[14] Shengwei Mei, Fei He, Xuemin Zhang, Shengyu Wu, and Gang Wang,
electric power system modeling and optimization,
An Improved OPA Model and Blackout Risk Assessment, IEEE Trans.
vulnerability analysis and renewable integration.
Power Syst., vol. 24, no. 2, pp. 814823, May 2009.
[15] M. A. Rios, D. S. Kirschen, D. Jayaweera, D. P. Nedic, and R. N. Allan,
Value of security: modeling time-dependent phenomena and weather
conditions, IEEE Trans. Power Syst., vol. 17, no. 3, pp. 543548, Aug.
2002.
[16] M. Rahnamay-Naeini, Z. Wang, N. Ghani, A. Mammoli, and M. M.
Hayat, Stochastic Analysis of Cascading-Failure Dynamics in Power
Grids, IEEE Trans. Power Syst., vol. 29, no. 4, pp. 17671779, Jul.
2014.
[17] M. Rahnamay-Naeini, Zhuoyao Wang, A. Mammoli, and M. M. Hayat,
A probabilistic model for the dynamics of cascading failures and
blackouts in power grids, in 2012 IEEE Power and Energy Society
General Meeting, 2012, pp. 18.
[18] M. Anghel, K. Werley, and A. Motter, Stochastic Model for Power Grid
Dynamics, in 2007 40th Annual Hawaii International Conference on
System Sciences (HICSS07), 2007, pp. 113113.
[19] I. Dobson, B. Carreras, and D. Newman, Branching Process Models for
the Exponentially Increasing Portions of Cascading Failure Blackouts.,
in Hawaii International Conference On System Sciences HICSS, 2005.
[20] Z. Wang, A. Scaglione, and R. J. Thomas, A Markov-Transition Model
for Cascading Failures in Power Grids, in 2012 45th Hawaii
International Conference on System Sciences, 2012, pp. 21152124.
[21] M. H. Athari and Z. Wang, Modeling the uncertainties in renewable
generation and smart grid loads for the study of the grid vulnerability, in
2016 IEEE Power & Energy Society Innovative Smart Grid Technologies
Conference (ISGT), 2016, pp. 15.
[22] H. Bludszuweit, J. A. Dominguez-Navarro, and A. Llombart, Statistical
Analysis of Wind Power Forecast Error, IEEE Trans. Power Syst., vol.
23, no. 3, pp. 983991, Aug. 2008.
[23] A. Field, Discovering statistics using SPSS. Sage publications, 2009.
[24] D. George, SPSS for windows step by step: A simple study guide and
reference, 17.0 update, 10/e. Pearson Education India, 2003.
[25] A. J. Wood and B. F. Wollenberg, Power system generation, operation
and control, John Wiley, New York, 1984.
[26] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas,
MATPOWER: Steady-State Operations, Planning, and Analysis Tools
1949-3029 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.