Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

Impact of Renewable Distributed Generation on Power Systems

M. Begovi, A. Pregelj, A. Rohatgi
School of Electrical and Computer Engineering
Georgia Institute of Technology
Atlanta, GA 30332-0250

D. Novosel
ABB T&D Technology Ltd
Baden, Switzerland

Abstract

pollutant emissions [1]. Utility economic benefits also

include loss reduction, avoided costs of energy
production, generation capacity, distribution and
transmission capacity investment deferral, reducing risk
from uncertain fuel prices, green pricing benefits, etc.

The traditional approach in electric power generation is

to have centralized plants distributing electricity through
an extensive transmission & distribution network.
Distributed generation (DG) provides electric power at a
site closer to the customer, eliminating the unnecessary
transmission and distribution costs. In addition, it can
reduce fossil fuel emissions, defer capital cost, reduce
maintenance investments and improve the distribution
feeder voltage conditions. In the case of small residential
photovoltaic (PV) and wind systems, the actual generator
locations and DG penetration level are usually not
apriori known. The following study attempts to calculate
the boundaries of the impact of randomly placed
distributed generators on a distribution feeder. Monte
Carlo simulations are performed, and boundaries for
overall improvements are determined. The study shows
that the knowledge of total penetration of small PV
systems is sufficient to estimate the effects of DG on the
feeder.

Usually, PV systems are designed to operate at unity

power factor, providing only active power to the utility.
This design maximizes the benefits of the individual
customer, since residential customers are usually charged
only for the active power that they draw from the grid.
However, PV systems may be operated at non-unity
power factor, and there are situations when it is
reasonable to allow PV systems to generate limited
amounts of reactive energy. The utility regulations [2]
dictate that PV systems should operate at a power factor
greater than 0.85 (leading or lagging), when output is
greater than 10% of rating, but specially designed systems
that provide reactive compensation or voltage support
may operate outside of this limit with utility approval.

2. Problem description
1. Introduction
This paper investigates the effects of the dispersed
generation (DG) devices onto the electric power
distribution system. The prevailing utility strategy, so far,
is to treat the DG devices as a parasitic source, and to
impose a set of strict rules designed to limit the effects
that those systems may have on the distribution feeder.
The reasoning for these actions is simple. Most US
utilities are currently price regulated, which simply means
that the actual utility's revenues (and profits) are dictated
entirely by sales. DG devices installed at the customer
level reduce sales and lead to decreasing profits.
However, DG systems inherently provide some
benefits to the utility. They may level the load curve,
improve the voltage profile across the feeder, may reduce
the loading level of branches and substation transformers,
and provide environmental benefits by offsetting the

The following paragraph illustrates the effect of DG

penetration on the actual load demand and voltage profile
of the distribution feeder equipped with switched shunt
capacitors. In radial networks, bus voltages decrease as
the distance from the distribution transformer increases,
and may become lower than the minimum voltage
permitted by the utility. Utilities usually combat this
problem by increasing the tap ratio of the distribution
transformer, and/or by switching on the shunt capacitors.
By providing a portion of energy on site, DG systems
reduce branch currents, which in turn leads to reduced
losses and increased voltage throughout the feeder.
We will consider the 69-bus, 8-lateral distribution
feeder, based on the data from [3], with few modifications
concerning the placement of loads. The feeder lateral
layout is shown in Fig. 1; the locations of loads, shunt
capacitors and one possible distribution of PV generators
are also shown.

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

PV generators scattered across the feeder and operating

at unity power factor behave as a negative active load,
effectively decreasing the load active power demand (Fig.
2a - Fig. 4a), and thus improving local voltage conditions
(Fig. 2b - Fig. 4b). The capacity of PV generators
scattered across the feeder as shown in Fig. 1, was scaled
from 0% to 40% of local loads at their respective buses.

Fig. 1. The test system [3].

The lowest voltage for a base case without

compensation is 0.94 at bus 54 for a summer day with
heavy load. Using sensitivity analysis buses 26 and 54
were identified as best places to add reactive shunt
support. The shunts considered are standard discrete
blocks of 0.3MVAr controlled locally using simple
low/high voltage limit regulations, which connect the
necessary number of blocks in order to maintain
controlled voltage between low and high threshold levels.

Fig. 2b shows the voltage at a remote feeder bus for a

typical sunny day with high load demand. Without PV
support, the shunt capacitors are turned on early in the
morning as the load demand increases and voltage drops
below the threshold level. PV generators improve the
voltage profile, but just delay the turn-on of the switched
shunt, since the PV and load profile characteristics are not
closely matched. While the PV generation peaks around
noon, the peak of the load demand occurs at sunset when
PV generation is not capable of maintaining the voltage
above the threshold value (0.95 p.u.). Similar situation is
portrayed in Fig. 3. Without PV support, the capacitors
turn on at 10:00am, but even 10% PV support is enough
to keep the voltage above 0.95. However, since the load
peak occurs after dark, when PV generators are not active,
the voltage profile worsens and the shunt capacitors turn
on. Note that this might have been avoided if the PV
system was equipped with battery storage (expensive
solution), or if the PV and load characteristics were better
matched. Fig. 4 illustrates that situation. Both PV and
load peaks occur around 11:00am. Even a moderate 10%
PV support is quite enough to keep the voltage within the
acceptable limits throughout the day; the shunt capacitor
never turns on.

Day 246: Total active power at bus 53 for different levels of PV penetration

Day 246: Voltage at bus 53 for different levels of PV penetration

300
0.97

0.965

250
Bus voltage [p.u.]

Active power [kW]

0%
10%
20%
30%
40%

200

0.96

0.955

0.95
150
00:00

06:00

12:00
Time of day

18:00

24:00

a) Active power at bus 53

0%
10%
20%
30%
40%
06:00

12:00
Time of day

18:00

b) Voltage at bus 53

Fig. 2. Effects of PV generation - sunny day with high load. PV provides only active power.

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

Day 289: Total active power at bus 53 for different levels of PV penetration

Day 289: Voltage at bus 53 for different levels of PV penetration

280

0.97
0%
10%
20%
30%
40%

0.968
0.966
0.964

240
Bus voltage [p.u.]

Active power [kW]

260

220

200

0.962
0.96
0.958
0.956
0%
10%
20%
30%
40%

0.954

180

0.952
160
00:00

06:00

12:00
Time of day

18:00

0.95
00:00

24:00

06:00

a) Active power at bus 53

12:00
Time of day

18:00

24:00

b) Voltage at bus 53

Fig. 3. Effects of PV generation - typical fall day. PV provides only active power.
Day 26: Total active power at bus 53 for different levels of PV penetration

Day 26: Voltage at bus 53 for different levels of PV penetration

0.966

Active power [kW]

240

0.964
0.962
Bus voltage [p.u.]

260

0%
10%
20%
30%
40%

220

200

180

0.96
0.958
0.956
0.954
0.952

160

0%
10%
20%
30%
40%

0.95
00:00

06:00

12:00
Time of day

18:00

24:00

06:00

a) Active power at bus 53

12:00
Time of day

18:00

24:00

b) Voltage at bus 53

Fig. 4. Effects of PV generation - sunny day with moderate load. PV provides only active power.

3. DG-enhanced feeder operation

The above figures provide only the qualitative analysis
of the effect that DG systems may have on the operation
of the distribution feeder. We now present a systematical
study which attempts to quantify these and several others
aspects of the DG-enhanced feeder operation.

3.1 Load and PV profiles

We use the test system shown in Fig. 1, for which the
hypothetical daily load profile is obtained from the actual
utility data for a Central-US small city. The measured
utility data spans one whole year in 10-minute intervals,
which is also the period considered in this study. It is
assumed that all of the loads at the feeder follow the same
active and reactive load patterns, however, those patterns
are different for active and reactive power consumption.

The PV generation was obtained by simulating an

ideally oriented and tilted fixed 1kW PV system at the
actual Central-US location using PV simulation program
PVGRID 7.1. The following derating coefficients were
used: shading 4%, dust 4%, mismatch 2%, DC losses 2%,
MPPT losses 4% and the inverter efficiency was assumed
to be 93%. This approach allows simple scaling of a
system, by simply multiplying the energy output values
by the nominal power of the actual system. Non-optimally
oriented and tilted systems may be simulated by
multiplying the output of the system by a number smaller
than one. The output of PVGRID is obtained in hourly
intervals and linearly interpolated to get the PV
production in 10-minute intervals. The weather input to
PVGRID was derived from the TMY2 database that
serves as a standard database for weather conditions in the
US [4].

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

Active power (P)

800
600
400
200
0
0

Hours/year

In the case of renewable DG systems, the DC energy

obtained from the renewable source is fed through the
power-conditioning unit (inverter). The majority of
contemporary inverters used in DG systems are current
source inverters (CSI) operating at unity power factor,
and hence the overall power factor of a distribution feeder
enhanced by DG and equipped with such inverters would
be lower. If, however, we assume that voltage source
inverters (VSI) can be utilized instead of CSI, we can
generate reactive power commensurate with the
remaining unused capacity at any given point in time, as
shown in Fig. 5.

the histogram of the annual active energy production of

the 2kW PV system equipped with 2kVA inverter, and
corresponding histograms of the maximum reactive
annual energy output of that inverter for cases when its
reactive power output is limited by Qmax and Qlim, as
explained above. The PV production was obtained using
PVGRID 7.1 and TMY2 database.

Hours/year

3.2 Inverter control strategies

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

1.2

1.4

1.6

1.8

Qmax determined by inverter's rating

2000
1000
0

Hours/year

Q max

0.2

0.4

0.6

0.8

Qmax s.t. inverter's pf>0.85

1000
500
0
0

0.2

0.4

Q lim

0.6

0.8

1
Power

1.2

1.4

1.6

1.8

Fig. 6. Histograms of annual inverter active energy

production and corresponding reactive production,
limited by Qmax and Qlim.
Pi

Fig. 5. The inverter modes of operation.

The active power output of the inverter Pi is

determined by the amount of power that is supplied by
PV modules, and inverter's efficiency. The majority of
DG inverters supply only active power, however, there is
no intrinsic barrier for inverters to supply reactive energy
as well. For a given active power output Pi, the inverters
reactive power output Qi is limited by its nominal
apparent power rating S, i.e. can be anywhere between
zero and Qmax, as shown in Fig. 5. The current regulations
dictate that PV inverters should operate with power factor
above 0.85 [2], but specially designed systems may
operate with lower power factor with utility's approval.
The reactive limit Qlim determined by this power factor
condition is also shown in Fig. 5. In addition, the
inverters reactive output is limited by the maximum
allowable voltage at its bus. The size of the inverter is
dictated by the maximum expected DC power over a
prolonged period of operation. In the case of PV, or wind
DG systems, the supplied DC power is determined by the
available solar or wind energy at that time point. The
actual DC power and inverters active power output will
therefore typically be less than the inverters apparent
power rating S, which provides plenty of opportunities for
the reactive power output. As an example, Fig. 6 shows

We will consider the following inverter control

strategies:
1. Inverter supplies only active power:
Pi = PDC , Qi = 0

(1)

2. Same as 1), but inverter also supplies maximum

available reactive power up to the limit QV determined
by the maximum allowable voltage at its bus:
Pi = PDC , Qi = min (QV , Q max )

(2)

3. Same as 2), but under constraint that its power factor

is above 0.85.
Pi = PDC , Qi = min (QV , Qlim )

(3)

3.3 DG systems placement

The utility usually does not have the apriori knowledge
of the locations and sizes of small distributed DG systems
(usually PV and/or wind systems) across its feeder.
Therefore, it is beneficial for it to know the boundaries
that a specific DG feeder penetration will have on some
specific quantities of interest, such as voltage profile,

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

active and reactive losses, and switching times of shunt

capacitors. The utility would also be interested in
knowing the exact locations that will provide best overall
impact on the feeder, in terms of those quantities.
We will consider four different PV penetrations:
190kW, 380 kW, 570kW and 760kW, which represent
approximately 5%, 10%, 15% and 20% of the total feeder
load. We assume that all PV systems are multiples of the
standard PV building block of 2kW, i.e. for a 760kW
PV penetration, 380 building blocks are scattered
across the feeder. In order to conserve space, only the
results for 760kW feeder penetration are presented; the
results for remaining feeder penetrations are, as expected,
appropriately lower.
Two different random spatial distributions were
considered for determining the PV generator locations:
uniform and Poisson. Uniform distribution tends to spread
PV generators evenly across the feeder. Result is that for
a fixed penetration level, almost all load buses have
relatively small PV systems connected to them. Using the
Poisson distribution with a suitable parameter , we can
force that the same total PV power is spread over fewer
load buses. Therefore, the PV capacity on those buses will
be higher. Using the Poisson distribution, but restricting
the generators to be spread to a maximum of 60% of the
load buses, we are imposing a situation with very few,
relatively large DG systems across the feeder. Fig. 7
shows one possible PV generator placement in those three
cases for the total feeder penetration level of 760kW.

Uniform [kW]

Various distributions of 760kW of PV power across the feeder

100

Poisson [kW]
Poisson (60%) [kW]

We are proposing a different strategy. Instead of using

a complete annual data set or just a couple of typical days,
let us extract the smaller set of points that can actually
represent the behavior of the feeder with PV penetration
throughout the year. The mathematical representation
follows.
Let X be our data set composed of all daytime vectors
xi of active (Pi), reactive (Qi) and PV (PPVi) power
throughout the year, i.e.:

X = x i | x i = Pi , Qi , PPV , i = 1,..., N
i

10

15

20

25

30

35

40

45

50
0
10

15

20

10

15

20

25

(5)

50

100

(4)

Define set D as a smallest set that contains set X. The

objective is to generate set M with minimal number of
elements such that:
M = {y i | y i D, y i represents D accurately}

50
0

the daily and annual variations of the DG performance

have to be accounted for. At first, 500 series of yearly
Monte Carlo simulations have been proposed. Each
simulation should have identical load conditions, but
different PV generator locations and sizes according to
the chosen PV spatial distribution (uniform or Poisson).
The whole procedure should be repeated for three
considered inverter strategies. This is a very computerintensive process, considering that a total of 26,280,000
(6 24 365 500) power flow solutions would have to be
calculated for each spatial PV feeder distribution and
inverter control strategy. Even if we restrict the data set to
include only daytime-hours, more than 14,000,000
simulations would be needed. Alternatively, we could use
just several typical days, but by doing so, important
annual aggregate values would be lost.

30

35

40

45

50

25
30
Bus number

35

40

45

50

100
50
0

Fig. 7. PV generators placements throughout the

feeder for uniform and Poisson distributions.

4. Performance analysis using clustering

techniques

where the statement "accurate representation" means a

suitably chosen mathematical criterion, which will be
defined later in the text. One method to generate set M is
to perform the cluster analysis of the set X. Clustering
algorithms identify natural groupings of data in a large
data set, sorting the data into several groups (clusters),
which can then be processed as single entities. The
conventional clustering algorithms assign each data point
to only one cluster. The m-clustering of a set X is defined
as a partition of X into m sets C1,...,Cm, where the
following conditions are met:
(1) C i , i = 1,..., m
m

( 2)

=X

(6)

i =1

In order to fully determine the actual effects of the

random placement of PV systems throughout the feeder,

(3) C i C j = , i j , i, j = 1,..., m

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

In other words, there cannot be zero-clusters (1),

missed data (2), or overlapping clusters (3). In fuzzy
clustering algorithms, each data point xi belongs to cluster
Cj to some degree specified by its membership grade uij.
The fuzzy clustering of X into m clusters is defined by a
mxN membership matrix U whose elements are
membership grades uij if the following conditions are met:

ij

= 1, i = 1,..., N

(7)
N

(3) 0 < u ij < N , j = 1,..., m

The membership grades close to unity indicate a "high
grade" of membership to the corresponding cluster. It can
be proven that the conventional hard clustering can be
considered as a special case of fuzzy clustering, with
membership grades restricted to take values only from
{0,1}.
In order to separate data into clusters, a proximity
measure between data vectors and between a point and a
cluster needs to be defined:

( )

(8)

where f is a suitably chosen function. One commonly

used proximity measure between two vectors is the
standard Euclidean distance, where f(x)=x. Each cluster
can be represented by a simple vector, called the point
representative of the cluster and denoted as wj. Therefore,
the distance between the vector and a cluster can be
calculated as a distance between the vector and the cluster
point representative. The objective of the clustering
algorithm is then simply to obtain the clusters point
representatives wj and membership matrix U. This is done
by minimizing the cost function of the form
N

J q (W , U ) = u ijq d x i , w j
i =1 j =1

(9)

with respect to W and U, subject to the constrains for

membership grades uij presented above. The parameter q
is called a fuzzifier (q>1) and the matrix W is defined as

W = w1T ,..., w mT

q
ij

i =1

i =1

(10)

, i = 1,..., N , j = 1,..., m

d x i , w j q 1

k =1 d (x i , w k )

j =1

d x i , x j = f (x i ) f x j

1
m

Using standard minimization techniques, we arrive at

the following two equations:
u ij =

(1) u ij : X [0,1], i = 1,..., N , j = 1,..., m

( 2)

N m
N
m

L(W , U ) = u ijq d x i , w j i u ij 1 (11)

i =1 j =1
i =1
j =1

d x i , w j
w j

(12)

= 0, j = 1,..., m

These equations are coupled and cannot, in general,

yield a closed form solution. They are solved iteratively,
by taking an initial guess for the cluster point
representatives matrix W, calculating the membership
matrix U, updating W, and continuing this process until
the termination criterion is reached. In the case of hard
clustering, the equations simplify significantly, since uij is
restricted to only two values (0 or 1). Hard clustering
algorithms, however, tend to have problems with noncompact clusters usually found in real-life data. The main
advantage of fuzzy clustering is that it captures the
imprecision encountered when describing real-life data,
providing more information about the data structure
compared to a non-fuzzy scheme. We have employed a
fuzzy c-means (FCM) clustering algorithm that uses the
squared Euclidean distance as a proximity measure and a
value q=2 for the fuzzification parameter.
The drawback of the fuzzy clustering algorithm is that
it requires the number of clusters to be known in advance.
However, several validity criteria can be used to
determine the optimal number of clusters, once the
clustering is performed [5]. They measure the quality of
clustering by assigning the value to it. The more the data
points are concentrated around the cluster centers, the
cluster structure is better and the criteria values are lower.
The Xie-Beni (XB) index is used for validation of
clustering produced by the fuzzy-c means algorithm when
the Euclidean distance is in use, and can be calculated as
XB =

J2
2
N D min

(13)

where J2 is the cost function defined in (9), N is the

number of data vectors, Dmin is the minimum Euclidean
distance between two cluster point representatives, i.e.

This approach leads us to the Lagrangian in the

following form:

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

Dmin = min wi w j

(14)

i , j =1,...,m
i j

computational effort we are undertaking is in fact

required.

m is the number of clusters and wk is a vector defining the

point representative of the cluster k. Note that the XB
index is a monotonically decreasing function of the
number of clusters m, as m gets close to N, indicating that
in that range the XB index is not a valid fuzzy clustering
validity criterion [6]. However, the optimal value mopt is
much smaller than N, and we will not reach the starting
point of the decreasing tendency. The optimal number of
clusters can be found by examining the plot of XB as a
function of m, which is shown in Fig. 8.
0.07

The Xie-Beni (XB) index

0.06

Fig. 9. The data and cluster center points (projection

onto the PV-P plane).

0.05

0.04

0.03

0.02

0.01

510 25

50

100

200
Number of clusters

300

Fig. 8. The Xie-Beni index as a function of number of

clusters.

The cluster point representatives are also shown sorted

by the load active power in Fig. 11, which shows
significant correlation between P and Q (as expected), but
no correlation between P and PPV, indicating that all the

Cluster point representatives sorted by P

P [p.u.]

2
1.5
1
0.5

20

40

60

80

100

120

140

160

180

200

20

40

60

80

100

120

140

160

180

200

20

40

60

80

100
120
Cluster #

140

160

180

200

3
Q [p.u.]

The point representatives of these 200 clusters form the

set M that will be used for Monte Carlo simulations. The
scatter plots of all data vectors with superimposed cluster
point representatives shown in Fig. 9 and Fig. 10 reveal
that the clustering algorithm has indeed maintained the
intrinsic data structure. Note however that XB index only
validates the goodness of clustering; it does not guarantee
that the results of analysis using the reduced clustered set
will provide an accurate description of the actual feeder
behavior.

Fig. 10. The data and cluster center points (projection

onto the P-Q plane).

2
1
0

1.5
PV [p.u.]

The structure of XB index in Fig. 8 indicates that the

benefit of increasing the number of clusters rapidly
decreases after a certain number of clusters is reached
(1-200), which can be used as a criterion for the selection
of optimal number of clusters. We have chosen the
optimal number of clusters as mopt=200. Since the total
number of data vectors is N=28137, this value is well
below the range where m and N are comparable.

1
0.5
0

Fig. 11. Cluster point representatives (m=200).

In order to validate that the analysis using the reduced

clustered set will in fact provide dependable results, we

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

have simulated a full year of feeder operation for a

particular PV placement distribution and inverter control
strategy using all available daytime-hours data points and
compared the results to the results obtained using only
cluster centers for several clustering scenarios. Table 1
shows the total annual consumption, losses and total
number of shunt capacitor control (on/off) actions for the
DG-equipped feeder using cluster sizes of 10, 20 and
28137 (all data points). All quantities (losses, total
consumption, number of switch shunt control actions) are
expressed as a percentage of quantities calculated without
DG support. The results indicate that reductions in total
cumulative annual feeder energy consumption and losses
obtained by placing DG throughout the feeder can be
estimated using relatively small number of clusters.
However, more accurate clustering is needed to
investigate the effects of DG penetration on the operation
of switched shunt capacitors.
Table 1. Total feeder consumption, losses and
number of shunt capacitor control actions.
Cluster size

10

200

5. Monte Carlo analysis

Five hundred Monte Carlo simulations have been
performed using 200 cluster point representatives for each
considered PV spatial distribution and inverter control
strategy. Fig. 13 and Fig. 14 show the minimum, average
and maximum recorded values for the substation-level
power factor and remote bus voltage duration curves for
two PV spatial distributions, assuming that the inverters
reactive power output is limited by its kVA limit.
Although the range of recorded values for both power
factor and voltage is, as expected, wider in the case of
Poisson distribution, the average values (solid lines) are
almost identical. The improvements, compared to the base
case without DG (bold solid line) are clearly noticeable.
Poisson distribution yields slightly better maximum
voltage improvement, showing that it may be
advantageous to concentrate PV systems on buses where
low voltages are expected. The average values can be
used for simple cost-benefit calculations.
1

28137

P [%]

83.34

82.37

82.88

Q [%]

84.63

83.83

84.27

P [%]
Feeder
consumption Q [%]

92.5

92.46

92.74

46.76

42.23

40.92

37.5

55.4

51.91

Losses

Switchings

[%]

Fig. 12 shows the substation-level power factor

duration curves for the same three clustering scenarios.
The solid line, obtained using 200 cluster center points,
closely tracks the bold solid line obtained using all 28137
data points, while the curve obtained with only 10 cluster
center points shows significant discrepancies from the
bold curve, indicating that more than 10 clusters are
needed for accurate modeling of feeder operation.

Power factor at the substation level

0.99
0.98
0.97
0.96
0.95
0.94
0.93
0.92

No PV
Uniform PV distribution (min-max)
Poisson (60%) PV distribution (min-max)
Average PF for both distributions

0.91
0.9

10

20
30
40
50
60
70
80
Percentage of time that PF exceeds curve [%]

90

100

Fig. 13. Power factor duration curves for different PV

spatial distributions.
0.98

No PV
Uniform PV distribution (min-max)
Poisson (60%) PV distribution (min-max)
Average voltage for both distributions

0.975

Voltage at bus 53 [p.u.]

Power factor at the substaion level

0.99
0.98
0.97
0.96
0.95

0.97

0.965

0.96

0.94
0.955

0.93
0.92

10 clusters
200 clusters
All data points

0.91
0.9

10

0.95

20
30
40
50
60
70
80
Percentage of time that PF exceeds curve [%]

90

10

20
30
40
50
60
70
80
Percentage of time that voltage exceeds curve [%]

90

100

100

Fig. 12. Power factor duration curve for several

clustering scenarios.

Fig. 14. Voltage duration curves for different PV

spatial distributions (bus 53).

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

The average and extreme values of total distribution

feeder losses, total energy taken from the transmission
network and the total number of shunt capacitor control
(on/off) actions over 500 Monte Carlo simulations for a
total PV penetration of 760KW are summarized in Table
2. All three quantities are expressed as a percentage of the
values that were calculated without PV support, showing
the improvement obtained by PV penetration. Without
PV, the active and reactive losses are 1.72% and 2.28% of

the total active and reactive energy taken from the

transmission system throughout the year, respectively.
Table 2 depicts results obtained for all considered spatial
PV system distributions (uniform and Poisson), and all
inverter control strategies (supplying only active power,
both active and reactive but limiting inverter power factor
to be higher than 0.85, and supplying reactive power up to
the maximum allowable by the kVA inverter rating).

Table 2. Total feeder losses, energy consumption and number of shunt control actions over 500 Monte
Carlo simulations for different PV spatial distributions and inverter control strategies.
Losses

Total penetration: 760kW (~20%)

No Q
generated

Uniform PV
distribution

Max Q
subject to
pf>0.85
Max Q
allowed by
kVA rating

No Q
generated

Poisson PV
distribution

Max Q
subject to
pf>0.85
Max Q
allowed by
kVA rating

No Q
generated

Poisson PV
distribution
(only 60%)

Max Q
subject to
pf>0.85
Max Q
allowed by
kVA rating

Feeder consumption

Switchings

P [%]

Q [%]

P [%]

Q [%]

[%]

Min

86.05

87.44

92.52

101.80

64.75

Avg

88.06

89.04

92.55

102.75

74.55

Max

90.30

90.77

92.59

103.89

82.73

Min

83.31

84.48

92.47

90.67

59.71

Avg

85.67

86.82

92.51

91.98

67.20

Max

88.19

88.89

92.56

92.83

78.42

Min

77.02

78.98

92.36

41.78

32.37

Avg

79.98

81.32

92.42

46.33

46.01

Max

83.14

84.10

92.47

50.81

64.03

Min

84.51

86.11

92.49

101.34

58.27

Avg

88.20

89.15

92.56

102.73

74.74

Max

93.31

93.32

92.64

104.60

87.05

Min

81.00

82.93

92.43

90.18

48.92

Avg

85.81

86.92

92.52

91.91

67.76

Max

91.85

91.84

92.62

94.06

82.73

Min

74.25

76.61

92.32

39.11

19.42

Avg

80.30

81.58

92.42

46.15

46.92

Max

88.62

88.50

92.56

54.15

70.50

Min

81.80

84.31

92.45

101.11

56.12

Avg

88.27

89.20

92.56

102.70

74.95

Max

94.38

94.46

92.66

104.81

89.21

Min

78.08

81.10

92.38

89.74

46.76

Avg

85.85

86.95

92.52

91.82

68.55

Max

92.39

93.37

92.64

94.27

87.05

Min

72.71

74.96

92.28

35.21

15.83

Avg

80.84

82.01

92.43

45.17

47.82

Max

89.78

90.43

92.58

57.24

76.98

The obtained results verify that PV penetration actually

enhances the feeder condition by reducing losses,
improving the power factor and enhancing the voltage

profile. Thus, voltage stability of the system that supplies

energy to the feeder is also improved.

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

Proceedings of the 34th Hawaii International Conference on System Sciences - 2001

6. Conclusions

7. References

The performance assessment of the DG-equipped

feeder is a challenging problem at the planning stage,
because of the uncertainties involved in both predicting
the DG placement and its future operation. We approach
the problem from the statistical standpoint and identify
the features that allow significant reduction of the
computation involved while retaining the important
information in the input data.

[1] T. Hoff, D. S. Shugar, The Value of Grid-Support

Photovoltaics in Reducing Distribution System Losses,
IEEE Transactions on Energy Conversion, Vol. 10, No. 3,
September 1995, pp 569-576.

In the illustrations used to demonstrate the salient

features of the proposed approach, we also investigate the
impact of DG grid-connected inverter control on feeder
voltage performance. This may be used in the future to
enhance voltage stability of the feeder and the network it
is connected to, which is the subject of ongoing
investigation.

[2] Recommended Practice for Utility Interface of Photovoltaic

(PV) Systems, IEEE 929-2000 Standard, 2000.
[3] M. E. Baran, F. F. Wu, Optimal Capacitor Placement on
Radial Distribution Systems, IEEE Transactions on Power
Delivery, Vol. 4, No. 1, January 1989, pp. 725-732.
[4] W. Marion, K. Urban, "User's Manual for TMY2s", National
Renewable Energy Laboratory, June 1995.
[5] S. Theodoridis, K. Koutroumbas, Pattern recognition,
Academic Press, 1999.
[6] Xie. X. L., Beni. G, "A validity measure for fuzzy
clustering", IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 13, Issue 8, 1991, pp. 841-846.

0-7695-0981-9/01 $10.00 (c) 2001 IEEE

10