17th IEEE Mediterranean Electrotechnical Conference, Beirut, Lebanon, 13-16 April 2014.

Effect of Aggregation Level and Sampling Time on

Load Variation Profile A Statistical Analysis
Intisar A. Sajjad, Gianfranco Chicco, Roberto Napoli
Energy Department
Politecnico di Torino
Torino, Italy
malik.sajjad@polito.it
on a higher level of aggregation the aggregated load pattern
represents the system response that may be more or less
flexible as far as the system operator or aggregator is
concerned. In literature, a lot of work can be found on the
effect of uncertainties associated with load demand to explore
demand side flexibilities [11-16] but it is difficult to find work
related to the effect of sampling time on the shape of
aggregated load patterns. Knowing the characteristics of the
aggregated daily load patterns is a key aspect to manage load
and supply side flexibility for economic system operation.
In this paper, the effects of sampling time as well as
aggregation level on the characteristics of the aggregated load
patterns are addressed on the basis of comprehensive
statistical computations. The analysis of a set of data referring
to extra-urban residential customers is carried out by using the
load variation profiles (Ps), containing the absolute values of
the variations between two successive points of the load
pattern taken at a given sampling interval on the aggregated
daily load pattern. The rationale of this kind of analysis is that
the Ps embed the information on the load variation trends,
and this collective information can be useful to determine how
flexible the demand in different periods of time can be.
Two probabilistic tests have been conducted to get
information about any similarity existing between different
observations with respect to data distribution. For this purpose
we have selected the two-sample Kolmogorov-Smirnov test
and the Wilcoxon rank sum test.
After performing the probabilistic tests it is necessary to
validate the results through other statistical tests that give
some numerical relationship between different sets of data
under study. For this purpose, two parameters are used to
confirm probabilistic results. Percentage relative standard
deviation (ARSD) is a parameter used to measure the
randomness in load variations, and the percentage normalized
load variations (NLV%) measures the average daily load
variation trend. The combination of all the above mentioned
methods can give a good picture above the behavior of load
variations with respect to sampling time and aggregation level.
The rest of the paper is organized as follows. Section II
describes briefly the mathematical framework and data
organization. Section III recalls the statistical methods used for
the analysis. Section IV contains the discussion on results and
Section V concludes the discussion and suggests some future
applications.

Abstract Electrical load patterns that represent the

consumption level are affected by different types of uncertainties
associated with customers behavior and with keeping acceptable
comfort level. The resulting aggregated load pattern indicates the
system response that may be more or less flexible in different
periods of time. Many research activities have been dedicated to
explore the flexibility of load demand using load patterns and
associated uncertainties but little work is found on investigating
the effect of sampling time and aggregation level on the shape of
the load patterns. Knowing the characteristics of the electrical
load patterns is a key aspect to manage load and supply side
flexibilities for most economic system operation. This paper
addresses the effects of sampling interval as well as aggregation
level on the characteristics of the aggregated load patterns. The
study is carried out on the basis of comprehensive statistical
computations on collected data using load variation profiles,
because these profiles embed the information on the load
variation trend. The findings of this study may be used for load
forecasting and management, generation allocation and economic
operation of smart grid system, especially for microgrids.
Keywords Load variation pattern, statistical computation,
aggregation, sampling time, flexibility

I.

INTRODUCTION

In distribution system studies, the characterization of

electrical load patterns is a primary aspect. The information
about time evolution of load patterns is of great importance for
the system operator or aggregator. Time evolution can be
effectively used for network studies, system automation and
control, demand response programs, utilization of load and
supply side flexibilities etc. Time of use tariffs and other
programs have been in place in the last few decades for load
pattern conditioning purposes [1-3]. At present, the interest
towards the analysis of load patterns is focusing on some
important issues like the study of possible impacts of demand
response, tariff differentiation and direct load control for
specific consumer groups [4-9].
The electrical load patterns that represent the consumption
level are affected by different type of uncertainties associated
with customers behavior and with keeping acceptable comfort
level at the customers premises [1, 4]. However, assessing the
time evolution of such systems is a challenging task. For a
small aggregation level (e.g. up to 20 consumers) the
variations in demand are very abrupt and significantly
dependent on life style and kind of family groups [4,10]. Yet,

978-1-4799-2337-3/14/$31.00 2014 IEEE

208

17th IEEE Mediterranean Electrotechnical Conference, Beirut, Lebanon, 13-16 April 2014.

II.

A sketch of the three dimensional representation of the

structure of the complete data set is shown in Fig. 1.

MATHEMATICAL FRAMEWORK AND DATA

ORGANIZATION

The aggregated load profiles for extra-urban residential

customers have been constructed by using Monte Carlo
simulations [10] and the probabilistic characterization of these
profiles has been discussed in [17]. This paper addresses the
effect of aggregation level and sampling time on load
variations by using aggregated residential load profiles. For
each set of aggregation level and sampling interval there are K
observations. By considering I different aggregation levels and
J sampling intervals for each aggregation level, the data can be
represented by using the expressions (1) - (6), where i, j and k
are the variables to describe the aggregation level, the
sampling interval and the observation number, respectively.

Sampling Interval (j)

text

text

P2J2

......

text

P1J2

text

text
...

PIJ1

text

PI22

text

text

PIJ2

PI2K

P2JK
...

PI21

P1JK

PIJK

In this study, the aggregation levels with 20, 50, 75 and

150 households are used for each data set, i.e., I = 4. The data
set under study consists of K = 100 observations for each
sampling time j and aggregation level i. Different sampling
intervals (from 1 minute to 60 minutes) have been used for
analysis but, for potential smart metering applications, only
the intervals with 1, 15, 30 and 60 minute are discussed.
The data samples obtained through experiments may
contain noise, bad data or outliers, and as such it is always
recommended to preprocess them before further analysis. The
data under study for this research work is obtained from
simulations [10] and it is assumed that is error-free.
III.

(3)

STATISTICAL METHODS FOR DATA ANALYSIS

Different statistical tests are carried out to study the effect

of sampling time and aggregation level on load variation
profiles. The behavior of all data samples has been tested to
find the data distribution between different observations and
trend of load variations for different sampling intervals and
aggregation levels.

2 (4)

A. Nonparametric Test
The purpose of probabilistic tests is to compare the effect of
sampling time on load variation patterns in terms of data
distribution. Nonparametric tests are used for comparison
purposes because different data samples may follow different
kind of probability distributions.
1) Two-sample Kolmogorov-Smirnov (KS) test
This is a non-parametric test and an empirical distribution
function is used to compare the data distribution of two
samples. The null hypothesis (0 ) for this test is defined
assuming that both samples belong to the same data
distribution. The result of this test is 1 if 0 is rejected at given
significance level , otherwise it is 0. With given i and j,
compared all observations k have been compared with each
other to verify the above test results. If x, y = 1,2, K, then by
using (5) this test can be formulated mathematically as shown
in (7) - (9).

(5)

1J

IJ

text

...
...

12

I2

P2J1

Fig. 1. 3 dimensional representation of data organization.

(2)

Where each column is a set of all observations at a particular

time instant, and each row is a load variation profile of a
particular observation. Eq. (4) can be rewritten as (5) and (6).

11
=
I1

PI1K

For all values of , Eq. (3) can be represented in matrix

form as follows:

ij = 2

P221
...
...

= [ , 2 , ]

P211

PI12

If the load profile starts from mid night then will exactly
equal to and (2) can be rewritten as:

2 2

P1J1

PI11

(): load variation at time instant t with respect to the

previous time step
(): power demand at time instant
:
sampling interval

:
time instant for the first sample
:
total number of samples
:
load variation profile for a typical day

...
...

Aggregation Level (i)

= [ , + , +
2 + ( 1) ]

1 2

P121

(1)

() = () ( )

ij = 2

P111

(6)

209

17th IEEE Mediterranean Electrotechnical Conference, Beirut, Lebanon, 13-16 April 2014.

, > ()

=
=1 =1(, )

(7)

150

Aggregation Level (No. of Houses)

1,
=
0,

(8)
(9)

, : empirical distribution functions

(): critical value at significance level
, : statistics of test
, : logical decision
:
total no. of fails to reject 0
In this study 5% significance level is used for making
decision and () can be found using standard tables [18].

=
=1 =1(, )

=1

1

=
=1( )

6500
6000

50

5500
5000

10 15 20 25 30 35 40 45 50 55 60
Sampling Interval (Minutes)

4500

Aggregation Level (No. of Houses)

10000
9500
9000
8500
8000

75

7500
7000

50

6500

10 15 20 25 30 35 40 45 50 55 60
Sampling Interval (Minutes)

6000

By using (13) and (14), Eq. (3) and Eq. (6) can be
and as shown in (15) to (18).
represented in terms of
Here the data is reduced to 2 dimensions as the 3rd dimension
is eliminated by averaging the data using observations.

(11)

= [
,

2 ,
]

(12)

11

12

= [ , 2 , ]
11
=
1

B. Average Relative Standard Deviation (ARSD)

and the
The mean load variation profile
corresponding standard deviation for are calculated
by using (13) and (14). The 95% confidence interval is used
for all combinations of and .

7000

Fig. 3. Results of Wilcoxon Rank Sum (WRS) test.

Both tests described above have been used in different

areas of research to judge the goodness of fit of different data
distributions [19-23]. These tests are applied on all the
observations for different sampling intervals and aggregation
levels using (7) - (12). The contour plots for these results are
shown in Fig. 2 and Fig. 3. Since the total number of
observations is 100, we have 10000 comparisons for each
sampling interval and aggregation level.

1
=

7500

75

No. of fails to reject H0

8000

20

Where and are the rank sum values for

and , respectively. The null hypothesis 0 is evaluated
using the expression reported in (11).

8500

150

1,
0,

9000

Fig. 2. Results of the Kolmogorov-Smirnov (KS) test.

= =1 , = =1 (10)

, =

9500

20

2) Wilcoxon Rank Sum (WRS) test

The difference between medians of two samples is
evaluated by using this non-parametric test. The null
hypothesis used for this test is that both data samples are
independent and belongs to the same data distribution.
Let , be the rank vectors of and ,
respectively, and x, y = 1,2, K. Then, for a given
aggregation level and sampling time, we can calculate the rank
sum for and using (10) - (12).

10000

No. of fails to reject H0

, = ( ) ( )

12

(15)

(16)

(17)
(18)

Here,
and are the matrices of organized
and
data sets.
The standard deviation provides the information about
the possible deviation or error with respect to the estimated
. A statistical measure for this deviation is
mean
introduced in this paper to see the effect of sampling time and
aggregation level. This statistical measure (ARSD) is
calculated using (19) and is organized in (20) for all
combinations of and .

(13)

(14)

210

17th IEEE Mediterranean Electrotechnical Conference, Beirut, Lebanon, 13-16 April 2014.

=1

(19)

12

11

60

Percentage Normalized Load Variations

(%NLV)
(NLV
%)

(20)

The results for this statistical measure are shown in Fig. 4

and are summarized in Table I. The smaller values of ARSD
indicate that the data is less deviated and all the observations
are following almost the same trend, and the higher values
represent the opposite case. This indicator is used to verify the
results of nonparametric probabilistic tests numerically.
Avg. Relative Standard Deviation
(ARSD)

1.1

0.9

30
20
20 houses
50 houses
75 houses
150 houses

10

TABLE II.

0.7
0.6
0.5

PERCENTAGE NORMALIZED LOAD VARIATIONS

Aggregation
level

# of
Houses

1.
2.
3.
4.

20
50
75
150

NLV% for sampling interval of:

1
minute

18.0
12.6
10.6
7.7

IV.

0.4

10

20
30
40
Sampling Interval (Minutes)

50

60

C. Percentage Normalized Load Variations (NLV%)

The probabilistic tests and ARSD indicate the randomness
in customers behavior, and the parameter NLV% measures the
trend for absolute load variations on average basis for a typical
day. All the aggregation levels are scaled down to 20 houses
for these calculations for comparison purposes. Eq. (21) is
is normalized using the
used to calculate NLV% and each
contract power which is normally 3 kW for the residential
customers analyzed.
% =

=1

( )

(21)

contract power (3 kW)

no. of houses in aggregation level 1 (20 houses)
no. of houses in aggregation level

Eq. (20) is applied on the selected data set and all the
results are shown in Fig. 5 and Table II.
TABLE I. ARSD FOR DIFFERENT COMBINATIONS OF AGGREGATION LEVELS
AND SAMPLING INTERVALS
Aggregation
level

# of
houses

1.
2.
3.
4.

20
50
75
150

ARSD for sampling interval of:

1
minute

1.086
0.900
0.855
0.803

15
minutes

0.805
0.753
0.738
0.690

30
minutes

0.761
0.681
0.628
0.554

15
minutes

42.2
29.7
25.5
21.4

30
minutes

44.6
35.1
33.2
29.9

60
minutes

56.6
51.0
49.8
47.7

DISCUSSION

The results of the non-parametric probabilistic tests are

shown in Fig. 2 and Fig. 3. From the KS and WRS tests it can
be seen clearly that as the sampling intervals increase, the
numbers of fails to reject 0 increase. The same behavior can
also be seen by increasing the aggregation level. It can be
noted that adopting a sampling time not higher than 20
minutes is a fair option with an aggregation of 20 houses.
However, if the aggregation level increases up to 150 houses,
this limit decreases to 5 minutes. This is a clear indication that
with increase in sampling time and aggregation level we are
actually losing the dynamics of individual customers and
ignoring the actual response of the system.
Table I and Fig. 4 summarize the results for ARSD. From
Fig. 4 it can be observed that for aggregations of 20, 50 and 75
houses the value of ARSD decreases exponentially for smaller
sampling intervals, and afterwards the behavior is almost
linear. For an aggregation of 150 houses it is almost linear for
all sampling intervals. These illustrative results strengthen the
observations of probabilistic tests. From Table I, It can also be
observed that approximately 50% of the dynamics about the
customers behavior is lost when we increase the sampling
time from 1 minute to 60 minutes.
On the other hand, from Fig. 5 and Table II, it is observed
that the magnitude of load variations increases by increasing
the sampling interval but decreases when the aggregation level
is increased. Table II summarizes the relationship between
sampling time and aggregation level. All the results are scaled
down to aggregation of 20 houses. With the increase in the
aggregation level, the load variation decreases to more than
50% for sampling time of 1 minute, but increases when the
sampling time varies from 1 minute to 60 minutes. With the
results of the analysis carried out it is not possible to draw

Fig. 4. Comparison of ARSD for aggregated residentail customers with

different sampling intervals.

1001

60

50

40
30
20
Sampling Interval (Minutes)

10

Fig. 5. Comparison of NLV% for aggregated residential customers with

different sampling intervals.

0.8

:
1 :
:

40

20 houses
50 houses
75 houses
150 houses

50

60
minutes

0.651
0.547
0.496
0.424

211

17th IEEE Mediterranean Electrotechnical Conference, Beirut, Lebanon, 13-16 April 2014.

conclusions for sampling intervals with duration of less than

one minute. Higher sampling interval means less possibility of
following the dynamics of the daily variations. We can say
that it will lead towards reducing the potential of estimation of
demand and supply side flexibilities. For this specific data set,
these mean daily variations are 2 to 3 times higher as we go
from 1 minute to 60 minutes. But with the increase in the
aggregation level, the trend of the results is opposite. The
more the aggregation level, the less the diversity in the mean
daily variations, and these are about 3 times less as we go
from aggregation level of 20 houses to 150 houses. These
results refer to extra-urban customers and may vary depending
upon the topological and demographic situation. In any case,
the results are indicative of the possible trends referring to the
absolute load pattern variations in time.
V.

[4]
[5]

[6]

[7]
[8]
[9]

CONCLUSIONS
[10]

Proper selection of sampling interval and aggregation level

is very much important to effectively use demand and supply
side flexibilities. In this paper, a comprehensive study based
on statistical parameters has been carried out to show the
effect of both aggregation level and sampling time on the
characteristics of the aggregated load patterns. The results,
referring to the specific data set of extra-urban aggregated
residential customers, can be interpreted in terms of minimum
number of residential customers to consider in the
aggregation.
It can be concluded from the above results that with higher
sampling intervals we are actually losing the dynamics of the
individual customers behavior and over-estimating the
diversity of load variations. But with the increase in the
aggregation level, we get less knowledge on the changes that
may be introduced in the aggregated load by the individual
customers behavior. From the above discussion it can also be
concluded that if we want to increase the aggregation level,
then we have to reduce the sampling time for accurate
representation of the load patterns aimed at investigating the
demand and supply side flexibilities.
The approach followed in this paper can be used to
determine the structure of a metering system for a microgrid
application in which, according to the current ICT
technologies, it is possible to decide a compromise solution
between aggregation size and sampling interval duration.
The findings of this research work will be applied for
assessing the demand side flexibility margins and utilize them
for scheduling of supply side resources for more economical
operation of system.

[11]

[12]
[13]
[14]
[15]

[16]
[17]
[18]
[19]

[20]
[21]

REFERENCES
[1]
[2]
[3]

C. F. Walker and J.L. Pokoski, Residential load shape modeling based

on customer behavior, IEEE Trans. Power Appar. Syst., vol. PAS-104,
no. 7, pp. 17031711, 1985.
I I.C. Schick, P.B. Usoro, M.F. Ruane and J.A. Hausman, Residential
end-use load shape estimation from whole-house metered data, IEEE
Trans. Power Syst., vol. 3, no. 3, pp. 986991, 1988.
A. Sargent, R.P Broadwater, J.C. Thompson and J. Nazarko,
Estimation of diversity and KWHR-to-peak-KW factors from load
research data, IEEE Trans. Power Syst., vol. 9, no. 3, pp. 14501455,
1994.

[22]
[23]

212

A. Capasso, W. Grattieri, R. Lamedica and A. Prudenzi, A bottom-up

approach to residential load modeling, IEEE Trans. on Power Systems,
vo1.9, no. 2, pp. 957-964, May 1994.
P. Stephenson, I. Lungu, M. Paun, I. Silvas and G. Tupu, Tariff
development for consumer groups in internal European electricity
markets, Proc. 16th International Conference and Exhibition on
Electricity Distribution, June 2001.
A. Molina, A. Gabaldn, J.A. Fuentes and C. Alvarez, Implementation
and assessment of physically based electrical load models: application to
direct load control residential programmes, IEE Proc. Gener. Transm.
Distrib., vol. 150, no. 1, pp. 6166, 2003.
G. Chicco, R. Napoli, P. Postolache, M. Scutariu and C. Toader,
Customer characterisation options for improving the tariff offer, IEEE
Trans. Power Syst., vol. 18, no. 1, pp. 381387, 2003.
D.S. Kirschen, Demand-side view of electricity markets, IEEE Trans.
Power Syst., vol. 18, no. 2, pp. 520527, 2003.
R. Malme, International energy agency demand side program Task
XIII: demand response resources position paper, Proc. IEEE/PES
Power Systems Conf. and Exposition, 1013 October 2004, vol. 3, pp.
16351640.
A. Cagni, E. Carpaneto, G. Chicco and R. Napoli, Characterisation of
the aggregated load patterns for extra-urban residential customer
groups, Proc. IEEE Melecon 2004, Dubrovnik, Croatia, 1215 May
2004, vol. 3, pp. 951954.
A. Abdisalaam, I. Lampropoulos, J. Frunt, G.P.J. Verbong and W.L.
Kling, Assessing the economic benefits of flexible residential load
participation in the Dutch day-aheadauction and balancing market,
International Conference on European Energy Market, 2012, pp. 1-8.
B. Halvorsen and B.M. Larsen, The flexibility of household electricity
demand over time, Resource and Energy Economics, vol. 23, no. 1, pp.
1-18, 2001.
H. Hildmann and F. Saffre, Influence of variable supply and
load flexibility on Demand-Side Management, 8th International
Conference on the European Energy Market (EEM), 2011, pp. 63 68.
B. Kladnik, A. Gubina, G. Artac, K. Nagode and I. Kockar, Agentbased modeling of the demand-side flexibility, IEEE Power and Energy
Society General Meeting, 2011, pp. 1 8.
R. A. Rosso, J. Ma, D.S. Kirschen and L.F. Ochoa, Assessing the
contribution of demand side management to power system flexibility,
50th IEEE Conference on Decision and Control and European Control
Conference (CDC-ECC), 2011, pp. 4361 4365.
R A. Martin and R. Coutts, Balancing act [demand side flexibility],
IET Journal on Power Engineer, vol. 20, no. 2, pp. 42 45, 2006.
E. Carpaneto and G. Chicco, Probabilistic characterization of the
aggregated residential load patterns, IET Generation, Transmission &
Distribution, vol. 2, no. 3, pp. 373 382, 2008.
K.S. Trivedi, Probability and Statistics With Reliability, Queuing and
Computer Science Applications, Wiley, New York, 2002.
P. De and J. Douglas, Testing Goodness-of-Fit for the Singly Truncated
Normal Distribution Using the Kolmogorov-Smirnov Statistic, IEEE
Trans. on Geoscience and Remote Sensing, vol. GE-21, no. 4, pp. 441
446, 1983.
W. Fanggang and W. Xiaodong, Fast and Robust Modulation
Classification via Kolmogorov-Smirnov Test, IEEE Trans. On
Communications, vol. 58, no. 8, pp. 2324 2332, 2010.
S. W. Samuel and H. A. David, On stochastic orderings of the
Wilcoxon Rank Sum test statisticWith applications to reproducibility
probability estimation testing, Journal of Statistical Planning and
Inference, vol. 137, no. 6, pp. 1838 1850, 2007.
J. L. Young, Finite rank sums of products of Toeplitz and Hankel
operators, Journal of Mathematical Analysis and Applications, vol.
397, no. 2, pp. 503 514, 2013.
J. W. Phegley, K. Perkins, L. Gupta and L.F. Hughes, Multicategory
Prediction of Multifactorial Diseases Through Risk Factor Fusion and
Rank-Sum Selection, IEEE Trans. on Systems, Man and Cybernetics,
Part A: Systems and Humans, vol. 35, no. 5, pp. 718 726, 2005.