SPATIAL ELECTRIC LOAD FORECASTING USING AN EVOLUTIONARY

HEURISTIC

E. M. Carreno∗ A. Padilha-Feltrin∗

A. G. Leal†

∗
Faculdade de Engenharia de Ilha Solteira, UNESP - Univ Estadual Paulista,
Departamento de Engenharia Elétrica CP 31
CEP 15385-000 - Ilha Solteira SP
†
ELUCID SOLUTIONS S/A
Av Angélica, 2318 Mezanino
CEP 01228-904 São Paulo SP

ABSTRACT 1 INTRODUCTION
A method for spatial electric load forecasting using elements During the last few decades, the growth of emerging
from evolutionary algorithms is presented. The method uses economies, such as some countries in South America and
concepts from knowledge extraction algorithms and linguis- Asia, has enabled their populations to improve living stan-
tic rules’ representation to characterize the preferences for dards; this phenomenon means that every year more inhabi-
land use into a spatial database. The future land use pref- tants of these countries look for new homes. This, in tandem
erences in undeveloped zones in the electrical utility ser- with increasing migration from countryside to cities (espe-
vice area are determined using an evolutionary heuristic, cially in Asian cities), has created a high level of city ex-
which considers a stochastic behavior by crossing over sim- pansion. Distribution utilities are currently facing new chal-
ilar rules. The method considers development of new zones lenges when planning a network expansion that offers a reli-
and also redevelopment of existing ones. The results are pre- able and economical service inside their service zone.
sented in future preference maps. The tests in a real system
from a midsized city show a high rate of success when re- This situation is more difficult for decentralized electric mar-
sults are compared with information gathered from the util- ket structures, where distribution utilities should now deal
ity planning department. The most important features of this with profitable operations while considering aging electrical
method are the need for few data and the simplicity of the structures owing to the lack of investments during the cen-
algorithm, allowing for future scalability. tralized era.

KEYWORDS: Spatial electric load forecasting, land use, Over the last few years, research in the distribution area has
knowledge extraction, distribution planning. increased substantially, and attempts to address these grow-
ing problems. In many countries, the distribution sector is
going through a deep restructuring owing in part to the search
for improved quality service levels demanded by consumers
Artigo submetido em 10/07/2009 (Id.: 01014) in new market-driven structures (Rudnick et alii, 2007). Be-
Revisado em 18/11/2009, 10/04/2010 cause of this, over the last few years, research in the distribu-
Aceito sob recomendação do Editor Associado Prof. Ivan Nunes Da Silva

Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010 379

tion area has increased substantially to address these growing this zone is also the one with higher energy consumption.
problems. The growth of the city demands growth of the urban core and
its surroundings. The urban core can grow both horizontally
To improve the distribution network expansion and planning and vertically.
by considering these factors, distribution utilities should con-
sider spatial electric load forecasting as business priority. This paper deals with a new method for spatial electric load
In the particular case of Brazil, a new regulatory law for forecasting, by considering new developed zones and rede-
Electrical Distribution Utilities, called PRODIST, imposes velopment of the existing ones, using concepts of knowledge
the use of Spatial Electric Load Forecasting in their studies extraction algorithms, and elements from evolutionary algo-
(ANEEL, 2008). rithms, with special emphasis on developing a simple but
powerful method with a reasonable use of available data, and
Spatial electric load forecasting is a process that tries to de- taking into account the stochastic nature of new consumers.
termine the future load growth for an electric energy distribu-
tion utility (Lee, 2002). This process, apart from answering
the question of how much load growth is expected in a deter- 2 OVERVIEW ON SPATIAL LOAD FORE-
mined time frame, also identifies the locations that are most CASTING
likely to receive the new loads.
In specialized literature, spatial electric load forecasting dis-
Electric load growth inside the service area of an electric util- cussion can be traced back to Van Wormer (1954), where
ity can be expected for two reasons: first, natural growth the need for a more detailed load forecasting considering the
because of the natural behavior of existing consumers; and shape of the cities was evidenced. Several works were devel-
second, new loads because of new consumers. The natural oped inside electrical utilities, and different techniques were
behavior of existing consumers is stationary, with low ex- used considering different factors. Most of these works are
pected growth; thus, the main reason for load growth is the compiled in (Lee, 2002), which is one of the fundamental
new consumers inside and outside the actual service zone. sources of information in this area.

Therefore, the location of these new consumers is of vital The first techniques used pattern recognition to determine
importance for the utilities so that they can adopt all the nec- preference maps (Lee, Anstos, 1979), where different fac-
essary resources for meeting the needs of the future system. tors are weighted to determine a specific land use type. This
Such information can be used to reinforce substations, elec- process was later improved using elements from fuzzy logic
tric structures, system reconfiguration, and allocation of new (Miranda, Monteiro, 2000). Different approaches in fuzzy
substations, among several other activities. logic have been developed for this problem, like fuzzy mul-
tiobjective decision-making (Chow et alii, 1998), cloud the-
Most of these new consumers will be located outside the ser- ory (Yang et alii, 2006), ] and recently, statistical approaches
vice area, continuing the natural growth of the cities, but oth- (Vazquez et alii, 2008).
ers will be located inside the service area in zones that are
redeveloped. The redevelopment process considers changing Later years have not shown special advances in spatial elec-
the use of some part of the land, generally to accommodate trical load forecasting; hence, it is necessary to take advan-
more loads. For example, a zone with horizontal housing tage of new advances in algorithms to develop new method-
could be transformed into a zone with vertical housing, or a ologies especially suited to the reality of new cities in emerg-
residential area could be transformed into a commercial area ing countries, considering the requirements of new users and
(Mo-Yuen, Hahn, 1996). comply with new regulations in the sector.

Strategic redevelopment is carried out because of major ur- The spatial electric load forecasting process begins with the
ban projects looking to improve certain areas of the cities; data recollection from different sources; these data go from
generally, this kind of redevelopment is planned politically, different maps of the utility service area with different in-
making it very difficult to forecast. One of the options is formation such as land use classification; geographical land-
using scenario simulations. scape; socioeconomic distribution; electric structure inven-
tory; historical peak load in different strategic points, like
Another kind of redevelopment is the growth of the urban substations and feeders. This information is complemented
core. The urban core acts as an activity center for the entire with that from the different entities entrusted by city planning
city, concentrating a large quantity of commercial and resi- and information from the utility planning department.
dential loads into a relatively small subzone, with financial
and cultural offers that attract the rest of the city inhabitants One of the major problems for distribution utilities to carry
for business or leisure. Because of the concentration of loads, out a spatial electric load forecasting is the lack of available

380 Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010

data; hence, the method presented in this work considers that The commercial consumer database available at most dis-
the utilities have access to basic information, like the com- tribution utilities identifies all the consumers by name, ad-
mercial consumer database and an inventory of all the elec- dress, class, consumer code, and historical energy consump-
trical structures in the system. All these data are organized tion. This information can be crossreferenced with the tech-
in a multidimensional database that allows extracting infor- nical databases using the consumer code, which should be
mation related to specific zones according to the needs of the the same in both databases. With both databases, it is pos-
chosen method. sible to allocate energy consumption spatially in the service
area.
After identifying the load growth using conventional meth-
ods (Alfares, Nazeeruddin, 2002), the new load is allocated In case the utility does not have historical records about peak
into the service area using the geographical and socioeco- load in substations, a method using statistical daily load pro-
nomic information from the spatial database. This is one of files can be used to estimate this value from the energy read-
the most difficult steps in the process because of the stochas- ings in the commercial database (Jardini et alii, 2000).
tic nature of the urban development.
For the pilot application of this method, a database was built
This paper presents an extension of an earlier work (Carreno, using information including the geographical distribution of
Padilha, 2008) by considering new developed zones and re- all the electric structures inside the service area; utility com-
development of the existing ones. The use of concepts of mercial database of all the consumers with energy consump-
knowledge extraction algorithms and elements from evolu- tion from the past years; statistically obtained load curves for
tionary algorithms is proposed. all types of consumers in different consumption levels; and a
land use classification map of the zone.
A large database of classification rules obtained automati-
cally by extracting the knowledge available in the spatial This information was crossreferenced, and a multidimen-
database of the city is used to identify user preferences for sional spatial database was created. This database allows
land use considering elements such as distance to nearby fa- visualizing specific data such as load growth for specific ar-
cilities, major roads, and electrical load density. eas, substations, feeders, transformers, utilization factors, in-
stalled kVA, spatial load density, and other valuable informa-
The redevelopment of the urban core is modeled by consid- tion.
ering spatial information from the city’s growth as well as
user preferences about land use. This proposal models the The spatial resolution adopted was 0.5 km2 to take advantage
urban core growth using the database of classification rules of the information available, but smaller resolutions are avail-
to analyze the preferences and dynamics of the urban core able on demand. Figure 1 presents the map of the service
related with urban growth, and this information is used to area with a total extension of 180 km2 .Each circle represents
create a preference map for urban core growth. The data, in one subarea of 0.5 km2 . Black circles represent currently
conjunction with the residential and commercial preference occupied areas and white circles represent unoccupied areas
maps, would generate a forecast to determine how the urban in zones with the possibility of being occupied. The zones
core will evolve. without circles represent the areas without the possibility of
development for different reasons.
The crossreference of the map’s information is made through
a stochastic evolutionary heuristic that creates new classifica- From the commercial database, it is possible to identify the
tion rules using the patterns already recognized in the spatial location of the different structures of interest such as schools,
database. hospitals, and commercial zones, among others. The ma-
jor roads and zones without potential can be identified from
The method is presented and tested with data from a real mid- aerial maps, and other information is obtained from the util-
sized city. The results present the preference and forecast ity planning department, like activity centers, and special ur-
maps of the study zone. The validation process of the results ban interest zones, to name a few.
is made by comparing the results with the ones obtained in
the utility planning department using the known procedure The information used in this work is readily available in most
of manual simulation from Lee (2002). of the electric utilities around the world, allowing fast imple-
mentation.
3 SPATIAL DATABASE
4 CLASSIFICATION RULES
The availability of data is the most important factor in any
methodologies dealing with land use determination in spatial To allocate the future loads inside the service area, it is nec-
electric load forecasting. essary to consider several factors; for example, a pollutant

Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010 381

 Attribute refers to the measure of the relations between the
different zones and the elements. If one element is distance
to a road, the rules are expressed by near, close, away, far-
away. For an element of residential electrical density, the
rules are expressed in low, normal, and high. Each Attribute
corresponds to a finite and exclusive set in the domain of the
corresponding object. For example: All the distances from 0
km to 1km inclusive to a school is considered near; All the
distances between 1.01 km to 5 km inclusive are considered
as medium; and so on. The range of values for each attribute
are selected based on an statistical analyses of all the data for
each object.

For each object, customers have different preferences, and

that is why a different set of attributes exist for each one, for
example, the attribute near does not have the same meaning
for a school or for the activity center of the city.

Figure 1: Service area map, aerial view.

5 CODING OF THE CLASSIFICATION
RULES
To extract and use the knowledge from the spatial database,
industry cannot be allocated near residential zones; residen- it is necessary to set up an acceptable coding to efficiently
tial consumers like to live near other residential zones, but store the information gathered during the process.
not too far away from commercial zones and not too dis-
tant from their workplaces and so on. These preferences are The classification rules (Input and Output) are coded as a
sensitive to location and culture; hence, they are specific for vector, where each position represents an element of interest,
each city. If static rules were adopted, it will be necessary to and the value stored represents a linguistic expression suit-
change them for each study. Some of these new consumers ing to the specific object. All the elements are related with
will be allocated in already populated zones having further the AND operator. Thus, a vector containing linguistic ex-
growth opportunities, while others will be allocated in new pressions for a set of objects represents a rule. For example,
zones that will be populated in the midterm with new infras- using the values {L N H} to represent three different linguis-
tructure. tic references as Low (L), Normal (N), and High (H) for each
of the objects considered in the classification rule, the rule is:
The local authorities have made a series of studies related
with the city growth; unfortunately, this information is not INPUT (IF Distance to object 1 is L, AND Distance to ob-
reliable because of its political dependence. Hence, future ject 2 is N , . . . , AND Residential load density is H) then
growth studies should use this information only as an indica- OUTPUT (Residential load is H, AND Commercial load is
tor, not as a rule. N , AND Industrial load is L). This rule can be represented
as (L, N, . . . , H) => (H, N, L), which can be written as the
One of the tools used to decide which new zones will be pop- vector [L N . . . H][HN L].
ulated in the mid- and long-term in the service area was the
classification rules, discussed first in (Chow et alii, 1998), For this particular application, the Input objects were chosen
and broadly used in different algorithms since then. In (Mi- as distance between the subarea and interest points, and load
randa, Monteiro, 2000), classification rules were introduced densities; the load densities are calculated in 3×3 and 5×5
to this problem to determine a land use score; these rules can subareas around the subarea in study, and the output objects
be expressed as: are loads for consumer class.

(IF Object (1) IS Attribute) AND (IF Object (2) IS Attribute) The input and output objects are not fixed and can comprise
AND . . . . THEN (LandScore IS Attribute) any combination of objects, so long as a linguistic definition
set for them exists.
Here, Object refers to a set of elements of interest to be con-
sidered in the process, like structures (schools, roads, high-
ways, activity center), nearby zones, or other elements of in-
terest.

382 Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010

6 KNOWLEDGE EXTRACTION X
SM = |RuleA(i) − RuleB(i)| (1)
To characterize the service territory, it is necessary to find out i=1..#Objects
the classification rules that will allow us to settle the future
land use preferences. One way to deal with the rules for the unpopulated area will
be to build the input part of the linguistic classification rules
These rules and the defining sets could be determined man-
and settle the output according to the same part of the rule
ually by the planner using rules supported by common sense
with the highest SM value among all the rules in R1.
and practice, however, these rules could work well for one
specific city based on the expertise and the knowledge of The process is described next:
the planner. The differences between cities around the world
demonstrate the need to settle a different set of rules for each
one. The method in this paper presents a systematic way to For each unpopulated area k:
extract a set of rules automatically for each city.
1. Build the input part of the classification rule;
Using the information available from the spatial database, it
is possible to create a database of known rules that can be 2. Calculate the SM value between the rule in analysis and
later used to settle the preferences for future areas, consider- all the elements in R1;
ing that the growth of the city is a stationary stochastic pro-
cess. The procedure for creating such database is as follows: 3. Assign to the output part of the classification rule in
analysis, the output of the rule in R1 with the highest
BeingR1 the set of rules for the populated areas with known SM.
output,
In this manner, however, the randomness of the behavior for
1. In the spatial database, identify all the elements of in- this particular problem is lost. Besides, owing to the nature
terest (Input), like activity centers, schools, major high- of the distribution of the unpopulated areas (generally out-
ways and roads, parks, among others; side the service area), the SM value could be low and not
significant enough to assign a logical output. Another prob-
2. Establish the linguistic expression for each element of lem arises when several rules with the same input part have
interest; different outputs; hence, it is not possible to choose one in-
stead of another only with this information.
3. For each populated subarea i: To overcome these problems, an evolutionary approach is
proposed by selecting the output part of the classification rule
a. Calculate the linguistic expressions for each element with more than one source of similar rules. The output part
of interest, considering the relation of the subarea of the classification rule is determined by selecting out a set
with the elements of interest and their surroundings, of similar rules (RSM), sorting it using the SM value for each
form the linguistic rule and store it in R1. one as the fitness function, and applying a uniform crossover
between two candidates selected by a tournament. In this
way, the output parts of the rules are chosen using the char-
With this algorithm all the knowledge present in the popu- acteristic of this genetic operator as controlled randomness
lated area is extracted and processed. and statistical survival of the fittest.

In some knowledge-extracting algorithms, the known rules In a tournament selection, two or more solutions are picked
will be mixed using different evolutionary operators to create at random and a tournament strategy is used to select one af-
new rules (Yi-Chung et alii, 2003). This approach is not used ter a number of rounds are played. In each round, a number
because most of the rules created will be useless; instead, of solutions come together to compete, and only the fittest
each unpopulated area is analyzed to find out its output, by solution will win and be selected (Chambers, 1999). For this
optimizing resources and minimizing computational time. particular application, two rounds are played, each one with
two solutions randomly picked using the SM value as the fit-
To compare how similar two rules are, the Hamming distance ness function.
is used. The Similarity (SM) (1) is defined as the Hamming
distance between two rules, that is, the number of positions The crossover operator is one of the most important operators
where the corresponding values of the rule Aare different in genetic algorithm. This operator is the one in charge of
than the ones in rule B (Hamming, 1950). propagating the genetic information of two elements of the

Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010 383

current population to the next, by combining their genetic
Table 1: RSM Set
code to generate a descendant.
Element SM Input Output
The complete process is detailed in the following algorithm: 1 5 [N L L L N H ] [LN H]
2 5 [H N L L N H ] [LLH]
For each unpopulated area k : 3 3 [N L N L N N ] [LN N ]
4 4 [L N L N N H ] [N LH]
1. Build the input part of the classification rule; 5 3 [L N L N N N ] [LHL]
6 4 [L N L L L H ] [LN H]
2. Calculate the SM value between the rule and all the ele-
ments in R1;

3. Build a set of similar rules (RSM) using the SM value;

Using a tournament selection, vectors (1) and (3) from the
4. Select two candidates to be crossed over using tourna- RSM Set presented in Table 1 are selected to be crossed over.
ment selection;
The output parts of rules (1) and (3), [LN H] and [LN N ],
5. Crossover the output part of the two candidates selected, are mixed using a uniform crossover obtaining the vector
and assign that value to the output part of the classifica- [LN H].
tion rule in analysis.
At the end, the output part for the rule [N N L L N H ] should
be [LN H].
The size of the RSM set was determined as 10 for a vector
of 13 elements (10 inputs, 3 outputs) by experimentation; With the results obtained until this point, it is enough to make
with values higher than 10 the SM values are too low, and an entry-level spatial electric load forecasting process that is
with lower values the diversification is low. At the end of able to identify all the new zones with possibility of growth.
the process, all the unpopulated areas are classified into dif- But to complement the results and lower the spatial error, it
ferent preferences for consumer class and expected load den- is necessary to consider some level of redevelopment inside
sity, creating a preference map where it is possible to identify the service area.
different areas of interest to the planner.
The redevelopment process considered is the urban core
Considering the stochastic nature of this process, the algo- growth. The rule database and the same principles of sim-
rithm is repeated several times to identify the interest areas ilarities between zones are used in this step; hence, there is
more clearly. no need for new data or different mechanisms than the ones
already developed. The process starts by identifying the ac-
To explain the idea, a brief example is presented next: tivity center by searching for the subareas with the highest
values of residential and commercial energy consumption.
The values {LN H} represent three different linguistic ref-
The location of this special subareas is easy to identify be-
erences as Low, Normal and High for each of the objects
cause in the spatial database a zone identified as the urban
considered in the classification rule.
core in the “elements of interest” already exists.
The classification rule for an unpopulated subarea under
The classification rules of the subareas in the urban core are
analysis is given by:
compared with the ones in a 3×3 neighborhood zone, and a
(Distance to major road is N , Distance to nearest school is score is calculated adding those values. This value represents
N , Distance to activity center is L, Residential density in the relationships between the urban core and its immediate
3x3 neighborhood is L, Commercial density in 3×3 neigh- neighbors. Next, the classification rules for all the subareas
borhood is N , Industrial density in 3×3 neighborhood is H). in the service zone are compared with the ones in the urban
core and its own 3×3 neighborhood zone. When presented
The vector [N N L L N H ] is the input part of the classifica- graphically, the results of this exercise will show how the
tion rule for the unpopulated area k. different subzones share similarities with the urban core, and
especially, how the different factors in the city will take the
For this vector, the RSM set size is determined as 6 for ex- urban core growth in a specific direction.
perimentation. The vector is compared with all the elements
in the set R1, and an RSM set is formed with 6 elements, as An outline of the complete method is presented in Fig. 2.
shown in Table 1. This process is done for each planning horizon. The results

384 Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010

 Input

• Distance to nearest school;

• Distance to nearest hospital;
• Distance to nearest major road;
• Distance to main activity center;
• Distance to secondary activity center;
• Residential load density in a 3x3 neighborhood zone;
• Commercial load density in a 3x3 neighborhood zone;
• Industrial load density in a 3x3 neighborhood zone;
• Residential load density in a 5x5 neighborhood zone;
• Commercial load density in a 5x5 neighborhood zone;
• Industrial load density in a 5x5 neighborhood zone.

Output

• Residential load;
• Commercial load;
• Industrial load.

Figure 3 presents the results obtained for the residential class.

Figure 2: Outline of the complete algorithm In this figure, the circles represent the zones of the service
area where some load is already present. The area expected
to be occupied according to the information obtained from
the utility planning department (validation set) is identified
by diamonds, and the area expected to be occupied is identi-
obtained in planning horizon n-1 are used as input elements
fied by squares.
for planning horizon n.
In Fig. 3, one area in the South is identified by the plan-
7 TESTS AND RESULTS ning department as being occupied for residential consumers
in the future but is not identified by the algorithm. This is
The method was incorporated into the UE viewer, a dis- because a large project is expected to be built in that part of
tribution network management software designed for Elu- the city, changing the classification of this area from indus-
cid Solutions, using data from a midsized city in Brazil trial to commercial. Because of the industrial classification
called Presidente Prudente, in the interior of São Paulo State, of the zone, the algorithm did not consider this area suitable
with around 200,000 residents and two major activity centers for residential occupation.
identified as major commercial zones.
In other tests, considering the change of the land use clas-
Some information obtained about the service area, especially sification and the inclusion of a large simulated load in this
information on future land developments, were not included zone, it was possible to match this expected behavior.
in the database. This would be used as a validation set for the
results. Another interesting result is the number of areas identified by
the algorithm that were not identified by the manual simula-
With the information available, the next objects were chosen tion of the utility planning department. This is because the
to form the classification rules for this particular system: zones forecasted by the planning department were based in

Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010 385

 Figure 5: Service Area Map with expected commercial pref-
Figure 3: Service area map with expected residential prefer- erence.
ence.

Figure 4: Service Area Map with expected industrial prefer-

ence. Figure 6: Spatial allocation for the three major classes.

present and future development projects, and not likely the

terrain characteristics. Therefore, the answer obtained from these areas. To do this, after calculating the expected growth
the algorithm is more complete because it is possible to iden- because of new consumers in the service area using tradi-
tify all the zones of interest in the long term. tional methods, the areas are assigned considering neighbor-
ing characteristics, assigning first the ones with a bigger load
The results for the industrial and commercial sector are pre- density nearby, obtaining the results presented in Fig. 7.
sented in Figs. 4 and 5, respectively, presenting complemen-
tary patterns between the three consumer classes analyzed The methodologies used to calculate the load growth are not
(see Fig. 6). part of the scope of this work. This load growth was cal-
culated using time series and population data, but any other
Finally, it is necessary to settle the temporal allocation of method could be used to determine this value.

386 Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010

 Figure 8: Redevelopment preference map.

8 CONCLUSIONS
Figure 7: Temporal allocation for residential consumers. A simple and efficient algorithm to settle future land use and
occupation in the spatial electric load forecasting process has
been presented by considering the identification of new zones
with potential growth and the redevelopment of the existing
ones. This application has been proved to be an important
tool in assisting distribution network planning engineers to
identify important zones with the possibility of growth inside
The first areas to be occupied are located in the southeast the service area. The algorithm uses simple procedures to ex-
part of the service area, an area with rapid and constant urban tract the knowledge related to user preferences about land use
projects for mid- and high economic levels, and in the north from a multidimensional spatial database. This knowledge is
of the city, where proximity to another small city and land stored in the form of linguistic classification rules.
intended to popularize housing are important factors. This
information was not introduced into the algorithm. It is possible that several classification rules with the same in-
put present different outputs. This is because of the stochas-
Figure 8 shows the redevelopment preference map based on tic nature of the process, which makes the use of fixed clas-
the urban core growth. Obviously, the areas with higher sification rules a problem. To overcome this difficulty, the
probability of redevelopment are the ones around the urban consumers’ stochastic characteristic is considered using el-
core, but it is important to note that the urban core is not ements from evolutionary algorithms to create new classifi-
growing axially around that point, but that it has a strong in- cation rules for the vacant areas using the knowledge previ-
fluence to grow in the southwest. This behavior is explained ously extracted.
by the growth in commercial and residential density in that
direction, especially for high-income housing. Meanwhile, The final result shows the efficiency of the method in cre-
to the north and east of the urban core, growth is not expected ating a forecast map similar in 95% to the utility planning
because of the low growth in electrical load density. department forecast, with low computational effort and min-
imum computer memory requirements, using few resources
To complement the validation of the method, the results were and data. The results obtained prove the efficacy of the ap-
compared with an internal forecast of the electric utility, proach and encourage more work in that direction.
made manually for the engineers in charge of the network
planning, using the manual method presented in (Lee, 2002) One of the main advantages of this method is the reduced set
, resulting in an error rate between 5% and 10% in different of data needed to execute it, using only the distribution util-
simulations the short term (5 years) and mid term (10 years). ity commercial database and a georeferenced data set of the
For the long term (20 years) there were no results to compare network elements. This approach could be helpful to utilities
to. These results are considered very good, specially because with limited resources and few available data. The multidi-
the manual simulation from (Lee, 2002) could take easily one mensional spatial database created for this application can be
month of work for several employees. used as support information to solve other problems related

Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010 387

to the planning and operation of the distribution network. Van Wormer, F. C., "Some Aspects of Distribution Load
Area Geometry," Power Apparatus and Systems, Part
REFERENCES III. Transactions of the American Institute of Electrical
Engineers, vol.73, no.2, pp.1343-1349, Jan. 1954.
Alfares H. K., Nazeeruddin M. “Electric load forecasting: lit-
Vasquez-Arnez, R.L.; Jardini, J.A.; Casolari, R.; Magrini,
erature survey and classification of methods,” Interna- L.C.; Semolini, R.; Pascon, J.R., "A methodology
tional Journal of Systems Science, v.33, no 1, p. 23-24,
for electrical energy forecast and its spatial allocation
Jan. 2002. over developing boroughs," in Proc Transmission and
ANEEL, Procedimentos de Distribuição - PRODIST 2008. Distribution Conference and Exposition, 2008. pp.1-6,
Available online http://www.aneel.gov.br/. 2008.
(In portuguese) Yang, X., Yuan, J., Zhang, T., Mao, H., “Application of Un-
Carreno, E.M.; Padilha-Feltrin, A., "Evolutionary heuristic certainty Reasoning Based on Cloud Theory in Spatial
to determine future land use," in Proc. Power and En- Load Forecasting,” in Proc. of Intelligent Control and
ergy Society General Meeting, IEEE , pp.1-6, 20-24. Automation, vol 2, pp 21-23. Jun. 2006.
Jul. 2008. Yi-Chung, H., Ruey-Shun, C., Gwo-Hshiung, T., “Find-
Chambers, L. Practical Handbook of Genetic Algorithms: ing fuzzy classification rules using data mining tech-
Complex Coding Systems, Volume III. USA, CRC niques,” Pattern Recognition Letters, no 24, pp 509-
Press, 1999. 519, Jan. 2003.

Chow, M., Zhu, J., Tram, H., “Application of fuzzy multi-

objective decision making in spatial load forecasting,”
IEEE Trans. on Power Systems, vol 13, pp 1185–1190.
Aug. 1998.

Hamming, R. W., “Error detecting and error codes.” Bell

System Technical Journal, New York, v.26, no 2, p.147-
160, 1950.

Jardini, J. A., Tahan, C. M., Gouvea, M. R., Ahn, S. U.,

Figueiredo, F. M., “Daily Load Profiles for Residential,
Commercial and Industrial Low Voltage Consumers.,”
IEEE Trans. on Power Delivery, vol. 15, no. 1, pp 375-
380, Jan. 2000.

Lee, W, H. Aanstos, J., “Some unique signal processing ap-

plications in power systems analysis,” IEEE Trans. on
Acoustics, Speech, and Signal Processing, v. 27, no 6,
pp 685, Dec. 1979.

Lee, W, H. Spatial Electric Load Forecasting, 2nd Edition.

New York, Marcel Drekker Inc, 2002.

Miranda, V., Monteiro, C., “Fuzzy inference in spatial load

forecasting,” in Proc. Power Engineering Society Win-
ter Meeting, pp 1063-1068 vol.2, Jan. 2000.

Mo-Yuen, C., Hahn, T., "Methodology of urban re-

development considerations in spatial load forecast-
ing," in Proc. Transmission and Distribution Conf, pp
452-457, Sep. 1996.

Rudnick, H.; Arnau, A.; Mocarquer, S.; Voscoboinik, E.,

"Stimulating efficient distribution," IEEE Power and
Energy Magazine, vol.5, no.4, pp.50-67, Jul. 2007.

388 Revista Controle & Automação/Vol.21 no.4/Julho e Agosto 2010