You are on page 1of 22

2 32 PROCEEDINGS OF THE IEEE, VOL. 71, NO.

2, FEBRUARY 1983

Spatial Electric Load Forecasting:


A Tutorial Review

Abmuct-A spatial laad forecast is a forecast of the future locationsthe last few years. During the last decade, a number of prac-
and magnitudes of electric load within a utility's service territory.Such
tical and efficient automated computerized methods have been
forecasts are a necespIly part of power deiive.ry system phning. Ihic
paper presents a comprehensive review of spatial load f o h g tech-
developed as an aid in delivery system planning. Many employ
niques. A discussion of planning needs, uncertainty, spatial growth optimization to refine the system design and determine a mini-
character, and forec+sting enor focuses attention on the essential mum costexpansion plan [84]-[89], [95], [ 108 ]. These
qualities of a spatial f o r d The vnrious approaches to forecasting automatic procedures produce substantial savings over more
are then reviewed, compared, and evaluated on a uniform basis. Em- traditional design methods,butrequireaccurateprojections
phasisisontheuseoftwo-dimeMionnlsignnltheoryrsaunifonn~e-
work for analysis of load, growth,m , and forecast model behavior. of future small arealoads [49], [ 6 6 ] , 1831, [ S I , [94]. The
desire t o gain maximum benefit from these procedures has fur-
ther spurred development of improved forecasting methods.
I.INTRODUCTION AND OVERVIEW In this paper we review the major aspects of spatial load fore-
N ELEClrRIC POWER system consists of a large number casting, concentrating on the needs and methods used in mod-

A of interconnected components that work in concert to e m powersystem planning. For this reason we concentrate
generateand deliver electricalpower to many points almost exclusively on computerized techniques that
scattered over a wide geographical area. Generation equipment available from modem utilityload and customer databases. We
utilize data

is generally concentrated at only a few large power stations. first briefly summarize planning and design of power delivery
The power delivery system,however, is composed of many systems in order to identify the needs for and uses of a spa-
pieces of equipment scattered throughout the utility's service tial load forecast. Then we discuss electric load and load
area, all sharing the burden of power distribution. Each unit growth, concentratingonthosefactors of load growth that
of the delivery system must be properlylocatedand sized are important in forecasting, particularly coincidence of load,
to serve the demand in the vicinity. Planning the future expan- spatial resolution,growth characteristics, timeperiod, and
sion of this system involves determining both capacities and forecasting error. In the final section, the bulk of this paper,
locations of future components. A first step in this planning we present the various types of forecastmethodsfromthe
is the forecast of the future electric demand with geographic perspectives established earlier and summarize each technique,
detail sufficient to plan equipment sizes and locations. Such its application, and shortcomings. We have attempted to ob-
load forecasting is called spatial, or small area, load forecasting, jectively review and compare the various methods of forecasting
and is the subject of this paper. with respect to one another and the planning needs of electric
The objectives of power system planning are to provide an distribution systems.
orderly and economic expansion to meet the electric utility's In determining typical applications and usages, and in evalu-
future electrical demand with an acceptable level of reliability. ating the state of the art, we are relying heavily on our own
The components of the delivery system-subtransmission lines, experience as well as four recent surveys of electric utility plan-
substations,distribution feeders,and laterals-generally have ning practices [ 901 - [ 93 ]. Where applicable, we have used the
capacities many orders of magnitude less than the load of the terminology developed under ElectricPower ResearchInstitute
entire system. Planning of this system involves determining (EPRI) project RP-570 [49]. Where EPRI has not developed
thecorrect sizes, locations, interconnections,and timing of appropriate terminology, we use our own.
future additions to this equipment. Such planning is a difficult
11. PLANNING AND FORECASTING NEEDS
task, compounded by recent trends of tightening design mar-
gins, longer equipment lead times, and increasing regulatory The primary purpose of power delivery planning is to deter-
scrutiny.These trends have resulted in increasing power in- mine an orderly expansion of the present system in order to
dustry emphasis on delivery system planning. Correct planning meet future demand and to forecast budget needs for financial
of future delivery equipment requires a forecast of the future planning. Planning falls into the two broadcategories of short-
geographic distribution of electric demand in a manner suffi- range and long-range analysis, which differ in both timespan
cient to differentiate between available alternatives of capacity, and goals. Short-rangeplanning has been the traditional ap-
location, and interconnection. The quality and accuracy of this proach for distributionsystems and is done only far enough in-
forecast has a large influence on the quality of the subsequent to the futureto cover equipment delivery and construction lead
delivery system planning. Therefore,methods to accomplish times for the next series of system additions. The goals and
such forecasts have been the subject of increasing attention in motivations of short-range planning are to determine the single
best plan for present construction commitments and to ensure
Manuscript received June 22, 1982;revised December 10, 1982. The that therequired equipment is operational when needed.
submission of this paper was encouraged after the reviewof an advance By contrast, long-range planning is motivated by a desire to
proposal.
The authors are with the Advanced Systems Technology Division of determine economic viability of the short-range commitments.
Westinghouse Electric Corporation,Pittsburgh, PA 15235. Recognizing that any unit of equipment will have a useful life-

0018-9219/83/0200-0232$01.00 0 1983 IEEE


WILLIS AND NORTHCOTE-GREEN: SPATIAL ELECTRIC LOAD FORECASTING 233

Fig. 1. A spatial load forecast locates electric demand geographically,


as shown in this load density grey scale plot of electric demand for a
25-by-25-mi part of a large metropolitan area.

Year 1 Year 10 Year 20


Fig. 2. Electric load grows in both magnitude and location as shown
here in displays of electric load ( Z axis) for Panama City, Panama, for
the years 1980, 1990, and 2000. In these diagrams the PaciTk Ocean
is in the right portion of the plot, the Panama Canal runs along the
4
lower left edge. Small area size is k m 2 .

time of many years, long-range planning attempts toassure that locations of load, and each perhaps requiring a different de-
all elements of the short-rangeplan are efficientlyutilized livery system design.
throughout their lifetime, not just in the first few years after The planning of most power systems is performedon an
installation. Therefore, long-range planning concentrates on a annual basis in anticipation of annual peak load. Spatial load
time period beginning at equipment lead time and extending analysis is accomplishedby dividing a utility system into a
intothefuture.Theuncertainty associated with long-range number of small areas and forecasting the load in each. Loads
forecasting is minimized by using a number of individual fore- for planning are typicallyreferred to on an annual basis by
casts, each a reasonable representation of future load growth. small area
These forecasts are used in a multiple-scenario study of long-
range needs. Short-range commitments can be studied to assure I&) = peak annual load for small area k, in year t ,
that they fit well into all reasonable long-range scenarios. There- k = I to k small areas.
fore, the forecasting for long-range planning needs a multiple
<t.
h

scenario capability-the ability to produce a set of reasonable Zk(t, t o ) = estimate of l k ( t ) ,made from data fromyear to
forecasts that cover theuncertainty of future load growth. In some cases, the small areas used for geographic analysis
The need for short-range planning is for a single most accurate are irregular in shape and size, corresponding to the service
forecast as a guide to determine construction commitments. areas assigned to particular delivery system components such
Spatial load forecasting involves prediction of both themagni- as substations or feeders. However, most modem techniques
tudes and locations of futureelectric load. The analysis of use an M X N grid of square cells, indexed by n , rn location,
location must be made with sufficient geographic resolution to that cover the region to be studied.
permit siting and sizing of delivery system components. A
spatial forecast is therefore not only aforecast of a certain In,&) = lk(t), where k = n + ( m- 1) * N ,
total magnitude of load but also a forecast of its geographic where m = l , M and n = 1,N.
distribution. The Same total could be distributed differently
in a variety of spatial forecasts-each a unique pattern of the A cellular spatial load distribution is shown in Figs. 1 and 2.
2 34 PROCEEDINGS
VOL. IEEE,
OF THE 71, NO. 2, FEBRUARY 1983

The planning of a power systems expansion always includes


a forecast of the spatial load distribution because any plan for
expansion (or no expansion) is based on some set of assump-
tions about future load growth, however poorly defied, cursory
or implicit. Surveys indicate a fairly even split between use of
a uniform grid and equipment-oriented area definitions. Area
size varies depending on application. Currently, most U.S. utili-
ties use 640-acre (mi2) or 160-acre (quarter-section) cells, or
irregularly sized areas associated with distribution substations.
The average short-rangeforecastperiod is four years ahead;
that of long-range planning eight to twenty years ahead.
The first spatial forecast methods using a digital computer
were developed andappliedin thelate 1950s. These were
done on an irregular, equipment area basis using multiple re-
gression curve fitting to extrapolate historical load trends. 0.f 10 1W loo0 10000

Computerized land-use based methods were first developed in NUMBER OF CUSTOMERS


the early 1960s, initially as methods in which the computer Fig. 3. A coincidence curve for sizing equipment gives the multiplier to
was used merely to storeand output forecasts based totally on use in determining peak load based on the number of customers.
user intuition [ 1 1J , [ 15J , [ 17J . Land-use based methods em-
ploy an analysis of the mix of customer types (land use) and For this reason the peak total system load occurs when the
density on an area by area basis. Not surprisingly, as the data- combination of customer demands is at its highest. Some cus-
base management, analytical, and data-handling speed of com- tomers may be demanding only a part of their peak demand at
puters has evolved, the methodology and quality of forecasting that instant. The ratio of the peak total system demand and
has kept pace. Development of computerized forecast methods the sum of the individual customer peak is called the coinci-
has accelerated in the last five years, driven by the availability dence factor C,
of optimization-based automatic distribution design programs
[ 49 J , [ 841 - [ 891, [ 951, requiring a high-resolution spatial load c, = sum ofpeak system load
customer peak loads
(1)
forecast. The largest single impetus in this field was EPRIs *

projectRP-570 from 1975 to 1979 [49]. This investigated a For most power systems this ratio is in the range of 0.3-0.7.
wide range of forecast methods, established a uniform termi- This apparently simple concept has subtle and far-reaching
nology, and identified many of the fields major concepts and consequences regarding the power systems analysis [ 791 - [ 82 I ,
priorities. [ 84 J - [ 921. The peak load on any piece of equipment occurs
Almost all modem spatialforecast methods areallocation when the combination of the demands of all the customers it
methods that distribute on a small area basis a previously fore- serves is at a maximum. This peak may not occur at the same
cast total system load which is an input to the forecast pro- time as the total system peak, or as the peaks of any other
cedure. There are two compelling reasons why this is done. equipment. For this reason, the sum of all feeder peak loads
First, considerable sophistication and effort is used by the aver- usually far exceeds the total system load.
age electric utility in projecting its total system load growth Coincidence factors are used in much of the analysis and de-
[ 75 J -[ 8 2 J . This attention to theforecast of the total load is sign of equipment loads because utility engineers normally will
motivated by the utilitys financial planning which needsa have timely, reliable load readings on only major equipment.
projection of electric power sales revenues. Also, forecasts of Values used in the engineering and planning of much of the
total system load are needed for the planning of generation. smaller equipment must be estimated in one manner or another
Consequently, the effort put into such forecasts is high, and from available data. Most electricutilities size distribution
the results from an engineering standpoint are excellent. While equipment with ademandcoincidence curve similar to that
total system forecasts do not spatially locate the load growth, shown in Fig. 3. The curve is based on observed behavior
they typically stratify, or break down, the forecast on a rate or probabilistic assumptions about the distributionof customer
or customer-class basis. Often theyinclude very detailed specif- peak demands andoff-peakdemands. As a function of the
ics on the expected average per-customer use of electricity. number of customers involved, this curve gives a coincidence
There is no need for a spatial forecast procedure to repeat this multiplier that is used to size equipment. For instance, if a
analysis. Second, within most utilities the various engineering planned feeder to a new subdivision were to serve 150 homes,
plans will be approved by management only if they are com- it would be sized as follows:
patiblewiththecompany financial plan-a plan developed
directly from the companys revenue forecast. Plans developed total load = sum of individual estimated household
from another, independent total system forecast would have a peaks for 150 houses
nearly hopeless uphill battle forapproval. feedercapacity = C( 150) * total load + margin (2)
111. LOAD AND LOAD GROWTHBEHAVIOR where C ( 150) is the coincidence value from the curve for 150
customers. The margin would be dictated by the utilitys plan-
Coincidence of Load
ning policies and design standards. If ten laterals will distribute
The electric load on a power system is the composite demand the feeder load among the 150 homes, each would be sized in
of many electrical devices owned by many different customers. a similar manner, using on theaverage C( 15) insteadof C( 150).
An important aspect of electric load is that customers do not Note that since C ( 15) > C( 1501, the sum of the lateralcapaci-
simultaneously demand their peak power; there is a great diver- ties would exceed the feeder capacity.
sity of when people, businesses, and devices demand power. Few spatial load modelsexplicitlyaddresscoincidence of
WILLIS AND NORTHCOTE-GREEN:SPATIALELECTRIC LOAD FORECASTING 235

SEVERALSQUAREMILES
40001
TOTAL SYSTEM LOAD

/
. .
.......
......
. . ......
........ .
........
. .
.....
I----
I

...
7

Fig.4. As a typical electric utility systemis subdivided into smaller and


smaller areas, its load growth behavior becomes less and less a smooth
continuous process, and more of the S-curve behavior,characterized
by a sharp, brief period of growth.

load in the manner that we have discussed above. Coincidence tric households. Growth of the numberof customers causes the
is acknowledged in one of two ways. First, it may simply be spread of electric load into areas that previously had no electric
ignored; the load-value used for a small area is the best avaii- load- cells that were vacant from the power systems stand-
able data on the annual peak demand in that area.Ignoring point. Changes in usage per customer occur simultaneously and
coincidence has no adverse consequences if the small area size largely independently of any change in the numberof custom-
is large enough that a typical small area contains at least several ers. Except for areas of urban renewal, over the past decade,
hundred customers. The coincidence factor for that number there have been only slight increases in usage per customer in
of customers will be very close to C,, thus a small areas peak most US. electric utilities. This trend is a result of increasing
load will be approximatelyits share of system peak, and prices, conservation, and improvementsinthe efficiency of
very little error is introduced by ignoring coincidence. electrical usage. A number of comprehensive methods exist to
Alternatively, the values used may include an explicit adjust- assess the changes in customer usage due to such factors. These
ment to account for coincidence. methods are adequately covered in current literature[ 75 ] -[ 821
and will be discussed in this paper onlyto the extent that they
Z&) used for area k = (C(e)/C(a)) impact spatial forecast models. Here the character of spatial
* (estimated peak load in cell k) (3) load growth will be examined intwo ways-analysis of temporal
growth trendsand analysis of load level transitions.
where a is the average number of customers in a cell, e is the When viewed from asystem total basis, a growing power
average number of customers in a substation or feeder service
system generally exhibits a smooth, continuous trendof annual
area(whichever equipment level is being planned fromthe peak load growth. Slight variations are caused by weather and
load forecast). minor factors thatvary from one yearto the next.By contrast,
This is the most common manner of addressing load coin-
the behavior of growth in a typical small area is not a smooth
cidence in modem spatial load models. The recent trend to-
curve but is more like that shown in Fig. 4-a sharp Ogive, or
ward much higher spatial resolution has made the effects of
Gompertz curve, commonly referred to as an S curve [ 2 9 ] ,
load coincidence significant. For instance, 10-acre cells often
[42], [49], [ 581. The majority of the small area load growth
contain only 10 or 20 customers. The coincidence factor for
occurs during only a few years, during which load goes from a
numbers of customersin thisrange is not close to C,. As
near zero t o a value close to a final saturated peak load. The
noted earlier, the crux of distribution planning is the overall
smooth constant trend of the total system load is due to it
system design and the sizing of its major components.An
being the aggregate of many small areas whose bursts of
estimate of the peak load in any one small area is not as im-
growth occur at differenttimes.
portant as that areas role in the overall system. By using such
Fig. 4 shows growth curve averages for a large North Ameri-
loads on a small area basis, any number of I k ( t ) can be directly
can city. As the cell size used for analysis is reduced, the typical
added to form load estimates for any arbitrary area. If the sole
growth curve shape gradually becomes less a smooth constant
purpose of the planning is to plan for large equipment (that
with over 500 customers), this is valid because the coincidence trend and morea sharperS curve shape. Typical growthcurve
curve (Fig. 3) is flat for thelarge numbers of customers served shape is a function of spatial resolution. Of course, each small
by major units of equipment. area will differ-most slightly, and a few significantly-from the
average growth behavior. The quantitative aspects-the average
rise time and its changes as a function of the resolution-differ
Load Growth Character
fromoneutility system to another.This basic behavior is
Load growth of a power system is due to two simultaneous observable in nearly every power system [49], [ 581, [64].
processes. First, growth is driven by increases in the number A second perspective of electric load growth can be studied
of customers in the utilityservice area. Second, it can be caused with a matrix approach to transitions among various levels of
by increases in the average demand usage per customer. New load [ 571. Define a set of load levels that partition the small
customers are added to a system due to migration into an area area load levels, the following sixteen-level set defined on the
(population growth) or theelectrification of previously nonelec- last year t 2 of a period t l , t 2
236 PROCEEDINGS OF THE IEEE,VOL. 71, NO. 2, FEBRUARY 1983

(4 (f) (9) 01)


Fig. 5. Transition matrices for electric load growth as observed at vari-
ous resolutions and time periods in a large city. (a) 2560 acres-2 yn.
(b) 2560 acres-4 y n . (c) 2560 acres4 yrs. (d) 2560 acres-8 yrs. (e)
640 acres-4 yrs. (f) 160 acres-4 yrs. (g) 40 acres-4 p.(h) 4 0 acres-
8 yn.

916 = load level exceeded


by of all cells Similarly, the
totalgrowth due to all cells that ended the
period in level i is the row sum
16
growth from level i = Ggj . (7)
j=l

Fig. 5 shows G matrices for a large metropolitan area with a


For any year t , all small areas in a utility system can be classi- fairly high growth rate. These matrices are displayed as grey
fied by their load l k ( t ) into one of these 16 classes. A transition scale plots of the element values to facilitate the easy recogni-
matrix of movement of small areas from one level to others tion of their essentialqualities. The form of thesematrices
during any particular period of time can be formed as a Q by varies considerably depending on both the cell size and time
Q matrix with elements period, reflecting the change in the character of growth as a
function of both cell size and time period. They highlight an
T i j = number of small areas that moved from interesting quality of spatial load growth;from the standpoint
level i to level j during theperiod from or the character of observed load growth, reducing cell size
year tl to year t z . and lengthening time period are equivalent.
Fig. 5(a) shows G matrices of growthin four-square mile
From this, a matrix useful in the diagnosis of the forecast areas over a two-year period. This matrix is diffuse, indicating
method applicability can be formed as that the majorityof growth occurred due to transitions of cells
G ~=JTi,i* (2, - Qi) = number of cells that moved from from among many different load levels-particularly in the in-
level i to level j , weighted by the termediate levels. Columnone-growth due to development
difference in average load be- of load in vacant areas-is nearly empty, indicating little load
tweenthetwo levels (Qi - ai). (5) growth in previously vacant areas. Little growth occurs in the
rows indicating growth to the higher load levels, the final or
The G matrix haselement values that, neglecting load-level saturation levels. If the time period is lengthened, as shown
quantization errors,are theportion of load growth due to in Fig. 5(b), (c), and (d), the formof the matrix changes. Less
transitions among thevarious levels. Total system load growth of the total growth is due to transitionsamong the intermediate
from year t l to t 2 is levels; more is due to transitions of cells into the higher load
levels. There is significant growth due to transitions of cells
that began the period with no load, as evidenced by the first
column of the matrix.
A similar change in character occurs if the time period is
The growth over the period, due only to growth of all cells held constant but the cell size is reduced. Fig. 5(a), (e), (f),
that started the period at level j is a column sum and (g) show, respectively, four-year G matrices for 2560,640,
160,and 40 acreresolutions. If cell size is reduced while
16
growth
from
leveli = Ggj . (6) period length is held constant, more growth occurs as transi-
i=l tions to the final load levels, andcolumn1-growth of cells
WILLIS AND NORTHCOTE-GREEN: SPATIAL ELECTRIC LOAD FORECASTING 231

N
t LLW
1 mile
u
1
-.
km. WL
+

--
- --
++-+
-.- --+.
-. .++
.
- - + - + -+
+-

(a) (b)
Fig. 6. The spatial distribution of error in three different forecasts of
the same group of forty-acre small areas. Flus and minus signs indicate
over or under forecasts in each area; the size of the sign gives an indi-
cation of magnitude. Broad lines indicate approximate boundaries of
substation service areas; thin lines indicate feeder areas. All three
forecasts were produced with Houston Lighting and Power Companys
ELUFANT forecast model, with variations to the parameter setup of
its high-frequency suitability model [ 331.

that began the period nowith load-becomes significant. The TABLE I


predominant growth is incolumn 1-growth of previously EVALUATION OF ERRORIMPACT BY THREE
DIFFERENTMEASURESVERSUS
I M P A C ON
~ SUBSTATION AND FEEDERPLANNING,
MEASUREDBY ACTUAL
vacant areas. reveals that from the perspective Of A N A L Y SOF~ THREE
DIFFERENT FORECASTS
AND THE POWERS Y ~ M
observed behavior, increasing spatial resolution or increasing
the time
muchperiod has same effect.
These two different ways to study load growth-curve shapes
ERROR
UEASURE - FORECAST
-A B - C

transition
and matrices-reveal the basic change character
in the R.M.S. 141
. 01
43. 4. 4 4
of the load growth as the resolution and timeperiod are varied. Average Absol. Value 1 2 . 31 1 . 31 2 . 3
Impact on Substations
At afairly low spatial resolution load growth is generally a measured by examinatiol, 9.8 2.6 2.4
gradual growth of each area, lasting many years, with few pre- measured by eq.9 2 . 6 2.5 10.0
viously vacantareas developing anyload. At higher resolu- -ct on Feeder iaads
measured by examination 12.0 3.5 8.2
tions growth behavior is much sharper and briefer, with many measured by eq. 9 12.1 3.7 8.2
cases of previously vacant cells developing load growth. If the
time period of observation is longcompared to the average
cellular rise time, then most of the growth will appear to be developed based on an actual forecast, the planner is of course
transitions to the final load levels, and will be mostly due to unaware of the magnitudes and locations of the errors in the
development of previously vacant areas. forecast. Define the error as the difference between the esti-
The two methods given above allow the applicability of vari- mated and actual load on a small area basis
ous forecast methodsto be studied quantitatively for any partic-
ular situation. For example, traditional load forecast methods
such as curve fitting and other trending methodsare incapable
of forecasting development in presently vacant areas. The col- = error in forecast load for area n, m . (8)
umn 1 sum of the transition matrix for a particular problem A number of traditional error measures have been applied to
is theamount of vacant area growththatsuchtrending the en,m including the rms, average absolute value, and Mean
methodscannot predict. Somemodemsimulationmethods and Variance and the68percent Rule developed under
work with asimplifying assumptionthatgrowthoccurs as
EPRI sponsorship [49]. Thesevariousmeasures have given
transitions directly from the vacant to a final level [ 361, [ 431.
conflicting, inconsistent rankings of different forecast methods
In cases where most growth appears in column 1 and the higher applied to the sameproblem I441. More importantly, none
load level rows, such an assumptioncauses little error. correspond to theerrorimpact when assessed bydetailed
study of the actual impact on planning [471, [ 63 1 . Tradi-
Analysis of Spatial Forecast Error tional statistical measures do not properly assess error measure
Before turning to the forecast methods themselves, we wish because they are not sensitive to the spatial qualities of the
to briefly examine the quantitative measurement of error in a errors, while a power system is.
load forecast and its relationship to the planning of a power Fig. 6 shows the forecast errors for the same portion of a
system.Generally,forecast methods are tested bypredicting utilitysystem,fromthreedifferent forecasts.Eachforecast
present small area peak loads, using data from a past year, per- has roughly 13.5-percent rms error, as shown in Table I. In-
haps ten years before. The accuracy in forecasting present loads spection of the three forecasts will show that they obviously
is taken as an indication of probable accuracy in forecasting have a far different impact onplanning. The forecast shown in
future loads over an equivalent future time period [ 231 - [ 271, Fig. 6(b) has errors distributed in a manner such that positive
[291,[331,[371-[421. When asystemexpansionplan is and negative errors usually cancel almost entirely within any
238 PROCEEDINGS OF THEIEEE, VOL. 71, NO. 2 , FEBRUARY 1983

1
predicted
actual
-
0
FEEDERS

W
vl
iL
0
n
vl
Y
a

0 01
0 5 1
FREPUENCY - Cyclewmi.
Fig. 7. Results of tests comparing actual impact of error on the estima-
tion of feeder and substation loads, with the impact as estimated by
signal analysis of system error sensitivity. Spatial frequencies repre-
senting error were added to small area load data and the consequent
impact measured and recorded 88 shown. Predicted response is the
normalized spectrumof the Pdd) fdter.

areas the size of feeder or substationareas. Thus errors in fore- This function, P,(x, y ) , is a circularly symmetric function and
cast (b) would have little adverse impact on planning feeder embodies the overall design characteristics of the substation
and substation capacities. When its loads are added manually design of the system. Actual impact of the cells errors on sub-
or by computer to form the equipment load estimate for any station load estimates is closely approximated by
reasonable set of feeder orsubstation areas (not just those
shown), most of the errors cancel leaving a neterror close N M
to zero; hence, errors of this type have little adverse impact
on planning. As shown in Table I, these errors cause less than
2-3 percent error in estimating substation loads, as determined
by detailed examination.
Forecast errors distributed as shown in Fig. 6(a) do notsimi-
larly cancel. Positive and negative errors tend to clump to-
gether in large contiguous areas, whose size in general is equal
to or larger than equipmentservice areas. The net effect is that a
substantial amount of error accumulates in feeder and substa-
tion areas. A forecast method that spatially distributed error (10)
in this manner could be very damaging to planning. The fore- where
cast shown in Fig. 6(c) is an example of a forecast that has
strong impact on feeders and little impact on substations. In P,(d)=O, forall d > P .
this forecast the errors are distributedso that they largely can-
cel within substation areas, but on the average only partially The Usvalue usually corresponds closely to the actual impact
cancel over areas the size of feeder areas. This forecast would of the errors determined by examination, as shown in Table 1.
mislead the planning of feeder capacities but generally lead to The P,(d) filter derived above from the substation area struc-
correct sizing of substations. ture, and others similarly derived for transmission,subtrans-
The difference in impacts of the three forecast error sets dis- mission, andfeeders give the response to spatialforecast
cussed above reveals an interesting aspect of system planning. errors of the planning of equipment capacities. Agreement
From the standpoint of spatial error impact, the power system with observed behavior is usually excellent and has been veri-
planning is a spatial low-pass filter. High spatial frequencies of fied on a number of systems as shown in Fig. 7 [63]. All such
the load forecast error, as in Fig. 6(b), have no impact, but filters are low pass; specifics vary from one system to another
lowerfrequencies, as in Fig. 6(a), do have an impact.Small depending on equipment specifications, design standards, and
area electric load can be viewed as a signal function of space. geography.
When the cells used in the small area analysis are the cells of a A load forecast must contain an accurate representation of
uniform grid the Zn, m ( t ) values are uniform discrete samples all the spatial frequencies to which the design is sensitive. The
of the spatial electric load signal. These values are not strictly spatial sampling of the electric load must be done with a rate
samples of the load level at the particular points, but the aggre- sufficient to represent the frequencies to which the design is
gate of electric load in each cell. This type of data is roughly sensitive. The filters spectra can be used along with the Nyquist
equivalent to that obtained by passing the spatial load signal criteria to determine the required forecast cell size, the sam-
Z(t) through a spatial low-pass filter, and sampling the output pling rate, needed to accurately represent the loads. Table I1
of the filter [ 1001. Aset of two-dimensional fiiters, whose shows sampling rates calculated for a large metropolitan sys-
derivation is fully explained elsewhere can measure impact on tem. As would be expected, the design of the smaller parts of
the particular level of a power system. For instance, for sub- the system(feeders)requiressubstantially more resolution
stations thefilters space domain representationis than that required for larger equipments.
The preceding discussion of error shows that the forecast of
P,(d) = probability that a randomly selected cell a distance the individual loads is not as important to proper powersystem
d away from a particular point is served by the sub- planning as the assessment of some of the overall spatial aspects
station serving that point. PJx, y ) is a twodimen- of the load distribution. For this reason, the authors refer to
sional function P,(~W). forecasting of loads for distribution planning asspatial, rather
WILLIS AND NORTHCOTE-GREEN: SPATIAL ELECTRIC LOAD FORECASTING 239

I I
Fig. 8. Two scenarios of electric load growth for a major city. Both rep-
resent a forecast of the same total amount of load but differ in the pat-
tern of its geographic location (most of the difference is in the lower
right portions of the patterns). The difference in the two patterns is
due to modeling of a new highway as along either of two proposed
alternate routes.

TABLE I1 not sufficiently identify problems caused by localized areas of


CELLSIZESNEEDEDFOR ADEQUATE
LOADMODELINGFOR THE VARIOUS high density load. Using expectations gives load patterns much
LEVEUOF A LARGE METROPOLITAN
POWER SYSTEM
smootherthannormal; essentially low-pass spatial-filtered
Lcvel of System Area Sizes Needed patterns. The various scenario load setsmust be carried through
2 the planning process individually, as shown in Fig. 8. Multiple
PIV hansmission 4 sq. miles (10 kn 1 scenarios will be discussed in more detail later in this article.
2
Transmission 1 sq. mile (2.5 km 1

SUbSatiOnS 160 acres (.63 kni 1


2 IV. LOAD FORECASTING
METHODS
2
Feeders 71 acres (.28 km ) In this section the defiitions and concepts developed earlier
will be used in a discussion of the various spatial forecasting
procedures. Spatial load forecast methods can be categorized
than just small area, forecasting. This highlights the need for by several characteristics. These categories are: 1) whether or
an integrated assessment of the entire load pattern. not they analyze historical data in the course of the forecast;
The discussion above is a workable definition of error assess- 2) the type of grid or area distinction; and 3) the type of fore-
ment, or quality, etc., when only one forecast is compared to cast algorithm. AnaZytic forecast methods perform analysis
actual loads. How does one measure the error inaset of of past or present data, identifying trends and patterns that
multiplescenarioforecasts? If one of thesealternative fore- are then used to project future loadgrowth. By contrast,
casts were absolutely correct, all theothers would by their nonanalytic methods are approaches that use a computer only
very nature be incorrect. Criteria for multi-scenario forecasts for the record keeping and report generation of an intuitively
are much harder to define. A common misconception is that generated forecast. Most of the newer forecast techniques use
planning can be done based on the small area expectations. a rectangular grid of uniformly sized small areas as the basis
These expectations are obtained by probabilistically weighting of the forecast. Many methods, particularly thetraditional
the various scenario loads for each cell according to the per- approaches, use irregularly sized areas defined by the service
ceived likelihood of each scenario, as areas assigned to different members of some equipment level,
such as substations or feeders.
I
E ( l k ( t ) )= ( P i zk(i, t ) ) pj (1 1) Fig. 9 shows the overall relationship of the spatial forecast
i=1 j
/ I= 1 algorithmswhich wewill discuss inthissection.Analytic
methods fall into several categories. The simplest in terms of
where there are I scenarios and pi = probability of scenario i, data and method are the various algorithms that extrapolate
Ik(i, t ) = forecast load for year t in cell k, for scenario i. Using past small area load growth, generally called trending methods
such E(Zk(t)) is invalid, as demonstrated by a simple example. [SI, [91, [ 101,1141,[251, W I , [@I, 1581,
[291, [ M I ,[691.
Suppose there are two scenarios, each representing theaddition Some are quite novel. Multivariate methods work with other
to the future loads of a major factory, which could go at either data in additionto small area load data [ 271, [ 281, [49]. The
of two widely separated but equally likely sites. Each of the types of possible algorithms vary considerably, but the term
two scenariosgenerated from assumptions thatthefactory multivariate alone has generally been used for methods that
would locate at either siteis a valid, reasonable forecast of what extrapolate simultaneously a number of variables on a small
might happen. Combining the loads probabilistically results in area basis. A very successful class of multivariable methods are
a scenario with half a factory at either site-something that Zand-use based methods that employ one or another form of
is quite unlikely. In general, combining small area loads in a urban system land-use simulation to predict future electric load
probabilistic manner results in a geographic pattern that does ~ ~ l , ~ ~ I , ~ ~ ~ l , ~ ~ ~ l , ~ ~ ~ l , ~ ~ ~
240 PROCEEDINGS OF THE IEEE, VOL. 71, NO. 2, FEBRUARY 1983

METHODS THAT METHODS THAT


NON-ANALYTIC WORK ONLY WITH ANALYTIC WORK WITH MORE

MULTIVARIATE

FILTERING
2 3-3 OR 3-3-3

Fig. 9. The different classes of spatial electric load forecast method are
shown here, plotted by general level of forecast quality (accuracy,
flexibility) vertically, and complexity (ease of use, data needs, resource
requirements) horizontally. Tree structure show the family relation-
ship of various methods, following the classification of methodology
described in the text. Year given
- for each method is the authors esti-
mate of its fvst use.

[431-[511, [601, [651, [661. Although these are technically specified system total load leveL Results from such methods
multivariate methods, they are generally regarded as a separate are generally poor, but requiredresources are minimal.
class because they often do not extrapolate in the strictest
sense. Trending
Trending is a generic term that encompasses all methods that
Nonunalytic Methods
extrapolate loadbased on pastloadvalues.Such methods
Nonumlytic methods for spatial load forecasting rely almost have two advantages; they employ simple, direct algorithms,
completely on user intuition. A popular method, dubbed the andtheyrequirea minimum of data.Nearlyevery utility
coloring bookapproach, will bedescribedhere as typical maintains records of major equipment loadings, so the required
of all such methods [ 71, [ 101, [ 181. A map of the utilitys data are always available at least on an equipment area basis.
service area is divided by a series of grid lines into 160 acre The most obvious trending method involves curve fitting t o the
cells. The amount of existing land use in several classes, for historical loads on an area-by-area basis. A K area forecast is
example, residential, apartments, commercial, and industrial, done as K independent extrapolations, using multiple regres-
is placed on themap by literally coloring each cell with colored sion to fit a function to each cell (area), typically
pencils,adifferent colorindicatingeach class. Future land- A

use, obtained from user intuition, is similarly coded onto the I k ( t ) = f k ( t , t 2 , t 3 , r-l) (12)
map. An average per-acre load density for each land use class or
is developedfromexperience,survey data, and the utilitys
h
overall system load forecast. These densities are used t o manu- =fk(log t , (log o, (log 0 3 ) . (13)
ally convert the coloredlanduses into aload for each cell.
Despite the apparent crudeness, this method or a variationhas Applied on a low-resolution, short-term forecast basis (typi-
been widely used as a long-range planning tool by over one cally on l-squaremileorlargerareasforthree-to-fiveyear
fourth of all utilities surveyed. forecast periods) multiple regression gives useful results. When
It is a natural stept o employ a computer as the data manipu- applied t o areas smaller than 1 square mile, curve fit and ex-
lation and reporting medium for such a scheme. A number of trapolationmethods haveproblemsdealingwiththe sharp
techniques doing just this were developed in the 1960s and SCurveshapestypicalof small areas1421, [49],[58].
early 1970s [ 8 ] , [ 111, [ 161, [ 171, [45]. Use of a computer Equipment serviceareas-theusualsource of anonuniform
allowed expansion of the methods in a numberof ways. More area basis-are time varying due to transfer of loads from one
classes of load, up t o twenty, could be used. A common fea- feeder t o another that occur when operating personnel reswitch
ture was the ability to scale or adjust the total growth to meet loads and feeder segments, or when new construction is done.
WILLIS AND NORTHCOTE-GREEN:SPATIALELECTRICLOADFORECASTING 241

Thesetransferscausevariations in recordedloads,creating
substantial error in trend identification.
Most applications of trending use a horizon yearload esti-
mate for eacharea
A
l&) = estimated eventual loadlevel for cell k, in year h,
a year well beyond end of forecast period.
The function f k ( t ) is fitted to both the historical data values
and this horizon year load. The contribution of such horizon
year loads t o improved forecasts has been well documented RESENT
[ 291 , [ 491 , [ 581. Short-term, low-resolution regression-based
forecasts are fairly robust as regards these horizon year esti-
4 . . . . .FUTUlE . .
...... . . . ,
YEAR
mates; results are not highly sensitive t o reasonable levels of Fig. 10. Clustet curve method fits a template, selected from a set of
error in the horizon year estimates [ 581, [64]. typical growth curve shapes, to each small areas load history, shifting
Aconsiderable amount of researchhasbeendevoted to the template t o minimize mismatch with the load history. The tem-
plate then provides the forecast. The method f M solves for the tem-
regression-basedtrending.Meinke [ 421performedextensive plates based on a cluster analysisof past histories of all cells.
tests ofhistoricaldataperiod,functiontype,andforecast
period. Similar tests of extrapolation sponsored by EPRI [251,
partially successful at forecasting load growth in cells with no
[29] reached roughly the same conclusions: trending is appli-
past or present electric load. Future growth in a vacant area
cable to short-term ( 5 years or less) forecasting, and the best
can often be correctly inferred by observing that a regions
extrapolation for most applicationsis a cubic logarithmic func- smooth, continuous load growth can continue only if vacant
tion (( 13) above) fitted t o only the most recent five or six areaswithintheregionthenbegin to grow.Fig. 11 shows
years of data. These studies were all at relatively low spatial a square mile divided into four areas, three of which have a
resolutions of 640or520 acres. Whethertheseconclusions
load. These three subareas can be trended-they each have a
are valid at higher resolutions hasnot been determined.
load history. The fourth area cannot be trended-its load his-
The industry surveys discussed earlier indicate a wide variety tory is zero. The entire square mile can be trended-usinga
of slight variations in applications of regression-based methods. horizon year load that is the sum of the four suburea horizon
Variationshaveincludedheuristicrulesforignoringbad year loads. The trend for the square mile can be compared t o
data points and numerous adjustment schemes for scaling the the sum of the trends of the three subareas that were individu-
sum of all trends to equal a global control total. Most adjust-
ally trended.The difference is the load attributable to the
ment methods fall into one of the three categories covered in fourth(vacant) area. Computer programsapplyingthis VAI
EPRIsponsoredresearch:1)adjustmentproportional to the concept have been appliedto bothregression-based and cluster/
area loads themselves; 2) adjustment proportional t o the small templatetrending [58], [67]. Extensivetestinghasproven
areagrowth or; 3) adjustmentproportionaltothecurvefit
that it is much more accurate than cell-by-cell regression curve
residuals (49).
Recall that earlier the discussion of the growth transition
3 3
fit, with typically only t o the average absolute error.
Since the early 19603, a number of trending methods have
matrices showed that for small cell sizes, or long forecast pe-
used a threedimensional function t o fit to thesmall area loads
riods, asignificant portion of growthoccurs in previously
vacant areas (the column 1 sum of the G matrices in Fig. 5). load = f(x, y ) , where x, y index the small area locations.
Trending on an individual cell basis cannot forecast develop-
ment of loadwhere thehistorical load is zero.Anonzero (14)
horizon year load for a vacant cell may force a nonzero extra-
polation, but it hardly provides a reasonable forecast. For this Generally,these methods severelyrestrict the form of f by
reasoncurvefittechniquesarelimited t o short-term,low- u priori assumptions about its character. Recognizing a strong
resolution applications, where the amount of load occurring in tendency for higher load density in the central core of a city,
vacant areas is only a small fraction of total growth. several methods use a function that is the sum of one or more
An interesting variation in extrapolation uses a clustering- radiallysymmetric,monotonicallydecreasingfunctions of
template matching method instead of regression [61] , [64]. distance from an urban pole, as shown in Fig. 12. Lazzari f i s t
This method uses a set of about six typical curve shapes, called used this approach in themid 1960s [ 1 1] , [ 131. Many similar
templates, t o forecast load on a cell-by-cell basis. Each cells approaches are now in use[8],[ 161, [34].
recent6-yearloadhistory is shiftedbackand forth against There are three features to such functions: 1) the locationof
eachtemplate.Thebestcombination of templateandshift the urban pole, ( x p ( t ) ,y p ( t ) ) ;2) the central load density, or
determines the cells forecast, as shown in Fig. 10. A simple pole height, D ( t ) ;and 3) a per-unit function of distance, d,
clustering algorithm is used t o derive the set of templates from from the center, g(d, t ) with g(0, t ) 1.O. As usually applied,
the historical small area data. The calculations are simple and g(d, t ) is amonotonicallydecreasingfunction of distance,
amenable to 16-bitintegerarithmetic,makingthismethod whose radius or point where g(d, t ) = 6 increases with time.
adaptable to the smallest computers. Due to the use of clus- As applied, the volume under the curve expandsin concert with
tering, the methodworks only on a uniform a d basis. the previously projected systemtotal, the central height expands
as necessary t o accommodate increases in load density, and the
Multiple Cell Trending diameter increases to accommodate an expanding city.
The most successful trending methods work not on a cell- For the cell at the center(x&), y p ( t ) )
by-cell basis, butongroups of cells. Asimple method of
multiple cell trending,calledvacantareainference(VAI) is
242 PROCEEDINGS O F THE IEEE, VOL. 71, NO. 2, FEBRUARY 1983

difference
is vacant area

trend of sum of .......


histories

trends

YEAR

/l/il/li/l
Fig. 11. Vacant Area Inference (VAI) techniques trend the load in small
areas that have no load or load history. This is done by comparing the
sum of the individual area trends of a block of areas to the trend of
the sum of their histories, as shown. The difference is the inferred
trend, assigned t o the vacant area. The critical element is the horizon
year estimates, w h i c h causes the difference between the sum of the
trends and the trendof the sum.

\ actu
y--.:
functwn nd

--.,_I *j
r u n t bnd data

.................

h
as applied,perspective P 2.,r....

-hnctM
i -C 1
0
FREOUENCV - CYCLES/YILE .5

Fig. 13. Spectrum of the spatial load distribution for a small city (pop.
220 000), and an urban pole function carefully fitted to the load dis-
tribution. As shown, addition of information on the restricted vacant
areas such as parks and cemetaries improves the match t o actual dis-
tribution. Note that most of the serious mismatch isin the higher
frequencies t o w h i c h planning is not highly sensitive (see Fig. 7).

DISTANCE FROM CITYCENTER tem total (obtained from theutilitys revenue forecast) defines
the forecast total volume of the function. The change in the
Fig. 12. An urban pole function models the tendency of electric load
density to increase in the center of a city. These simple functions are location of the center over the calibration period is extrap-
often very successful at modeling the overall spatial distribution of olated linearly, as is any change in o(t).For any future year
load, which can then be refrned by other means. thesethreeextrapolatedchangescanbeused t o determine
a new diameter,completelydefiningthefunctionsvalues.
and Implicitly,thismethodforecasts the growthofpreviously
vacant areas.
l i j = D ( t ) - g ( d ( x p- i) +( y p- j)2, t). (16) When fitted to past data and extrapolated into the future,
The volume of the function is the total load, on a discrete functions of this type typically produce S-curve shapes for
cellular basis. each cells forecastloadgrowth.The S curves obtained
from such a method can be compared to the observed range
ofcurveshapes, as shown earlier in Fig. 4. They will bea
i = j j=l
reasonable match t o the S-curve behavior at some (usually
large) area size. In a sense, that area size is the resolution limit
An important and easy addition often made t o these meth- of the method; it cannot generate local growth curvesthat are
ods is t o restrict the load in certain cells t o zero, where u priori any sharper.Even thoughitmaybe applied on a higher
information indicates no load growth is possible. Cells repre- resolution basis, the method is really modeling each cell as a
senting such areas as cemeteries, parks, swamps, and land with part of a larger areagrowth curve, and has limited resolution.
sharp gradients, etc., can all be defined as I,.,,,(t) = 0 for all t . Analysisof the spatial frequency content of this function
This results in a much improved forecast [ 341, [651. alsoprovidesinsight into its applicability. Fig. 13 shows the
A forecast of loadusing this technique is straightforward. functions spectrum and that of actual load density for a small
The function is fitted to load values for the most recent year, city to which it has been carefully fitted. The control total
and also for the loads for several years before, finding in each d e f i e s the spatial dc component of load. Use of u priori data
year the center D ( t ) and the form of g ( d , t ) . The forecast sys- on known vacant cells such as cemeteries provides someof the
high frequencies in the methods spectrum. Most of the error uik(f) is variable i in year f, for cell k. I is the total number of
is in the band of spatial frequencies to which substation plan- variables. Some of the data may be load data.
ning (see Fig. 7) is not sensitive. The nature of these additional data variables varies consider-
As the total systemload (the volume under the function) ably among approaches. All multivariate methods differ from
increases, the resolutionof this method viewed by eitherof the trending methods, in that they work as a series of iterations,
two methods describedhere, decreases. When used to cover extending data in time
a larger city, the functions limit of accurate spatial resolu- Vn(t + a ) = G( Vn(t),C ( t + a ) ) (18)
tion decreases. The same numberof parameters and functional
complexity are being used to cover a much larger geographic when V ( t ) is a vector composed of all the u i J t ) and C(f + a )
area; resolution suffers. Since this method loses its resolu- denotes a set of control totals and data. a is the forecast
tion when its size is expanded, there is a largest city or region iteration period, typically 1 to 3 years. G is the multivariate
size for which it is really suited, a function of the resolution forecast function. Multiyear forecasts for periods more than a
needed. Recognizing thatmany larger citieswere farfrom year long are doneas several iterations.
circularly symmetric, several forecast methods use a number The definitive computer program for multivariate forecasting
of different urban pole functions [ 161, [ 341, [ 651. A more was developed by Wilreker, Strintsis,and Long under EPRI
complex geographic distribution of load can then bemodeled, sponsorship and is now widely known within the electric power
and a large city can be modeled while maintaininghigher industry as Multivariate [491. This program is flexible
resolution. and can be set up to perform most types of multivariate load
The authors feel that urban pole trending methods andvaria- analysis. The mathematical formulation is covered elsewhere
tions on themto be discussed later may bethe most undenated in great detail [27], [281. Thecomputer program itself is
form of spatial load forecasting. Their simplicity causes many available from EPRI [491. The concept behind this method is
people to dismiss them without adequate testing. They have simple.
proven particularly suitable for forecasting in small cities with 1) The space is not the I data dimensions. Multivariate small
a well-defined center of town. area data are usually highly collinear. A principal components
rotation transforms each cells vector intoP dimensional space,
Other Trending Techniques P < I . These P dimensions are linearly independent and statis-
An early form of multi-area trending involved disaggregation ticallynormalized, allowing the next steps to be performed
of the total load forecast, by trending area loads against the validly [ 271.
system total load, rather than in time 2) The cells are not individually trended. They are clustered
A into sets. Each cluster is used to develop a linear function of
I k ( t ) =Fk(Z(f), (17)
data)
other movement within the cluster, the average trajectory of its
where L ( t ) is the total systemload (aggregate) forecast for cells. The resulting model is several piecewise (cluster) linear
year t . There were a great many variations of this theme [lo], approximations of a nonlinear process [ 281.
[ 161 , [ 67 I . Results are roughly equivalent to regression-based 3) Once established, these trajectories could be projected for
trending, and depend on the formof solution of the equations many years into the future. Instead, the forecast trends each a
Fk( ) from historical data. Methods that solve for all Fk simul- short distance in the space, representing change over only a
taneously, rather than one ata time, are more accurateand can few years. The forecast proceeds iteratively, recalculating the
be improved by using generalized, or Markov, regression [ 3 5 ] . entire process every few years. The reason is that as cells move
Thereareentire bodies of apparently applicable trending in the space, they could change clusters and hence trajectory
methodology that have to the authors knowledge not been characteristics.
widely applied to spatialload forecasting. These andother 4) At any point in the process the cells vector in P space
methods of trending have not begn vigorously pursued for can be interpreted into electric load using the principal com-
several reasons. First, trending methods cannot relate changes ponents loading matrix and a simple linear function [491.
in load to causal factors, such as the completion of a new This procedure producedreasonable results in a series of tests
highway, changes in theeconomy,etc. Thismeans trend sponsored by EPRI [44], [ 491 when driven by control vari-
methods are not amenable to multiplescenarioforecasting. ables (the C(t + n) in (18)) froma land-use projection program
Secondly, a once-important virtue, economy of data needs, is of thetypeto be discussed later. This land-use projection
not as critical as it was in the past. Many utilities now have provided the information on change from vacant to positive
computerizedtransformer loadmanagementand customer counts of residential, etc., customer type in yeart + a. Viewed
information system data bases that can provide detailed, rele- from the perspectives of load growth dynamics discussed ear-
vant data [ 891 -[ 941. Third, the continuing increase in eco- lier, the method canhandle extreme types of Y c u r v e be-
nomical computational power is allowing practical implemen- haviorbecause of the piecewise linear model based on the
tation of more involved analysis techniques.Thesereasons clusters. By itself, it cannot accommodate vacant area growth,
have focused most development and research along other lines. a major shortcoming. The burden of forecasting vacant areas
is shifted onto the control variable forecast, which almost
Multrivariate Extrapolation totally determines the quality of the procedures long-range,
Multivariate methods of spatial load forecastingmake use high-resolution accuracy,
of more than just the I&) peak annual load values, using in- As small area size is reduced, the number of customer types
stead from one to as many as sixty additional data measure- in an average cell decreases. Viewed at a one-mile resolution,
ments on eachsmall area. These variables form a data vector most cells in a city or town will have a great many types of
for eacharea. load; usually each will have some fraction of load for about 50
percent of all the customer classes but will vary greatly in the
---
W t ) = ( U l , k ( t ) ,U 2 , k ( t ) U&k(t)). mixture in each cell. The clustering method used by Multi-
244 PROCEEDINGS OF THE IEEE, VOL. 71, NO. 2, FEBRUARY 1983

is easy, improving the credibility of the results over that of


other methods, an important point when the forecast is to be
used t o justify major system expansion. The largest advantage
is that analytic land-use techniques are the most accurate and
useful spatial forecast methods available. Their advantage over
other techniques increases with a lengthened forecast period
and increasing spatial resolution [451, 1491, [541, [651.
All modem methodswork on a square-grid basis, with data on
the land-use type and density (amount) in each small area. An-
cillary data on geography and other factors are often included.
There are C classes of land use, c = 1, C , typically 9 < C < 20.
These classes include such distinctions as residential, commer-
cial, and industrial(usually threeorfour subclasses within
each).
Defiie
Fig. 14. Overall structure of a land-use based spatial load forecast
method. & ( t ) = amount of land-use type c in cell k for year t .
These a i ( t ) are asubset of the Uk(t) vector for cell k. The
variate will fiid a rich statistical mix on which to work and additional data vary from one approachto another but usually
will distinguish a number of clusters. As cell size is reduced to include zoning and data on the transportation network.
160,40, or 10acres, an average cell will be composed of fewer, The units of measurement of the ag(t) are usually one of
at most a few, customer types. Thus at high resolutions, the three types: 1) the actual number of customers of type c in
individual VK(t) vectors used will be sparse. Consequently, cell k ; 2) the number of acres of land devoted to class c in cell
clustering will not work. During a test at 25 acres cell size by k ;or 3) the percent of land area in cell k devoted to class c.
the authors, the cluster definitions inMultivariate collapsed All land-use methods work in the same basic manner, differ-
to little more than the customer class definitions. Cells were ing only in the method and sophistication of each of the several
best distinguished only by land-use type and density-distinc- steps shown in Fig. 14. A land-use based method is first used
tions that are clear without theuse of a clustering method. in a calibration phase in which present and historical data are
RecentlyThompsonandGonen have developed a novel analyzed to determine trends and locational patterns. .These
multivariate approach to distribution loadforecasting that trends and patterns are then used to forecast the cellular data
employs tables of production rules, similar in concept to the on land-use type and density on a cell-bycell basis. The result-
developmental Lindenmayer systems, automata, and grammars ing projected land use is then transformed into electric load
used to describe the growth of biological systems [ 521, [62]. withthe application of the usage modelwhichseparately
This multivariate model is novel in that it simultaneously pre- models load of each land-use type.
dicts the growthof both load and thedelivery system network There are three approaches to the conversion of the land-use
configuration. It has given promising resultsinlimited tests data to electric load: 1) a single factor per class; 2) a daily load
for forecasting rural electric load growth. profiie (24 hourly loads) for each class; or 3) use of several
24-h curves for each class, each curve representinga part of
Land-Use Based Forecasting the classs load.
Land-use based methods involve an intermediate forecast of 1) The first method uses a single factor of load density per
the growth of land-use type and density as a f i t step in the unit of each land-use type, and may be a function of the year
forecast of electric load (Fig. 14). This type of analysis has a L,(t) = electric demand of one unit of land-use type
number of practical andoperational advantages over other c in year t .
approaches. Chief among these is the complete decoupling of
the two causes of load growth. Changes in the electrical usage The load in cell k, in year t is
on aper-customer basis is tallied by a usage modeL Change
in the number and location of customers,which drives the
geographic spread of electric load, is handled by the landuse
forecast. Other forecast methods do not explicitly make this
2) The second method uses 24 hourly load readings rather
distinctionandconsequentlysufferintheir ability to infer
valid trendsandpatternsfrom historicaldata. Another ad- than a single value
vantage of land-use methods is that they can use the utilitys ) electric demand of one unit of land-use type c,
LJh, tE
total system forecasts to provide the bulk of the data needed in hour h of the peak day, in yeart
for an electric usage model that is related to land-use classes.
As mentioned earlier, the techniques used to project total sys- and the 24 hourly loads for the day, for cell
the become
tem loadprovide extremely sophisticated projections of the c
expected average behavior by customerclass. A third advantage load in cell k at h hour of the day = (ag(t) * L C @ , r)).
is that land-use methods areparticularly well suited to multiple c=1
scenario studies. They relate to the basic causal aspects of the
(20)
growth. By contrast, the trending andmultivariate methods
discussed earlier cannot realistically vary growth modeling. This second methods advantageis that the24-h curves explicitly
The concept of land-use/customer class appeals to intuition. address the fact that classs loadspeak at different times of
Therefore, subjectiveevaluation of a land-use based forecast day.
WILLIS AND NORTHCOTE-GREEN: SPATIAL ELECTRIC LOAD FORECASTING 245

3) The most complete load models used in spatial load anal- XIC landuse needs. For example, a parcel of land near a rail-
ysis are end-use type models, in which a set of up to 200 load road trackmay be idealfor industrial purposes and yet marginal
types are used. Each is a 24-h daily curve that describes the at best-due to noise, etc.-for residential use. A parcel of land
load behavior of a particular subclass of load. For example, only a half mile away may have opposite suitabilities.
residential might be represented by subfunctions for lighting, A spatial load forecast procedure can use these three con-
cooking,waterheating,spaceheating, and air conditioning cepts of balance, gravity influence, and suitability to refiie
loads, etc. Multiple-scenariovariations to a particular subfunc- estimates of the likelihood of future land-use class develop
tion, representing alternate scenarios of future usage, can be ment on a small area basis. These concepts are very amenable
included in the spatial loads by merely varying and recalcu- to computerization, particularly in simulation approaches. In
lating the load functions,without having to recalculate the some cases spatial load growth models have addressedthese
land-use forecast. three concepts by directly applying urban models developed
for transportation and city planning [ 221, [ 231 , [ 481, most of
Land-Use Forecasting which forecast land-use growth on a census tract basis. Relat-
Computerized analysis of land usage on a spatial basis has ing census tract output of such models to the cells of a uni-
grown into an established field of socioeconomic studies that form grid or to the nonuniform equipment areas can become
has at one time or another looked at mans use of land from extremely complicated in itself [ 3 3 ] , [35]. Urbanmodeling
nearly every imaginable perspective. Three concepts of analy- concepts applied in techniques specifically designed for electric
sis, borrowedfromurban modeling,are key ingredientsin load forecasting have performedmuchbetterthan existing
manyspatialurban models: 1)theconcept of basic causal urban models applied to electric utility planning [ 1 11, 1231,
growth and balance; 2) the gravity model of influence; and [261, [381, [491, [601, [651,[661, [941.
3) the concept of small area growth suitability. In the early
1960s, Lowrydevelopeda model of a metropolitan areas Spatial Frequency Ranges of DemandlSupply
growth as driven by basic industry-that portion of the local All land-use based spatial forecastmethods,bothanalytic
employmentsectorthatmarketedoutsidethe region. The and nonanalytic, can be viewed from a spatial land demand
expansion of the basic industry would increase employment, and land-use supply modeling perspective. Many have not been
leading to expansion of the residential sector, leading to an so identified in the literature, but forconsistency and illumina-
increase in the retail (commercial) sector of the local economy tion of similarities we will use this perspective. In order for
[ 961 . This concept can be interpreted as several basic rules for growthor change to occur, theremust bea simultaneous
balancing the forecast growth of any city, town, or rural spatial match of both demand and supply. For example, if
area: a) a region must provide sufficient room and services to residential growth occurs in a cell, it is because both a demand
house and sustain the populace; b) there must be a market area for housing and a supply of suitable housing land was simulta-
to deliver food and other necessary items of sustenance to the neously present in that cell. A fascinating aspect of landuse
populace; c) land, activity, and electric load must be devoted modeling as applied to electric load forecasting is that if
to industries that provide the local economy with basic em- demand and supply are appropriately defined, they are respec-
ployment; and d) these needs must be balanced in ratio. One tively the low and high spatial frequency aspects of growth.
does not find a residential area of small size in a city with a This concept was f i t explicitly used for electric load modeling
tremendous industrial base; the industrial workers must live by one of the authors in the late1970s [ 301, but it is clearly
somewhere close to their employment. Similarly, the residen- an implicit part of most land-use approaches.
tial segment of the city necessitatesa market segment large In order to fit into this perspective, demand must be viewed
enough to feed,clothe,entertain, andotherwisesustain it. as the restrictions on where someone seeking expansion for a
Therefore,anycity,town,or region will displayreasonable particular land-use type searches to findlandmatching his
and somewhat predictable ratios of land use that grow in this needs. Examples would be the individual seeking tofind a
coupled, driven manner. new home, or a business desirous of a new location. Both have
The second concept, the gravity model, assumes the growth a specific, exact, defintion of w h r is suitable for their needs
influence (coupling) of two areas is proportional to magnitude (good neighborhood, close to schools, not close to competing
of the influencing factors and inversely proportional to a func- industry,etc.). However, in additionto these suitability
tion of distance (not necessarily inverse squared) [ 95 1, [ 961. needs,eachrestrictshissearch to a certain, perhapsfuzzily
In Lowrys original model of Pittsburgh, the expansionof resi- defined, region. Regardless of whatsuitableland might be
dential growth was assumed to be driven by a monotonically available outside of this search area, the search for a new
decreasing function of distance centered in the industrialareas; site is restricted to this region. The prospective home buyer
there was more demand forhousing close t o these employment may examine only a quadrant of a city due to a desire to be
centersthanfar away fromthem. Distance meant travel relatively close to his employment. A business may likewise
time byroad and commuter methods, not straight line distance. desire to locate a new retail outlet only in an expanding
Much of the spatial distributionof a city can be understood by quadrantofthecity. Each is willing t o accept almostany
analysis of the street, freeway, rail, and water transportation parcel of land in a wide area if it matches his requirements
systems [96], [98], [loll, [102]. Insomecases,thesimpli- better than any other parcel in that area, but will not accept a
fying assumption that travel time is proportional to straight- parcel outside the area even if it is better. In a large city, the
line distance can be used with good effect, particularly on homeowner is usually indifferent to a change in exact location
small cities lacking dominant geographical barriers. of one or two miles if he finds a particular parcel of land that
The third conceptof urban modeling is land parcel suitability. better matches what he considers suitable. But he would not
Individual parcels of land are considered suitable only for cer- accept a change in location of twenty miles for even a larger
tain land uses. Suitability as used here can be thought of as a improvement. The key concept is that demand, defined as the
numerical measure of a particular parcei of land meeting spe- locational restrictions on growth dueto where people desire to
246 PROCEEDINGS OF THE IEEE, VOL. 71, NO. 2, FEBRUARY 1983

FILTERCE REClDCNTIIL

. . _ . . . . . . ......
. . . . ..

Fig. 15. Spatial diatribution of 1980 residential electric load on a.40-


acre small areabasis for Calgary,Canada. (a) Actual load data. (b)
Data after filtering that attenuates frequencies modeled as due t o land
suitability effects. These same data are plotted in Fig. 16.

live and work in rehtion to the city or region, is indifferent to concepts at a 25-acre basis for the western half of Houston,
small changes in distance. So defined, a quantitative measure TX, based on growth from1973to1979(a period where
of "demand" varies slowly on a spatial basis; it is composed of population roughly doubled).Supply was determined as the
low and intermediate spatial frequencies. best fit tothe small areagrowthcharacteristics using the
Thehomeowner and the business have specific suitability ELUFANT spatial land-use model to be discussed later [26],
definitions of the qualities that would make a parcel of land [30],[33]. Eighty-one percent of all growthoccurredin
suitable for their purposes. Each land,usetype hascertain areas designated as a highly suitable supply (most highly rated
characteristics that arecritical, or at least important,toits 12 percent of cells). The supply was calculated with a suitabil-
suitability. Commercial retail generally wants to find a parcel itymodelthathad spatial frequencycontent of only 4-mi
of land that is close to amajor traffic corridor.Residential wavelengths and shorter. Only 43 percent of such highly rated
needs to be close to other residential (good neighborhood), cells displayed any growth. The demand was 0 in the others.
convenient to transportation, nearamenities(schools,etc.) Demand in this experiment was calculated by a transportation-
but not near industry or railroads. These characteristics tend typeurban model that modeled growth demandwithfunc-
to change rapidly in space. One particular parcel of land may tions that were inthespatialfrequency range dc-to-4-mi
be ideal for a land-use purpose, and yet aparcel only mile wavelength. When combined, demand and supply calculations
away may be very marginal for that purpose. Therefore, quan- explaineda total of 93 percent of the growth locations and
titative measures of land suitability for specific usages are of magnitudes of electric load, verifying the basic concept of high
relatively high spatial frequency. If one views the set of all frequency supply and low frequency demand.
land parcels that are suitable for a particular land-use class as Often, when small area load data are filtered,rejecting those
the supply of land for that purpose, thensupply is spatially of frequencies in the suitability-supply band, an obvious urban
high frequency. pole structure of low frequency emerges. Figs. 15 and 16
Regardless of how suitable any site is, it will develop only if show the correlation between amount of residentialload on
there is also demand for developmentthere. If demand for a per-cell basis and distance fromthe major employment
housing is restricted,for whateverreason, to only the East centersfor Calgary, Canada. When the cellular residential
region of a city, little development will occur in the West. The development is filteredwithaspatialfilter thatattenuates
excellent sites in the West nonetheless rate high as a supply of frequenciesin the "suitabilityband," an obvious correlation
land suitable for that purpose. A parcel of land may be suit- emerges from what was otherwise an apparent random scatter.
able for a particular land-use class, but growth is not guaran- The result clearly shows the simple, almostlinear tendency
teed just because it is suitable. Likewise, demand alone does for residential locationto be near theemploymentcenter.
not guarantee growth; there must also be a supply of suitable The fdter rejectedspatial distribution caused by suitability,
land. It is the combination of both that results in growth. Be- leaving only that caused by demand.
cause demand is mostly low frequency,and supply is high
spatial frequency, the authors interpret them respectively as Calculation of Spatial Demand and Supply
the low- and high-frequency models of growth. As discussed above,demand is largely transportation ori-
In 1979 one of the authors (Willis) tested these frequency ented on the basis of convenient travel to employment. The
WILLIS AND NORTHCOTE-GREEN: SPATIAL ELECTRIC LOAD FORECASTING 247

UNFILTERED DATA FILTERED DATA

MILES MILES

(a) (b)
Fig. 16. Residential electric load versus distance to downtown on a 40-
acre basis, is plotted here in scatter diagrams. (a) No particular pattern
emerges from the actual data, due t o cell-by-ceJl variations in load
density caused by factors not related to distance from the urban core.
(b) Filtering of the data results in data values which show the expected
correlation, attributable t o land-use demand factors.

better transportation urban models,such as Putmansmodel (5 - distance in cell widths to the nearest railroad),
of transportation influence [ 1011, work with zones of about
tensquare miles or larger. Hence,theyare
models of growth.Little
cepts,procedures,or
low-frequency
if anymodification of thesecon-
calculation methods is needed for ap-
{
D~ = if it is less than 5 cell widths away
0, otherwise.
The distance weighted sum of all of one class or item in the
plication in spatial load models. Analysis can be performed on area is called a surround factor.
a large area basis and the low-frequencyresults interpolated
for a high resolution [ 4 3 ] . S:,j = sum of class c development near cell i, j , weighted
Inour discussion of trendingmethods given earlier, we by a function of distance W,(d).
showed that for many situations, the simple methods of the
urban pole-typemodelall the low-frequency components of
spatialgrowth.Theseare the samefrequencies thatarethe x=1y=1
demand band. Therefore, such urban polemodelsapplied
in a land-use method often sufficiently fulfill the needs of a Here, W,(d) ismonotonically decreasing withdistanceand
demandmodel. Because urbanpolefunctionsarelow fre- W ( d ) = 0 for d >8. Both W,(d) and 8 will vary depending on
quency, in the authors viewpoint they areland-use-demand the factor being calculated.
models even if implemented as part of the trending approaches Combinationsandfunctions of several proximityand sur-
discussed earlier. round factors are used in many land-use based models to ex-
By contrastwiththe low-frequencymodels, the high-fre- plain suitability on a high-resolution basis. For instance, the
quency models used in spatial load modeling have not applied suitability for residentialdevelopment may be a function of
the quantitative methods used in urban modeling, where there proximity t o schools,proximity to commercial(shopping),
is a perceived difficulty in calculating a measure at high spatial amount of surroundingresidentialdevelopment-all positive
resolution [ 1011. Define suitability for land-use type c, in cell factors. Negative factors may include proximity t o railroads,
k as amount ofsurroundingindustry,andproximity t o airports.
Calculation of the suitabilitiesbased on these factors is straight-
p i = suitability of type c in cell k, a measure of how cell k forward. It is thedetermination of theproximityandsur-
matches the needs of land-use type c. round factors that is difficult. Direct calculation proved very
The exact manner of calculation of the p k varies among pres- lengthy in some of the first electric load models in which it
ent forecasting methods. Usually, & is a function of both the was applied [ 3 0 ] , [ 3 6 ] .
cells datavector V k ( t ) and the data vectors of other small Factors such as Dn and Wn above can be calculated rapidly
areas next to or near to it. Many modern spatial load models in the spatialfrequencydomain.Notethatsince W ( d ) is of
determine suitabilities using very extensive calculations on the finite duration, monotonically decreasing t o 0 beyondsome
basisof twotypes of factors-proximityand s ~ r r o ~ n d distance b (consistent with the gravity concept) ( 16) may be
factors. A proximityfactor is an inverse distance to the rewritten as
nearestitemofsome set ofamenities. Anexamplefrom
Houston Lighting and Powers ELUFANT model [ 3 6 ] is used
in industrial suitability analysis.
Dk = 5 if there is a railroad in cell k a convolution of the u, with a radially symmetric FIR oper-
248 PROCEEDINGS
IEEE,OF THE VOL. 7 1 , NO. 2, FEBRUARY 1983

ator. Calculation can be done in three steps: 1) Transform Most important, multiple-scenariovariations cannotbe per-
boththe as image and W ( d ) intotheir Discrete Fourier formed due to lack of access to the urban model itself.
Transforms (DFTs) by use of a Fast Fourier Transform (FFT) Fischer developed a unique land-use method for Dallas Power
method; 2)multiply thempoint by point; and 3) inverse and Light that used preferences based on zoning, road, and
transform the products using the FFT. The result is the Sg. highway networks [SO]. It is the only 1-3-1 method the authors
This procedure is fast, making such calculations practical [ 361, have encountered. The model has an extensive and very inter-
[47], [ 711. The authors have recently implemented a similar esting method of tying historical load data on an equipment
methodthat calculates Dk factors within8 percent using area basis, to the small area land use. It was recently replaced
theFFT, alogarithmic operator and a trickinterpreta- with a 2-3-3 method developed by the authors [ 1051.
tion [ 1061, The EPRIsponsored Multivariate model used EMPIRIC, a
census-tract based urban analysis model whose census tract
Land-Use Based Forecast Models projections of growth, allocatedmanually amongthe small
areas in each tract, were the control variables referred to in
As discussed above, all land-use based spatial load forecast the earlier discussion of Multivariate. EMPIRIC performed the
models consist of three basic components: 1) demand model; spatial land-use demand analysis, Multivariateaccomplished
2) supply or suitability model; and 3) land-use to load model, both high-resolution suitability analysis and the load model in
varying only in the level of sophistication of each. one step. This method was basically a 3-2-1 model and gave
Demand models have been implemented in three basic ways: good results in tests onPhoenix, AZ [ 491.
1) no spatial demand model other than a control total by Brooks developed a now widely used 2-2-1 model using two
class [171, [SO]; urban pole functions which are coupled by different ratios to
2) urban poles spatial model [ 131 , [ 341, [ 651 ; each of 10 land-use or customer classes. Suitabilities for each
3) urbanltransportation model [ 301, [66]. land-use class are calculated from two input data factors. The
combination of these two factors with the demand function
The third type is normally done on a large area (zonal) basis. yields a growth allocation value for each cell for each land-use
The second is composed of simple functions of inherently low class. The resulting values are the relative probability of de-
spatial frequency but normallycalculated at high resolution velopment in a Monte Carlo allocation of the total system
for the sakeof simplicity. growth byland-use class among all K cells. Atype-1 load
Supply models are high resolution and also fall into three model is used. This method has given excellent results in fore-
basic categories: casting for small metropolitan areas [ 341 , [ 1071.
1) manual input suitability values based on user intuition Willis developed a nonanalytic land-use method during the
[Ill, [171; early 1970s that was subsequently evolved into a 3-3-1 model
2) combination of zoning and manual input suitability [ 341 ; called ELUFANT [301, [33], [361,[461, [471,used at Hous-
3) suitability measures calculated from detailed spatial fac- ton Lighting and Power from 1974 to 198 1. It operated on a
tors, usually based onproximity and surroundfactors grid of over 250 (&
000 cells of 25.6 acres square mile). A
detailed suitability model operatingatthisresolution calcu-
[361, i 1051.
lated suitabilities based on 20 proximity and surround factors.
The load model can be done in any of three forms as dis- A Lowry urban model, operating on 16-miz square zones, was
cussed earlier: used to allocate the total system growth (an input) to these
large zones. Once calculated, the suitability factors were
1) a single load value for each class [ 171, [ 341 ;
2) a 24-h curve for each class [45], [65]; matched with demand factors using a complex clustering-pat-
tern recognition analysis borrowed from the EPRI multivariate
3) a 24-h curve for each of several end-use subclasses of
each class [65], [66]. model [30], [36]. A type-1 load model was used, with load
values estimated from historical substation load data using a
The utilitys total system load forecast provides much of the novel method developed by Sadler [35].
information needed for any of these three load models. ELUFANT, one of the f i t attempts at large-scale high-
Because all land-use based methods employ some form of resolution land-use simulation, explicitly modeled demand and
analysis in each of the threecategories, it is convenient to refer supply as low- and high-frequency processes. The original
to any land-use based forecastmethodwiththreenumbers model (1975)had a gap between the demand models 16-miz
corresponding respectively to the basic type of demand, resolution and those frequencies easily and efficiently handled
supply, and load models. Thus a 2-1-2 model is a technique bythe 25-acre high-resolution model. Gregg subsequently
that uses urban pole demand functions, manually input small developed an improvedlow-frequencymodel for ELUFANT
area suitabilities, and 24-h class-by-class load curves. We will to correct this problem [431.
use this categorization in the discussion of models that have A 2-3-3 land-use based method has recently been completed
been used in the utility industry. by the authors organization, under sponsorship of the Canadian
A number of electric utilities have obtained the output of ElectricAssociation [65]. This model uses a demand model
low resolution, census-tract based urban models used for with multiple urban poles d e f i e d as the center of growth of
municipalplanning and combinedthiswithmanually deter- each class. A suitability modelcombines manually input zoning
mined suitabilities in a semi-analytic 3-1-1 method [ 8 ] , [ 151, with ten proximity and surround factors calculated in the fre-
[ 201, [48]. This approach gives better results thanpurely quency domain. For the demand model, the center of each
manual,coloring bookmethods because theurbanmodel urban pole, its diameter and its slope are found by a statisti-
balances growth as discussed earlier. Problemsincredibility cally guided iterative procedure utilizing filtering of the type
and accuracy have been reported [ 161, [22], [46], [65], due demonstratedin Fig. 15. Calibration ofthe high-frequency
to the dependence on an urban model external to the utility. model is basically heuristic, but straightforward if carried out
WILLIS AND NORTHCOTE-GREEN:
FORECASTING
LOAD
ELECTRIC
SPATIAL 249

with theaid of: 1) data fitering that attenuates low frequencies; (cell size required) is set by the requirement of spatial fre-
and2) displaymaps of calculated surroundandproximity quenciesneeded by the planning,as described earlier in the
factors. A type-3 load model uses 96-point (15-min) daily load discussion of error. The type of demand model is really dic-
curves, similar to those used in load management studies [ 771 - tated by the complexity of the metropolitan area. Large cities
[ 821. This model, called CEALUS, has proven extremely easy such as Los Angeles or Houston are so extensive that a simple
to use and accurate in recent tests on Calgary, Canada. urbanpolemodel is generally not sufficient.Amodelwith
Theauthorsrecently developed a very flexible computer two, three, or even four urban poles is fairly easy to calibrate
program for land-use based spatial forecasting. This program, and apply. An 11 urban pole demand model of Houston gave
called SLF-2, iscontrolled by a user-specified macro program, results equivalent to the transportation analysis urban model
of instructions for the basic forecast steps [ 1051 . This macro portion of ELUFANT, but was no easier to use because deter-
program capability allows it to be quickly tailored to a partic- mining the factors for eleven urban poles was more difficult
ular problem/data/situation. SLF-2 is now in use on a dozen than calibrating a transportation model. The choice of demand
utility systems, implementing 1-1-3, 2-1-2, 2-2-2, 2-3-3, and model type is really based on which is easiest to use in the
3-3-3 forecast models. Procedures used in SLF are similar application at hand.
to those of other land-use simulation methods with three ex- Land-use methods are ideally adapted to multiple-scenario
ceptions. First, calculation of suitability, not just the surround forecasting. Earlier, we discussed how a type-3 loadmodel
factors, is done in the frequencydomain. Suitability is modeled could be used for scenarios involving the application of load
as linear functions of the surround and proximity factors, and management and other changes in usage. In addition, land-use
the suitability functions spectrum can be directly calculated simulation models have data and parameters in both the de-
in the frequencydomain. Second, calculation of low-frequency mand and supply models adaptable to representscenarios
effects is done at low resolution and interpolated as needed. of significant shifts in social, economic,and transportation
Third,the program will automatically analyze multiple sce- parameters [231,[431,[491,[65).Inthis respecttype-2
narios in the manner to be discussed later. Several high-speed demand models are somewhat harder to use than type 3, but
fitering procedures, which reduce two-dimensional symmetri- have been made to perform very well [33], [45], [ 1071.
cal filters to one-dimensional approximations, are used t o It was shown earlier that errors-differences between forecast
increase speed [ 711, [ 1041. and actual loads-had an impact, ordifference of design that
Most land-use methods use only a single year of data history was assessible by a filtering analysis. The same basic assessment
(the most recent)as the base for the entireforecast. Contrasted of impact can be done to compare two or more alternate sce-
with this are multiple-year methods that base the forecast on nario forecasts. The cellular differences in loads of two fore-
several years of land-use data. These twotypes of method casts can be filtered in the same manner used to evaluate error
differ in how they are calibrated against past data. In single- impact. If a significant portion of their differences is passed,
year methods, thedemand and supply models are calibrated or it means the scenarios will have an impact on the planning.
fitted to bestexplain all land-use distribution in the only The authors spatial load forecasting program, SLF, uses this
year of data used. In multiyear methods the models functions concept in two ways. First, the differences among scenarios
are adjustedto explain the changes between two or more years. are filtered and the scenarios are ranked by expected impact
The single-year methods aresurprisingly effective and currently or design difference for each level of the system-a guide for
outperform multiple-year methods by quite large margins [ 591, the engineer in designing systems for each. Second, significant
[ 651 . The critical factor limiting multiyear methods is data input and model parameters are pertubated in fl0-percent in-
error in year-to-year land-use data [45] , [66]. As an example, crements and the resulting changes in loads are filtered. Only
data from 1977 and 1981 for a utility with a customer growth those parameters that resultin significant filtered impact
rate of 4 percentannually, wouldrepresenta17-percent are varied, singly andin combination, to establish some of
increase in total land-use density (measured as number of cus- the models scenarios [66]. Seldom do changes to the high-
tomers). If each land-use cellular data base had 5-percent ran- frequency(suitability)model of growth result in filtered
dom data error-about the best level economically attainable- impact, since these changes are high frequency, and the P,(d )-
differences in the twoyears data due only to error are roughly type filters are low pass. Therefore, most scenarios of impor-
10 percent [ 701. Total difference-error plus actual changes tance represent variations in the land-use demand and electric
caused by growth-is in the orderof 27 percent, of which over load (usage) model.
f is error. Calibrating a multiyear model to this difference in- Spatial frequency analysis of spatial load distribution itself is
troducessubstantialerrorintothe forecast.At the present interestingand auseful step indeterminingapplicability of
time, the quality of available data usually gives better results if forecast procedures. Fig. 17 shows the spectra, in one dimen-
single year models are fittedto all land-use location in only the sion, of four cities in North America. The high and intermedi-
most recent year of data [59], [ 701. The desire to improve ate frequency components vary among these examples largely
this situation has led the authors to research LANDSAT data due to the differing terrain and geographic restrictions in the
and automatic interpretation as a source of historical land-use various cities. Our last figure, Fig. 18, shows spatial frequency
data [371, 1551. It is hoped that interpretation errors from ranges of a number of concepts we have talked about in this
automatic procedures would be highly correlated from year to and the previous section, the averages of data and parameters
year, minimizing the problem discussed above. So far, results on four cities in North America. It is presented in no more
have been disappointing [ 591. quantitative form because we are not certain the resultsare
quantitatively general. Abroad range of high spatial fre-
Applicability of Land-Use Based Methods quencies are of vacant land types. The availability, or lack of
The applicability of the demand and supply models of land- availability, of land forgrowth is critical to high-resolution
use growth is interesting when viewed from a spatial frequency forecasting.
standpoint. The high-frequency suitability models resolution Analysis of the frequency content of the electric load, and
250 PROCEEDINGS OF THE IEEE, VOL. 71, NO. 2, FEBRUARY 1983

5
~RE&r- C~CLES/MILE .5

Fig. 17. Spectra of spatial load distribution o f fourcitiesin North


America. Two-dimensional spectra have been reduced to one dimen-
sion using a method based on an inverse to the McClellan Transform
[ 1061. Each spectrum is normalized to dc component equal t o 1 .O.
Low-frequency differences are due mwtly to the size and ove-rall lay-
out of the cities. Mid-and high-frequency differences arelargely a
function of terrainand geographic restrictions. Houston, built in a
region that is geographically rather featureless, has lower medium-
and high-frequency content than Phoenix, a city built around s e v a a l
small mountains and a flood zone, or Syracuse, a city that is built
around a lake and several other geographic restrictions.

I I I I I
DC 05 10 15 20 25
SPATIAL FREOUENCY
CYCLESIMILE

Ti. 18. Qualitativerange of spatial frequencies of needs, effects, and


data involved in spatial load forecasting.

the various data sources and available forecast models, can be of analysis consistent with the planninggoals, planning period,
used to tailor and calibrate a land-use approach to best fit a andplanningmethod.Theforecastmethodmust also be
particular situation. As mentioned earlier, evaluation of spatial compatiblewithboththe available data base andcomputer
resolution needs using the P,(d) and other filters provides in- and human resources.
sight on the required small area size, andimplicitly,onthe Applicability of anyforecastmethod, aswellas accuracy
limit of spatial frequencies that must be accuratelycovered by and quality, varies greatly depending on data, planning goals,
the forecast model. The spatial frequency content of the elec- and resources. For this reason quantitative comparison of the
tric load itself gives an indication of the degree of importance performance ofall theforecastingmethods covered inthis
of the low- and high-frequency models of growth. If the high- paper is difficult. In Fig. 9 we attempted to plot the various
frequency content of the spatial load pattern is low, the error algorithms by both forecast quality and overall complexity of
introduced into a forecast by ignoring high frequencies may be method, data, and resources. Table 111 gives the relative error
tolerable. Alternately, it may be possible t o establish that the levels we use in our own workwhenselectingfromamong
high frequencies of interest can be tied to one particular data forecast methods for any particular study. It was developed
source, such as vacant land (Fig. 13),or to a very simplified based on our own work over the past 12 years. Within each
model. Considerable attention to the high-frequency modeling category of performance measure, wehave defined trending
is justified because cost in termsof data, computer and human with regression based curve fitting as 1.O.
resources, and complexity increases rapidly as spatial resolu- We admit to a very strong preference for the land-use based
tion is increased. approach, largely because of its intuitive appeal, its multiple-
Applicability of a forecasting methodto aparticular problem scenario capabilities,and itscompatibilitywiththegrowth
is dependent on a number of factors, only some of which are characteristics of electric load. The higher costs, mostly due to
related to the forecasting. It is important not to lose sight of data needs, are easily justified by the performance advantage
the ultimate goal, which is not the forecast but the planning over other techniques. Although present industry applications
of a well-designed, economical power delivery system. There still make heavy use of load extrapolation, the trend is clearly
fore, the load forecast should be selected to provide the type toward land-use based methods [ 901-[ 941. The reasons we
WILLIS AND NORTHCOTE-GREEN: SPATIAL ELECTRIC LOAD FORECASTING 251

TABLE 111 land demandmodel, land supply model, and electricload


RELATIVE
PERFORMANCE
OF GENERIC OF FORECASTING
CLASSES METHOD model.
Planning Spatial load forecasting is not an end in itself but only the
Impact
Cost
as of Cost
of
MS Measured Required
Required first step in the planning of a power system. Most of the con-
Method T m -
Error by Eq. 8 Resources
Data
straints, goals, and priorities that have shaped theformof
TRENDING - theseforecasting methods and theirapplication are due to
Regression)(ref.
1.0 29) 1.0 1.0 1.0
needs of other planning tools. Spatial load forecast methods
VAI-TRENDING (581 .6 .55 1.0 1.3 andoptimization-based techniques are dependentonone
another, neither can produce large improvements without the
LOAD POLEURBAN FIT (13) .5 .40 1.5 2.0
other. The recent recognition that these two together can cut
NON-A)(ALYTIC LAND USE
delivery system costs by 10 percent has spurred development
(17) .7210.0 .55 10.0 of both. We believe that the historical development rate and
8.0
capabilities of spatial load forecast methods looks much like
10.0
ANALYTIC LAND
.30 USE
.302-2-1
(34) the S curves of small area growth itself, with the present
state of the art midway up the high slope part of the curve.
ANALYTIC LAND USE 3/2-3-3 20 .15 12.0 10.0
(36)
We expect intense development and increasing application of
such methods during the remainder of the 1980s.
USE LAND
ANALYTIC 3/2-3-3 .15 .075 15.0 12.0
with multiple filtered
scenarios based on
parmter pertubations
ACKNOWLEDGMENT
(66)
The authors wish to thank A. B. Cummings, J. Gregg, S. D.
AS measured on feedersusingthebestapplication o f theforecasting
nethod. Havemann, T. W. Parks, C. R. Sadler, D. L. Wall, and C. L.
Brooks for their many helpful suggestions and comments.
have given above provide most of the motivation for this shift.
Another and equally important reason is that a land-use based REFERENCES
method can be heavily tied to a detailed,customer class-oriented [ 1 ] E. L. Kanouse and J. W. Reinhard, The function of land use
forecast of a utilitys total systemload.Therefore, use of a surveys in power system planning, in Proc. American Power
Conf., 1955.
land-use based method improves thecoordinationbetween [ 2 ] F.C. Van Wormer, Some aspects ofdistribution load area
thepower delivery planningforecast andthe overall total geometry,ZEEE Trans PowerApp. Syst., p. 125, Apr. 1955.
system forecast. [ 31 W. J. Denton and D. N. Reps, Distribution substation and
primary feeder planning, ZEEE Trans PowerApp. Syst., p. 484,
The advantages of land-use based methods have focused June 1955.
most present research activity on them,almost to the complete [ 41 R. H. Sarikas and H. B. Thacker, Distribution system load char-
acteristics and their use in planning and design, presented at
exclusion of all other approaches.These techniques will un- the AIEE Winter Meeting (New York, Jan. 1956), AIEE paper
doubtedly become the powerindustrys standardapproach 57-168.
during the remainderof this decade. [ 5 1 S. J. Vest, Estimating kW demand for future loads on rural
distribution systems,AIEE Trans., p. 652, Aug. 1957.
[ 6 ] D.N. Reps, Monte-Carlo simulation of residential transformer
loads, in Proc AIEE Winter Meeting (New York, J a n . 1962),
V. CONCLUSION pp. 62-289.
[ 71 H. Jung and K. Smalling, Determination of area load on LILCO
Spatial load forecastingis currently undergoing a rapidchange system using kWH, presented at the Penn. Elec. h o c .Meeting
as more traditional methods are replaced by newer techniques (Long lsland Lighting Co., May 7, 1962).
[ 81 L. Cavic, People and land-use determinefuture loads, presented
that properly mesh with modem optimization-based planning at the Penn. Elec. Assoc.
Meeting (West Penn Power Co., May 7,
and design techniques. Theapparent characteristics of load 1962).
[ 9 1 W. 0.Zimmermann, Short range distribution load forecasting,
growth depend on time period and spatialresolution of the presented at the Penn. EIec. Assoc. Meeting (Long Island Light-
data and planning. Therefore, application of particularfore- ing Co., May 7, 1962).
cast methods is best evaluated by relating their qualities to the [ 101 Forecast loads four ways,E&c. World, p. 41, June 1962.
[ 11] A. Lazzari, Land we data improves load forecast, Elec. World,
characteristic dynamics of load growth at the resolution and June 18, 1962.
time period being studied. Signal theory is invaluable in spatial [ 12) W. R.New, Load forecasting on the TVA system, in Proc.
load analysis as it providesacomprehensiveframework for AZEE Winter Meeting (January 1963), pp. 63-222.
[ 131 A. Lazzari et al., Computer speeds accurate load forecast at
study of planningsensitivity, loadgrowthcharacter, model A B , Elec. Light Power, pp. 31-40, Feb. 1965.
structure and performance, anderror. Accuracy in a load fore- [ 141 W. Klein, Load forecasts by grid squares, Public Service Com-
cast is best judged byevaluating errors ina spatialfiter manner pany of New Jersey, Public Service Electric and Gas Company
Rep. 36.67.14, July 19, 1967.
to estimate impact on planning. The planning of the various [ 15 1 G. W. Austin, Load and growth maps aid distributionplanning,
levels of a power system differ in thesensitivity to spatial error Transmidon and Dirtribution,Sept. 1967, p. 44.
distribution. [ 161 H. D. Turner, Development of electric load forecasting from
billing, land usage, and other basic data, presented at the Elec-
Spatial load forecasting methods operate with one of three tric Utility RogramTeam Common (Houston, TX, Dec. 10,1968,
approaches-trending of historicalload growth, multivariate Memphis Power Co.).
[ 171 W.G. Scott, Computer model forecasts future loads, Elec.
extrapolation, or developing a simulation of landuse growth World, p. 114, Sept. 1972.
as an intermediate forecast. The land-use based methods are [ 181 J. F. Rad, Substation and feeder load forecasting, T r a m i s -
suited to long-range study needs, both because they can per- don and Distribution,p. 28, June 1974.
[ 191 W. G. Scott, Computer model offers more improved load fore-
form valid multi-scenario studiesand also because they are casting,EnergyZntmfional. p. 18, Sept. 1974.
well suited to high-resolution, long-range growth dynamics. [ 201 L. C. Barry and W. G. Scott, Modeling techniques aid planning,
Elec. World, Mar. 1973, p. 101.
The various methods of land-use based load forecasting differ [ 2 11 A. P. Barsalou, Load forecastingof terminal station, substations,
only in thedegree of sophistication of three basic components: and distribution, presented at the Canadian Power Engineering
252 PROCEEDINGS OF THE IEEE, VOL. 71, NO. 2, FEBRUARY 1983

sodety Meeting (Manitoba Hydro Power Company, Winnipeg, [ 49 1 Research into load forecasting and distribution planning, Elec-
Ont., Calla&, Oct. 19, 1975). tric Power Research Institute, Falo Alto, CA, EPRI Rep. EL-1 198,
I221 -, Long range suburban di8triiution planningtechniques, 1979.
presented atthe IEEE Canadian Power Enginedng Society [ S O ] R. L. F i h e r , LANDUSE Users Guide, Doctor of Electrical
Meeting (Manitoba Hydro Power Co., Winnipeg, Ont., Canada, Engineering Internship Rep. Texas A&M Univ., May 1980.
Feb. 8, 1977). [ 5 1 1 H.L. Willis, Microcomputer based simulation of small area
[23] F. M. F i e et al., Application of the EMPIRIC urban model to urban electric demand growth, in Roc 12th Annual Pittsburgh
small area load forecasting, in Roc 8th Modeling and Simub- Con$ on Modeling and Simulation (Pittsburgh, PA, Apr. 1981,
tion Conf.(Pittsburgh, Pa, Apr. 1977). ISA).
[ 241 V. F. Wilreker, Landsat applications to small area electric load 521 J. C. Thompson and I. Gonen, A developmental system approach
forecasting, in Proc 8th AnnualModelingandSimulation Con$ to modeling electric energy demand growth in radial network,
(Pittsburgh, PA, Apr. 24, 1977, EA). in Proc. 12th Modeling and Simulation Conf. (Pittsburgh, PA,
[ 251 E. E. Memge e t al., Electrical loadscan be forecasted for distribu- Apr. 1981, SA).
tion planning, in R o c . American Power Conf. (Univ. Illinois, 531 R. F.Wolff,ComputeninT&D,Ekc World,p.67,Apr. 1981.
Chicago, IL, Apr. 1977). 541 H.L. W ill
is e t al., Computerized small area load forecasting,
[26] H.L. Willis, Urban model application to electric demand fore- in Proc 7thPowerSystems Computcrtion Conf. (Lausanne,
e g , . in Proc. 8th Annual Pittsburgh Con$ Modelingand Switzerland, July 1981).
Srmulamn (Pittsburg, PA, Apr. 24, 1977, ISA). 5 5 1 H.L. Willis, LANDSAT applications to spatial electric demand
[27] V. F. Wilreker e t al., Spatialiy regressive small area electric load forecasting, in Proc. Western Remote Sensing Con$ (Monterey,
forecasting, in Proc. IEEE Joint Automatic ControlConK I977 CA, May 1981, NASA).
(San Francisco, CA). [.~561 M. A. Abu-El Magd and N. K. S i b a , Two new algorithms for
[28] R. W. Long et al., A geometricmultidimensionalclustering on-line modeling and forecasting of the load demand of a multi-
algorithm, in Proc. ZEEE Joint Automatic Control Conf.,(San node power system, ZEEE Trans Power App. Syst., p. 3246,
Francisco, CA, July 1977). July 1981.
[ 291 E. E. Menge, Improvements in time series analysis increase ac- [ 57) H. L. Willis and T. D. Vismor, Discrete transition matrix analysis
curacy of small area demand projection^," in Proc. ZEEE 1977 ofspatialelectric demandtime series, presented at the 5th
Joint Automatic Control Conf. (San Francisco, CA, July 1977). Annual Joint Time Series Conf. (Houston, TX, Aug. 1981).
[ 301 H.L. Willis, Small area electric load forecasting by dual level [ 5 8 ] H.L. Willis and J.E.D. Northcote-Green, A hierarchical re-
spatialfrequencymodeling, in Roc ZEEE Joint Automatic cursive method for substantially improving trending of small
Control Conf. (San Francisco, CA, July 1977). area load forecasts, ZEEE T r a m Power App. Syst., p. 1776,
[ 31 1 B. M. Sander, Forecasting residential energy demand: A key to June 1982.
distribution planning, presented at theIEEE PES Summer Meet- [ 591 T.D. Vismor et al., Landsat data applications to small area
ing (Mexico City, July 1977), IEEE paper A77642-2. electric demand forecasting,presented at the 13th EEI/APA
[ 321 W. G. Scott, Engineering management and control of distribu- Workshop on Remote Sensing Applications to Electric Utitities
tion systems via minicomputer, presented at the IEEE Control (Palo Alto, CA, Sept. 24,1981).
of Power System, Conf. (Oklahoma City, OK, Mar. 1978). [ 6 0 ] C.L. Brooks, Overview of load forecasting, in Proc. 13th
[ 331 H. L. Willis, J. Gregg, Y. Chambers, Spatial load forecasting for Modeling and Simulation Conf. (Pittsburgh, PA, Apr. 22,1982).
system planning, presented atthe American Power Conf. [ 61 1 A. E. Schauer et al., A new load forecasting method for distribu-
(Chicago, IL, Apr. 1978). tion planning. presented atthe13th Annual Modeling and
[ 341 C.L. Brooks and J. E. D. Northcote-Green, A stochastic-prefer- Simulation Conf. (Pittsburgh, PA, Apr. 22, 1982).
ence technique for allocation of customer growth using small [ 6 2 ] J. C. Thompson and T. Gonen, Simulation of load growth de-
area modeling, in Proc. American Power Conf. (Univ. Illinois, velopmental system models for comparison with field data on
Chicago, IL, Apr. 1978). radial networks, in Proc. 13th Modeling and Simulation Conf.
[ 35) C. R. Sadler and H. L. Willis, Boundary location errors in spatial (Pittsburgh, PA, Apr. 22, 1982, ISA).
allocation models, in Roc. 9th AnnuaZPittsburghModelingand 631 H. L. Willis, Load forecasting for distribution planning-Error
Simulation Cons (Univ. Pittsburgh,Pittsburgh, PA, Apr. 22, and impact on design, to be presented at the 1982 IEEE/PES
1978, SA). Summer Meeting (San Francisco, CA, July 18-23, 1982, paper
[ 36 1 H. L. Willis, Y. Chambers, and J. Gregg. The design of a spatial BZSM470-3.
energy model for electric utility system design, in Proc. 9th 641 H.L. Willis et al., Forecasting distribution system loads using
AnnualModeling and SimulationConf. (Pittsburgh, PA,Apr. curve shape clustering, to be presented at the 1982 lEEE/PES
1978, ISA). Summer Meeting (San Francisco, CA, July 18-23,1982), paper
[ 371 H.L. Willis and V.F. Wilreker, An improved method for spa- B2SM385-3.
tialiy related load forecasts for power system planning, presented [ 651 Final report on CEA Project 079DlS6-Urban distribution load
at the 1st Annual Lawrence Symposium (Berkeley, CA, Sept. forecasting-Canadian Electric Association, 1982.
1978). [66] R. W. Powell e t al., Advances in distributionplanning tech-
[ 381 A. B. Cumminkp and J.E.D. Northcote-Green, Method for fore- niques, presented at the 1983 CEPSI Conf. (Bangkok, Thailand,
casting loads in small areas, presented at the Pacific Coast Elec- Nov. 30, 1982).
tricalAssoc. Meeting (Salt River Project, Tempe. AZ, Mar. 16, [67] E.P. Cody, Loadforecasting method cuts time,cost, EIec.
World, p. 87, Nov., 1982.
1979).
[39] J.E.D. Northcote-Green, Spatial effects on feeder optimization (681 F. Y. Murad and J. A. Perry, Computerized distribution load
in distribution planning, in Proc 10th Modeling andSimulation projection in planning, presented a t the Southeastern Electric
Conf. (Pittsburgh, PA, Apr. 1979, ISA). Exchange Meeting (Hot Spring, AR, Apr. 13, 1970).
[40] C.R. Sadler e t al., Multi-scenario spatial forecast modeling, [ 691 H. L. Willis and H. N. Tram A cluster based V.A.I. method for
in Proc. 10th Annual Pittsburgh.Con$ on Modeling and Simuh- distribution load forecasting, to be presented at the 1983 IEEE
tion (Pittsburgh, PA, Apr. 1979, ISA). PES Winter Power Meeting (New York, NY, Jan. 31,1983).
[41] H.L. Willis, Spatial load fdtering for power distribution plan- 701 H. N. Tram et al., Load forecasting data and data base develop-
n+g. in Proc. 10th Annual Pittsburgh Con$ on Modeling and ment for distribution planning, submitted to the 1983 Power
Stmulation (Pittsburgh, PA, Apr. 1979, ISA). Industry Computer Applications Conf. (Houston, TX, May 19,
(421 J. R. Meinke, Sensitivity analysis of small area load forecasting 1983).
models, in Proc. 10th Annual Pinsburgh Modelingand Simula- 71 ] H. L. Willis and T. W. Parks, Fast algorithms for small area load
tion Conf. (Pittsburgh, PA, Apr. 1979, ISA). forecasting, submitted to the 1983 Power Industry Computer
[43] J. Gregg et al., Improvements in spatial load forecasting, in Applications Conf. (Houston, TX, May 19,1983).
Proc. 10th Annual Pittsburgh Conf. onModeling and Simuhtion 721 A Guide to Distribution Load Forecasting, Canadian Electric
(Pittsburgh, PA, Apr. 1979, EA). Association, Sept. 1977.
I441 C.L. Brooks, The sensitivity of small area load forecasting to 731 S. E. Collier, A comparison of load research project results with
customer class variant input, in Proc. 10th Annual Modeling the REA demand tables, in Proc. 1982 Rural Electric Power
and Simuhtion Conf.(Pittsburgh, PA, Apr. 1979, ISA). Conf. (Oklahoma City,OK,Am. 25,1981). IEEE CH 1733-5-D4.
[45] Summary of six years experience with ELUFANT forecast and [ 74) J. &r& e t al., A Markov ~ & ~ e s appged
s to forecasting, pre-
data base program, Houston Lighting & Power Co., Houston, sented at the IEEE PES Summer Meeting (Vancouver, B.C.,
TX, Tech. Rep. EP-79-3, 1979. C m d a . July, 1973), C73 475-1.
[ 46 1 H. L. Willis and J. Gregg,Computerized spatial load fomxsting.
TMn and Distribution, p. 48, May 1979. Non-Spatial Load Forecasting
[47] H.L. Willis and J. V. Aanstoos, Some unique signal proceging
applications in power system planning. ZEEE Tmm Acoust., [ 751 Load forecasting bibiliography: Part 1,ZEEE Tram PowerApp.
Speech, Signal Process., vd.ASSP-27, p. 685, Dec. 1979. Syst., p. 53, Jan. 1980.
[48] M. Munasinghe, The Economics of Power System ReliobUityand (761 Loadforecasting bibiliogrpphy: part 11, ZEEE Trans Power
Planning. Washington, D.C.: The World Bank, 1979. App. Syst., p. 3217, July 1981.
WILLIS AND NORTHCOTE-GREEN: SPATIAL ELECTRIC LOAD FORECASTING 253

[ 771 J. H. Broehl, An end-use approach to demandforecasting, [ 9 3 ] Industry distribution planning survey, Houston Lighting and
IEEE Trans Power App. Syst.,p. 2714. June 1981. Power Co., Houston. TX, tech. rep. EP-80-8, 1980.
[ 781 Bibliography on Load Management-IEEE Committee Report, [ 941 R. F. Wolff, The new electronic frontier-Distribution design,
IEEE Trans. PowerApp. Syst.,p. 2597, May 1981. Elcc. World, p. 65, May 1982.
[ 791 R. B. Adler et al., A distribution system/endvser cost model, (951 System design using dispersed generation andload manage-
IEEE Trans. PowerApp. Syst., p. 3590, July 1981. ment, Final Rep. Roj. DEDS1941, U.S. Dep. Energy (in
[ 801 R. F. Hamilton, The summation of load curves, AIEE Trans., preparation).
vol. 61, p. 369, 1942.
[ 81 ] Modeling and analysis of electricity demand by time of day,
EPRI Roj. EA-1304, Palo Alto, CA, Dec. 1979.
[ 821 Energy Demand and Consumption by Time of Use, Electric Other References
Power Research Institute, Rep. Project EA-1294, Palo Alto,
CA, Dec. 1979. I. Walter, Location and Space Economy. Cambridge, MA: The
MIT Res, 1956.
CalibratingandTestinga Gravity Model f o r A n y Size Urban
Distribution Planning Area, D.O.C., Bureau of Roads, U.S. Gov. Printing Office, 1963.
[ 831 C. J. Baldwin et al., Techniques for the simulation of subtrans- I. S . Lowry, A Model of Mezropolir; Rand Corp., Santa Monica,
mission and distributionsystem expansion,in h o c . IEEEPower CA, Memo. RM-403541964.
Industry Computer Applications Conf.(May 1969), p. 575. Final report: EMPIRIC land-use forecasting model, Final Rep.
[ 841 E. Masud, An interactive procedure for sizing and timing dis- for Eastern Massachusetts Regional Planning Project, 1970.
tribution substations using optimization techniques, presented C.R. Sadler, Practical aspects of urban model data error, in
at the IEEE PES Winter Meeting (New York, NY, 1974), paper Proc. 9th Modeling and Simuiation Conf. (Pittsburgh, PA, Apr.
T74142.142-6. 1977, HA).
[ 8 5 ] D. M. Crawford and S. B. Holt, A mathematical optimization S . H. Putman, Integrated policy analysis of metropolitan trans-
technique for locating and sizing distribution substations, IEEE portationand location, U.S. Dept. ofTransportation, DOT
Trans. PowerApp. Syst., vol. 94, no. 2, p. 230, Mar. 1975. P-30-80-32, Aug. 1980.
[ 861 L. S. Neal, An interactive computer approach to distribution S . H. Putman, UrbanResidentialLocationModels Boston:
circuitmodeling,presented at the Power Distribution Conf. Mirtinus Nijhoff, 1979.
(Austin, TX, Oct. 1977). R. J. Bennett, Spatial Time SeriesAnalysi& London: Pion Ltd.,
[ 87 1 Y. Backlund and J. A. Babenko, Distibution system design using 1979.
computer graphics technique,presented atthe1979 IEEE J. H. McClellan, Design of 2-dimensional digital fdters by trans-
Power Industry Computer Applications Conf. forms, in h o c . 7th AnnualPrinceton Conf. on Information
[ 88) D. L.Wall et aZ., An optimization model for planning radial Science (Princeton, NJ, 1973), p. 247.
distribution networks,presented atthe IEEE PES Summer Computeraideddistribution planning and design, Westing
Meeting (July 1978), paper F78653-8. house Advanced SystemsTechnology,Pittsburgh, PA, Rep.
[ 891 T. Gonen, A research progress report for electrical energy dis- DB52-235, J a n . 1982.
tribution systems, Elec. Eng. Dep. Rep., Univ. of Oklahoma, H. L. Wiltis, Rapid calculation of proximities with FFTs, sub-
Norman, OK, Sept. 1978. mitted to IEEE Trans. S y s t , Man, Cybernetics.
[ 901 Survey of industry distribution planning needs, in EPRI Roj. C. Silva etal., Advanced distribution planning for Panama City,
RP-570, EPRI Rep. EL 1198, Palo Alto, 1979. presented at the 1981 IEEE Central America Power Conf. (Pan-
[ 9 1 ] Survey of Canadian distribution load forecasting, CEA Roj. ama City, Panama, July 9, 1981).
079D136, 1981 (project report forthcoming). J.E.D. Northcote-Green et al., Long range distribution plan-
[ 921 Survey of industry distribution planning needs and variations, ning-A unified approach, in h o c . 7th IEEEIPES T t D Conf.
Westinghouse AST Rep. AST-1667, 1982 (proprietary). (Atlanta, GA, Apr. 1979), IEEE Publ. 79CH1399-5-PWR.

You might also like