Regression

Multivariate
Modeling
An empirical or regression modeling approach is simple to develop and easy to use
compared to detailed hourly simulations of energy use in commercial buildings.
S. Katipamula Therefore, regression models developed from measured energy data are becoming
P a c i { iNc o r t h w eNs ta t i o n aL la b an increasingly popular method for determining retrofit savings or identifiing opera-
R i c h l a nWd , A9 9 3 5 2 tional and maintenance (O&M) problems. Because energy consumption in large
commercial buildings is a complex function of climatic conditions, building character-
istics, building urolr, system chaiacteristici and type of heating, ventilation, and air
T. A. Reddy conditioning (HVAC) equipment used, a multiple linear regression (MLR) model
D re x e provides better accuracy than a single-variable model for modeling energy consump'
Uln i v e rs i ty -tion.
D e p t.o f C i v ial n d AIso, when hourly monitored data are available, an issue which arises is what
A r c h i te c tu Eran gl i n e e ri n g time resolution to adopt for regression models to be most accurate. This paper
P h i l a d e l pPhA i1a9
, 104 addresses both these topics. This paper reviews the literature on MLR models of
buitding energy use, describes the methodology to develop MLR models, and high'
tights the usefulness of MLR models as baseline models and in detecting deviations
D. E.Claridge ii energy consumption resulting from major operational changes. The paper first
T e x a s A & M U n i v e r s i t ydevelops the functional basis of cooling energy use for two commonly used HVAC
Energy System Lab systems: dual-duct constant volume (DDCV) and dual-duct variable air volume
Col l e gSeta ti oTnX , 77843 (DDVAV). {lsing ihesefunctional forms, the cooling energy consumption in five large
commercial buitdings located in central Texas were modeled at monthly, daily, hourly,
and hour-of-day (HOD) time scales. Compared to the single-variable model (two'
parameter model with outdoor dry.bulb as the only variable), MLR models showed
a decrease in coefficient of variation (CV) between I0 percent to 60 percent, with
an average decrease of about 33 percent, thus clearly indicating the superioriry of
MLR models. Although the models at the monthly time scale had higher cofficient
of determination (R2) and lower CV than'daily, hourly, and HOD models, the daily
and HOD models proved more accurate at predicting cooling energy use.

I Introduction and Review of Literature ter change point model (piecewise linear model) to model elec-
tricity use of a grocery store. The basis of using a four-parameter
Regression analysis of measured data is becoming increas- change point model to model the energy consumption in a gro-
ingly popular as a means of developing baseline models of cery store is that below a certain To there is no cooling load,
energy consumption in commercial buildings, both for de- but there is still a refrigeration load and thus the change in the
termining retrofit savings and for identifying O&M problems. slope.
They are simple to develop and easy to use compared to other The single-variable regression models (based on just the I,)
modeling approaches, such as calibrated DOE-2.1 (Diamond, can adequately model energy consumption in residential build-
and Hunn, 1981; Copeland, 1983; Bronson et al., 1992; and ings (Fels, 1986) and, in some cases, commercial buildings
Kaplan et al., 1992) or simplified systems modeling (Katipa- (Kissock et al., 1993). However, the energy consumption in
mula and Claridge, 1993). large commercial buildings is a complex function of climatic
Regression models can be based on monthly (Fels, 1986), conditions (dry-bulb and dew-point temperatures, solar, . .),
daily (Kissock et al.,1992; Ruch and Claridge,1992), or hourly building characteristics (loss coeffi cients, heat capacity, internal
energy consumption data (Seem and Braun, 1991; Dhar et al., loads, .), building usage (l2-hour, or 24-hour, amount of
1994). A two-parameter linear model (E, : a * bT,,) is often fresh air intake, . .), system characteristics (total mass flow,
used to regress the daily or the monthly energy consumption hot and cold deck supply temperatures,economizer cycle,. '),
as a function of the outdoor dry-bulb temperature (2,) (Kissock and type of HVAC equipment used. Because some of these
et al., 1992), because significant variation in the energy con- parameters are difficult to estimate or monitor in an actual build-
sumption in buildings is due to the variations in the envelope ing, they cannot be used explicitly as variables in the regression
and ventilation loads, which are a strong function of I,. The models. Although the building use and operational parameters
Princeton Scorekeeping'Method (PRISM) uses a three-parame- (e.g., supply temperatures, fresh air intake) change from hour
ter model to model residential energy consumption (Fels, 1986). to hour, they are effectively constant from day to day. Therefore,
This model is based on the physical observation that the gains their contributions are minimal for models with monthly and
in a building preclude the necessity for heating until the outdoor daily time scales. The daily models are reasonable and reliable
temperature drops below a "balance temperature," which is if several months (six months or more) of data are used. lJnfor-
lower than the thermostat set point by an amount equal to the tunately, if only fewer than three months of data are available.
ratio of the gains to the loss coefficient of the building. Schrock the daily regression models may lead to significant errors in
and Claridge (1989) and Ruch et al. (1992) used a four-parame- prediction of the annual energy consumption (Rachlin et al.,
1986; Kissock et al., 1993: and Katipamula et al.' 1995a).
A number of researchers have shown the benefits of MLR
Contributed by the Solar Energy Division of The American Society of Mechani- analysis (Boonyatikarn, 1982: Sullivan and Nozaki, 1984; Les-
cal Engineers for publication in the ASME JounNa.l op Solen ENencv ENct-
NEERINc.Manuscript received by the ASME Solar Energy Division, June 1997;
lie et al.. 1986; Mazzucchi, 1986; Haberl and Claridge, 1987:
tinal revision, Apr. 1998. Associate Technical Editor: G. Vliet MacDonald, 1988; and Abushakra et al.. 1995). Boonyatikarn

'120 177
Journal of Solar Energy Engineering Copyright O 1998 by ASME A U GU S T ' 1998,V ol. |
 (1982) used MLR model to analyze energy use of 50 commer-
cial buildings in Michigan. The MLR model using 10 indepen-
dent variables (system type, refrigeration type, fuel type, num-
ber of hours equipment was used, exhaust air, supply air, cool-
ing degree-days, percent opaque surface, average shading in
winter, and average shading in summer) explained 93 percent
of the variation in the energy consumption.
Sullivan and Nozaki (1984) used the MLR approach to ana-
lyzethe effects of various fenestration parameterson the heating
and cooling requirements of commercial buildings. The input
databasefor the MLR analysis was created by a series of simula-
tions using the DOE-2.1B energy simulation program. Leslie
et al. (1986) used MLR models with heating degree-days, pro-
duction levels, and labor forces levels to model the monthlv
energy use for a military base. Mazzucchi (1986) applied MLR
models with 4 and customer count to model energy use in
restaurants. Haberl and Claridge (1987) used a daily MLR
model to predict the energy consumption at a recreation center.
Abushakra et al. (1995) used a MLR model to model the electric
demand in a large commercial building. The MLR model used
ambient conditions (f", wind speed and direction, barometric
pressure, and relative humidity) and electricity demand (chillers,
pumps, fans, lights and receptacles).
Although the investigators cited above have shown the bene-
fits of MLR models, none of them developed the models based
on engineering principles of the systems used in specific build-
ing. Models based on engineering principles capture the physi-
cal interactions taking place between the HVAC system and the
building. BecauseMLR models based on engineering principles
require knowledge of the HVAC system operation and its inter-
action with the other building systems, they are more difficult
to develop. Another major disadvantage of any MLR model is
the selection of independent variables. For the MLR analysis
to be robust, the variables should be independent of each other;
however, in reality, most variables used in modeling energy use
are correlated to some extent.
Katipamula et al. (1994, 1995a, and 1995b) developed MLR
models of energy use from measured energy data based on
engineering principles. The major conclusions from those stud-
ies are 1) the CV of the MLR decreasedbetween l0 percent to
66 percent when compared to the CV from a single variable
model; and 2) models at the monthly time scale had higher R2
and lower CV than daily, hourly, and HOD models, but the
daily models and HOD models proved more accurate at pre-
dicting cooling energy use. This paper summarizes those siud-
les.
The main objective of the paper is to provide a methodology
to develop MLR models based on engineering principles and
to highlight the effectiveness of MLR models as baseline models
and in detecting deviation in the energy consumption due to -
major operational changes. The major sections in this paper are
l) methodology for developing MLR models based on engi-
neering principles for modeling cooling energy consumption in
- Nomenclature
co : specific heat of air qr,": internal sensible heat gains
E: €Dergyuse qsor= total global horizontal solar radia-
GJ : gigajoules : 10eJoules tion
,d : mass flow rate Uo: overall heat transmission coeffi-
h, : latent heat of vaporization cient
1= indicator variable w : specific humidity of air
I: dry-bulb temperature w. : specific humidity of air leaving
7- : dry-bulb temperature at minimum cooling coil
flow condition
T. : cooling coil leaving air dry-bulb Subscripts
temperature c : cooling
Tr, : heating coil leaving air dry-bulb d = internal zone
temperature dp: autdoor dewpoint temperature
178 | Vol. 120, AUGUST 1998
large commercial buildings: 2.t applying rhe MLR moclels to
severallarge commercial buildings; and 3) selectionof suitable
time resolution for regression models.
2 Methodology of MLR Model Development
The development of regression moders is often based on both
theory and experience. In this section, the theoretical basis of
MLR model developmenris presenred.To simplify the analysis
the buildings are assumed to be conditioned 24 hours u duy;
therefore, the thermal storage effect is not significant. For details
on the functional basis of thermal energy consumption, refer
to Katipamula et al. (1994), Reddy er al. (1995), ind Knebel
(1983).
2.1 Cooling Energy Consumption With DDCV Systems.
In a building with a DDCV system, the airflow rate is constant
and the comfort condition of each zone is met by mixing the
hot and cold air streams. Therefore, the energy use of a building
with a DDcv system is always greater than rhe building loadi
(including ventilation loads) (Reddy er al., r995). In a multizone
building, it is not unrealistic to have one zone calling for heating
and another zone concurrently calling for cooring-. Therefore-,
the physical model should account for simultaneoui heating and
cooling.
Previous studiesby Knebel (1983), Katipamula and claridge
(1993), and Liu and claridge (1995) have shown rhar buildings
can be physically modeled as a two zone building, one exterior
(perimeter) zone and one interior (core) zone, with adequate
accuracy. Again, to simplify the analysis, both zones are as_
sumed to have same s€t point temperature, the total cooling
energy consumption of the building with a DDCV system is
given by
E,: Irh{p(To - Tr) + (U"(7" - T,) * 4,or * 4,.,.") * e,,,,af
-
:l - w,) (r)
* m,h.(w^
"\?(To - T,)
The contribution of the individual components of i can be
quantified by studying the functional relaiionship of each of the
major components of Eq. (1). Most commercial buildings in
hot and humid climates reset the hot deck supply temperature
(7'r') with ro and the cold deck temperature (2.) is a ionstant.
Because of the reset in 2,,. the E, may exhibit a piecewise linear
relationship as a function of To (Katipamula et al., lgg4).
Equation (1) can be directly used in the regression analysis
if measured values of all the variables are avallable. However,
Tn,7,, T^, T, w^, and lrc are seldom measured on a continuous
basis; therefore, an equation that is functionally equivalent to
Eq. (1) but with reduced variables is as follows:
E": q. * 0rT, + grl + p3IT,
* jo(w^ - w,) * /sei * goQ"or Q)
e: external zone
ft : heating
ln : mixed air
min : minimum flow for a VAV system
o : outdoor dry-bulb
s: coil surface
/ : total
u: ventilation or outdoor air intake
z: zone
Transactions of the ASME
h

where d, gt,0z, gzi 00,0r, and P6 are regressioncoefficients
and / is the indicator variable introduced to model the change
point behavior. The value of 1 is set equal to I for To-greatet
ihan the change point temperature and is equal to zero otherwise
(Daniel et al., f-qgO).If the change point does not exist in the
data, then the magnitude of the coefficients B2 and B3 will not
be statisticallY significant.
In most laige iommercial buildings, a major portion of the
- w,)
latent loads is due to ventilation; therefore, the term (w,,
- ?.)*, where T7o is^the
in Eq. (2) can be replacedby (Tq
outdoor air dewpoinf temperature and Z, is the mean surface
temperature of the cooling coil. The term is set equal to zero
when it is negative, as representedby the superscript "+"'
This is n"c".rary because when 7,1,is lower than I, there is no
latent cooling load even if the mixed air is saturated'
E, : d * 0rT" + gzl + P3IT, + goT\o * gsQ, * FeQ,or (3)
2.2 Cooling Energy Consumption With DDVAV
systems. tn a uuitaing with a DDVAV system,.the airflow
rate is modulated to meet the comtort condition of each zone.
Therefore, the energy consumption closely parallels thebuilding
loads (including ventilation loads) (Reddy et al., 1995). The
total cooling of a two-zone building with a DDVAV
"n".gy
system is given bY
E, : l(U"(T.
- Tr) + 4,or * 4t.,.")* q,.,.al
(T^ - T')
x * rn.h,.(w^ - w,) (4)
(T, - T,)
This expression, unlike the one for DDCV, shows a second-
- T.) is a con-
order reiationship with Z, because the term (7,
- T.) is a function of To (Katipamula et
stant and the term (T^
of (T^ - T") will be significant if
al.. 1994). The contribution
rhe amount of fresh air intake is high. In a DDVAV system,
the mass flow is modulated to meet the zone sensible cooling
load; however, there is a minimum flow rate below which there
is no modulation. Therefore, when the flow reaches the mini-
mum flow, the system essentially becomes a constant volume
system and the energy consumption shows a change in slope.
Again, due to lack of some measured values (such as occupancy,
inftrnal gains, erc.) a functional form equivalent t9 Eq. (4) is
needed fir regression. Also, if we assume that the outdoor
air intake is small (less than 20 percent of the total rated air-
- T,) will be close to unity'
flow), then the term (I- - T,)l(7,
of Eq. (4) in terrns of measured
The functional equivalence
variables is
E, : d * 0rT, + 0rl + P3IT' + gqfio + 7sQ, * 1oQ'or (5)
: a * grTZ + 7rTi, * 1zQiT" * 0+q'orT,i
(7)
when T, > T'" at min. flow
3 Application of MLR Models
In the previous section, a methodology was presented for
developing empirical models from the fundamental governing
equationsTor cooling energy consumption in commercial build-
ings with DDCV oi UOV4V systems. This methodology can
be adapted to other system typg-s (single-duct constant
volume with ieheat, single-duct variable air volume with re-
"uiity
heat), and other end uses (heating and electric energy use).
Relevant equations are presented in Reddy et al' (1995)'
In this section, summary results of modeling the cooling
energy consumption in five large commercial buildings at a
dailfiime scale are presented. For detailed discussion on effect
of time resolution on statistical modeling refer to Katipamula
et al. (1995b). The physical characteristics and operational de-
tails oi the five buildings are given in the Appendix. Among
the five buildings there are three buildings with -QpCV and
(EC)
three with DDVAV systems. One of the five buildings
was retrofitted from a DDCV system to a DDVAV system;
therefore, the building has data for both DDCV and DDVAV
system operation. Measured data over one entire year were used
to develop MLR models; the data included To, Too, QsosQi lrrd
E". The hourly values of E , qi and 4solwere summed over the
day to get d;ily values. The independent variab.le.l' T, and
*"r" averaged over the day to obtain average daily values.
fr'r,
""ftfln
analysis assumes that the regressor variables are inde-
pendent of eich other. Multicollinearity betw-eenthe regressor
variables results in large uncertainty bounds for the regression
coefficients, leading to model uncertainty' A rule -of thumb
(Draper and Smith, teat; is that multicollinearity effects may
U" lmportunt if the simple correlation between two variables is
largeithan the correlation of one or either variable with a depen-
deit variable. Earlier studies by Wu et al. (1992), Reddy and
claridge (1994), and Katipamula et al. (1994) showed that col-
lineariiy is not significanfbetwe"n To, fro and q; at a daily and
hourly iime scale, but can be significant between 7o and q,or'
In most commercial buildings, the cold deck supply tempera-
ture is held constant, at 12.8oC,throughout the year. Therefore,
the average surface temperature of the coil, 4, for all buildings
analyzedlexcept EC wiih DDCV) is assumed to be 12.8"C. In
the bC, a contlol problem (Katipamula, and Claridge' 199?
resulted in mean i values varying between 10'C and 21"C.
Therefore, for the EC with the DDCV system, the average coil
surface temperature was assumed to be 15.6"C'

The above equation is similar to Eq. (3) for the DDCV system.
However. if itre outdoor air intake is high then the functional
equivalent of Eq. (4) is (Katipamula et al', 1994):
E, : d * grT. + 7zTXo* ?zQi * ?cQ'ori
when T, = T'" at min. flow
3.1 DDCV Modeling. Daily cooling energy use in three
(3)' The
buitdings with DDCV systems was modeled using Eq'
,u,,'rnufu of the "stepwiie" regression results is shown in Table
l. In a siepwise regression, the contribution of each individual
variable to the model can be assessedby comparing the partial
(6) R2 values. A partial R2 of 87.1 percent for I, (Table l) means

Table 1 Summary of results for DDCV MLR models

EC WEL MSB
Partial .R' Model .R' Partial .R' Model .R' Partial Ii" Model f{'

Variable (To) (%) (To) (To) (%) (To)

To 87.1 87.1 6.6 6.6 4.5 4.5
ril 4.0 91.1 90.3 96.9 88.6 93.1
0.8 91.9 0.0 96.9 0.9 94.0
Qi
0.0 91.9 0.0 96.9 0.1 94.1
I
0.0 91.9 0.0 96.9 0.9 95.0
I *To
0.0 96.9 0.0 95.0
Qsol 0.0 91.9

A U GU S T 1998' V o l' 120 / 179

Journal of Solar Energy Engineering
 -EC It
-.....wEL IE
l8 -_ MSB 16
l6

l4
l4
t2

s"
vlo
t2

10
S"
-lo l0

E OE
u8 6
6
6
4

0 0 gr ltro I
o ---;---a; qr r.To r qd To t*'
Addition of Independent Variables to the Model
Addition of IndependentYariablesto the Model
variableswith DDVAV
Fig.2 changein cv with additionof independent
with DDCV
Fig. 1 Ghange in cv with addition of independent variables system
sYstems

DDCV models, all three DDVAV models show a presence

that the outdoordry-bulb temperatureexplains87.1 percentof of a change point in the data (both 1 and IxTo ate statistically
the variation in the cooling energyuse. significant).
For the EC building, T,lxplains most of the variationin the "In
g"n"rul, the cooling energy consumption of a building with
cooling energyuse followed by fio and internal gaini' At EC'
a ndVRV system is pioportional to the thermal loads on the
the hoideckilmperature is constantand the ratio of the glazed building. tn i VRV system, the air flow rate is modulated to
surfaceto the totil surfaceis small (about 22percent);therefore, meet th-e comfort conditions in the zone' while the airflow is
there is no changepoint in the model, and the contributionof constant in the DDCV system. The modulation in the airflow
the solar variablJto the model is statisticallyinsignificant(i.e', reduces simultaneous heating and coolingi thus, the cooling
partial R2 : 0).
'- energy consumption is a stronger function of the thermal loads
A .nuior portion of the conditionedspacein both.WEL and tti" building. Therefore, the independent variables TXoand q,
MSB UuitOlngs is madeup of laboratories;therefore,the outdoor on
-rrigtr is much variation in cooling energy consumption
air intake is (>70 percent).unlike the EC building, the explain twice
wiih the DDVAV system as compared to the DDCV system.
Do explainr *&" variati-onin the cooling energyconsumption In case of DDVAV, the decrease in CV with the addition of
tdan i, in both WEL and MSB buildings for 1fr9 following
independent variables is greater than DDCV models (Fig. 2).
."u.onJ' 1) outdoorair intake is over 70 percent(laboratories),
2) thereis more day-to-dayvariation of frpthT &:jytd 3) there
is a correlationbetweenfio andTo.Also, in the WEL building, 4 Effect of Time Resolution on Statistical
Modeling
does not show a change point because the reset in the data are available,an
the data when hourlyor ls-minutemonitored
hot deck temperatureis not large. However, the data from the issue that arises is what time resolution to adopt for developing
MSB building showsthe presenceof a changepoint' . iegression models (Katipamula et al., 1995b). The operational
The generaitrendsac.ois the threebuildings are similar, with 'hout p#ameters (such as ?:, T;,, outdoor air intake, Qi)change from
T" or fioexplaining most of the variation and the other variables to hour but are effectively constant on a monthly or even
onfy .oit ibuting t'iigtttty. The contributionfor solar radiationis a daily basis. Therefore, monthly and daily energy consumption
statisticallyinsignificant in all three buildings for two reasons: when compared to the hourly consumption show much less
it rh. r1nuttfraJtion (lessthan 20 percent)of externalsurfaceis scatter. Diimond and Hunn (1981) compared simulated energy
giii"g, and2) to a lesserextent,the correlationbetween7" and consumption to the measured data and found that on a monthly
with the addition time scale the deviations were higher than on an annual time
[*, ry"anAonet al., 1991).The decreasein CV
variables to the model is shown in Fig' 1'
oi'ina"p.ndent scale. They concluded that some effects average out over the
reducing the deviations.
3.2 DDVAV Modeling. Daily cooling energy use in year, -
with DDvAV systems was modeled using One methol to account for the hour-to-hour variation in the
three buildings-summary by the daily operational
Eq- fSl The of the stepwiseregressionresults is internal gains (which are influenced
model schedulelf the building) is to sort the hourly data into several
stro*n'in Table 2. Ali variableswere selectedby the giouping would be first to sort the data
T, explains more variations in the subgroups. A logical
*iit ,tt" exceptionof g,or.
intJweekauys und weekends and then sort these into 24 hourly
c oolingene rg y c o n s u mp ti o n th a n a n y o thervari abl e.U nl i ke

Table 2 Summary of results for DDVAV MLR models

--pc BUR WIN
Model B' Partial fd" Model J(- ranlar f[
Partial It'
Variable (%\ (To) (To) (vo) (To) (To)
83.2 83.2 79.1 79.1 84.1 84.1
To
n1+ 7.2 90.4 8.1 87.2 D.O 89.6
tdo

4.L 94.5 2.3 89.5 0.6 90.2

Q;
1.1 95.6 2.3 91.8 0.2 90.4
I
I *To 0.8 96.4 0.8 92.6 0.9 91.3
0.0 96.4 0.0 92.6 0.0 91.3
Qsot

Transactions of the ASME

18OI Vol. 120, AUGUST 1998
 20
- Monthly
l8
""' Daily
--- Hourly
l6
HOD
l4

S"
-lo E"
ur ur0
6

0
To Top* gi I*To I gsol To Top* gi I*To I gsol

(f)
(c)

20
Monthly It
Daily
Hourly l6
HOD
l4

E" Stt
-lo

ur0 ur
5

2
0

To Tap* gi I*To I 9sol To Tap* ft I*To I gsol

(e)

nl
- Monthly I
tr]
""' Daily
--- Hourly l6
t2
HOD
l4

xQ
i-'
l0

g
s"
-lo

u6 ur
6
4
4
2
0

To Top* gi I*To I gsol To Top* ei I*To I gsol

(d)
(a)

(b) WEL DDCV, (c) MSB DDCV'

Fig. 3 Ghange in CV with addition of independent variabtes to the model. (a) EC DDCV,
(df EC DDVAV, (e) BUR DDVAV and (4 wlN DDVAV

(lata sub-groups each. Monthly, daily, hourly and hour-of-day mula et al., 1995b).The general trends are similar to the results
rHOD) cooling energy use in three buildings (EC, MSB, and presented in the previous section. Also, the trends fbr models
\\'EL) with a DDCV system was modeled using Eq. ( I ) (Katipa- developed for the three buildings with DDVAV system (EC.

Journal of Solar Energy Engineering AUGUST 1998, Vol. 120 | 181

 models
Table 3 Comparison of the predictive ability of different

--Todeling
Period
rr^-+iGnof inn I Prediction Indices MonthlY Daiiy Hourly HOD
site 'vo 4.8 5.8 5.1
'92 Jatr. - Lrec. CV 8.4
MSB Jan. - Dec. -2.2 ,0.6 -0.7 -0.9
(DDCV)
MBE
-Jan. '93 9.1 9.1
'92 - Dec. CV 9.9 I.O
BUR fan. - Dec. -8.2 -6.5 -6.4 -I.l

(vAV) MBE
- uec. 'YJ CV 15.3 L5.2 t5.7 ID.O
WIN F-Pec.'92 Jan.
-t2.6 -12.0 -10.7 - 1 1 . 8
(vAV) MBE

h"#*ii"rT5;,;:ff ji:"''#[*ii
ffi :ffi ;ffi;,+:-#"tr"#.ii'i;;fl:f
iit;ft' withonlv ls93'theuairfldoworate:il*:lliti.Y')tJleautv
r'i!r'., In60"tn: andlower-cVtt'-'otr'"' modeli' *tv tnevne
righ and
*;ilm:';;'ru.;;;e;ii"" ri'lher
i:1"+::':J*:'"'#ll'j;:;*"fl,k;::t6yT:,ffi1'Jiff
flt**"Xl"Li"J'THlilfi$.,"'^ffiF?'::l'ifl";
5 concrusions
;;#l,l*ti:;,lt;fr;TflS^""fi:'1""'$l'.1fJt#i
tt"k,l',il?'itl',llT[J1""#'?kmut
#J,p.'jn#t*;i\B,i-l*trr[-f$i
both-had.lower
ill',3]l,.a more wirhapiecewise
accurarery linear
muliiple
i;i:;*:*llll*:';lr":','".,";qi::il'ilx'*
efj*'.l5*9"$l*"ru.''t&lTiJ'il1'lll'ili interactions
ofrhi ohvsical rn
prace
takine
"f,ffi;1f"il."flr;t'r"i1*xil""fi; cv thandailv
lower
riiid'iil;l,ilr.,rl*:'r[k#,i11,"#
mlJr;."iii :',"','"n;;x;"::"11""'"x1,iH:"$f;i'"BB
use;a" rxi;r,l xin::':mm:'T:;lill';iiill'-Tf
en€rgv
:ir.utllll:;x*:":ilni"- rharroo,ins
l.i,lt"l.r":1;';*t*;r*l;'m6*;31ffi1d: i**p"t+i+***-**'1il"Ti1"tr
;lsJ1?H::::1":'il:t=.-[ll"#il'o;*f,, .#i:]: ""1 rhan
formore the
e0percen,.or '-.oaccounted
r. and

mav
il::il il;,h,;;;de,s no, ' $L:ffi$tr5r-;a1q1,'1
H,H.l*#nkfltl#+i.":,pifFilrL#!fi
i"T*li j'ffiiiill"*llilI -,'*;u,i.nio.m,r,
consumption accurately. Also, on fi;bj;lhifffi:iffi:',f,: A
, l3J,:.1,i{:{J""'Hi'*iif*.r
frri';';{1"ili1+$,","?,i}:Hlf""'.1}xl':Lt"#fi
#if;fu{iny;yl,m;tJ:,9:,ffiffT#,#i,;::'#: ' Xihdtg"tsf,"el"_.**.*,:*
nTj,,}".y
""T3'"ll"li*.0,*n,'",:.:"p,i-orli,J":r,f:rffiiI',i3;$ ru,ure
predic,ing f;#f,tlg*;.*:i;j;ffj:J",
ent time scales.the modelsidenti use(Katipamula ut ' 'nnll'au
wereusedto predict
' "n"'nu
"ooting"n"tgy-"ottt"totbrr
"t
f9:-:.-:"::11 -
,n- thehourly
:*i:ll*lt*:i*l."}"T;fii'lrffllffitli;.ff"[fi H:P"'H:::lil!,li'J5if,::+11!!lngef
.l
r*""fi;*'*{d;'t'1itt}r$fi.iffi". $1Tfi'iiqffigg*yr
II *T'J3;"3i*,1'"*""t1,'.?3,f";;;;;iu"",-,(Mse)on
*tTi'l$ll"t
i#:iiCiy|*"hh:[.";.1;'#""",Ti1
oneadditional
have
#';'it'u. oo,,u,nr. vearor :;il"f i;'ffiiliv-,noa'ii *u' u"*"'n l-2p"'""nt

Itl; i'*,,Hil*l*,m:'';lfrti{d',1,#'$ffi
ii3ii.*riun.
:1','rl{mflx*lf#jittt**'"
""4
ij*g-r"*':l"i:l;
Th' dailv.models
wN uuitalngr. !l',
b:::

llrll"l;3liJ,'illllt;fifliif!]1ilj rabre
hisherthanthatof theDDCVMLF
theadvantase
4 summarizes ofthe
anddisadvantages

ffi,l""Pl"t'3;li;;.i?'i:"#fi',i:i::t'6'*x"$;HtXxf effort
11":"f"T3'$""" 48
it involvesdeveloping
because
itr" .o_ntnty ,"qoii"" tt" ,ri*i-um
air. Because7,, andfi,-" i'igiiv "oil"iut"o'ui
f,""iiiip*aii,"a"*r',",r'Jii#fiii':*"1"H"**til: f,i'*".",1i,1,?"1'il,tfiiq;fi11:.: lifill,"i:ii;
[*l**t""'l*tff"Yi*:'S#i:".fJ*':];':fx ilTfi,",t"lX.*::.l*U*lf::"ffiff-#'n
of the ASME
Transactions
1g2t vol.12o, AUGUST1998
 Table 4
Modeling
Effort
Metering and
Monitoring
Data Needed
for Robust 12 months $6 months
Modeling or more or more
Applicability
to Savings [n some In most
Measurements cases cases
Prediction
Uncertainty
o&M
Opportunities
Detection
Dynamic
Control
tion. An earlier study by Rachlin et al. (1986) suggestedthat
eight months of monthly data are sufficient to model residential
or small commercial energy consumption. However, to model
energy use in large commercial buildings on a monthly time
scale requires 12 months or more of data. To accurately predict
annual energy use using daily regression models, at least three
to six months of data are required (Kissock et al., 1993; Katipa-
mula et al., 1995).No analysishas yet been done to determine
how much data are required for hourly or HOD regression
models. Because the daily models require three to six months
of data. this period may be valid for hourly or HOD models as
well.
6 Acknowledgments
This study was funded by ( I ) Pacific Northwest National
I-aboratory which is operated by Battelle Memorial Institute
tbr the U.S. Department of Energy under Contract DE-AC06-
76RLO 1830. and (2) by the Texas State Energy Conservation
Of'fice of the Intergovernmental Division of the General Ser-
vices Commission (State Agencies Program) as part of the
LoanSTAR monitoring and Analvsis Program.
7 References
Abushakra. B.. Zmeureanu. R., and Fazio. P., 1995, "Evaluation of Conven-
tronal Inverse lvlodels For Predicting Electricity Demand of A Large Institutional
tluilding." Proceedings of the Second International Conftrence on lruloor Air
Qttttlitt, Ventilation. and Energt Consen'utiotr in Buildings, May 9- 12, Montreal.
C a n a d a .p p . 5 5 9 - 5 6 6 .
Boonyatikarn. S., 1982, "lmpact of Building Envelopes on Energy Consump-
titrt and Energy Design Guidelines." Pntceedings of the ASHRAE/DOE ConJer-
(nt ( ()tt Thermal Pe(ornnnce o.f thc Exterior Ent'elope of Buildings II, pp.469.
l]ronson, D.. Hinchey. S.. Haberl. J. S., O'Neal. D. L.. and Claridge. D. E..
1991. "A Procedure tbr Calibrating the DOE-2 Simulation Program to Non-
\\'cather Dependent Measured Loads," ASHRAE Tnuisttcttont Vol. 98. Part l,
pp.636-652.
(-opelandC.. 1983, "Retrofit Energy Studies Using the DOE-2 Computer Simu-
lrttion Program." ASHRAE Trunsactions. Vol. 89, Part lA, pp. 3.11-351.
Daniet. C.. and Wood, F. S.. with assistanceof Gorman. J. M., 1980. Fitting
I'.tluttriuts to Datu: Computer Arrailsls of Mult$actor Data,2nd ed., John Wiley
runrlSons, New York. NY.
Dhar. A.. Reddy, T. A.. and Claridge. D. E., 1994, "lmproved Fourier Series
'\pproach to NilodelingHourly Energy Use in Commercial Buildin g," Prcceedirtgs
,tl tlre 1994 AslvlE/JSES/JSES Internttrionul Solur Energy Conference, San Fran-
c i s c o .M a r c h ? 7 - 3 O , p p . 4 5 5 - 4 6 8 .
D i a m o n d . S . C . , a n d H u n n . B . D . , 1 9 8l . " C o m p a r i s o n o f D O E - 2 C o n r p u t e r
I)rogranrSimulations to Metered Data For Seven Comrnercial Buildings." ASH-
l l . \ l i T r u n n t - t i o r r s ,V o l . 8 7 . P a r t l . p p . 1 2 2 2 - 1 2 3l .
Journal of Solar Energy Engineering
Advantage and disadvantages of different models
Monthly Daily Hourly HOD
Minimum Minimum Moderate Difficult
Noner Required Required Required
&6 months &6 months
or more or more
AU All
cases cases
High Low Moderate Low
Most
Limited Possible Appropriate appropriate
No Possible Yes Best
Draper, N., and Smith, H., 1981, Applied Regression Analysis,2nd ed., John
Wiley and Sons. New York, NY.
Fels, M. (Ed), 1986, "Special Issue Devoted to Measuring Energy Savings,
The Princeton Scorekeeping Method (PRISM)," Energr* and Buildings, Vol. 9,
No. I and 2.
Haberl J. S., and Claridge, D. 8., 1987, "An Expert System for Building Energy
Consumption Analysis: Prototype Results," ASHPl/.E Transactions, Vol. 93. Part
I, pp. 979-998.
Kaplan, M., Jones, B., and Jansen, H., 1992, "DOE2.lC Model Calibration
With Monitored End-Use Data," The ACEEE 1992 Summer Study on EnergT'
Efficiency in Buildings, pp. l0.ll5-10.125.
Katipamula, S., Reddy, T. A., and Claridge, D. E., 1995a, "Bias in Predicting
Annual Energy Use in Commercial Building With Regression Models Developed
From Short Data Sets," Proceedings of the 1995 ASMilJSES/JSES International
Solar Energy Conference, Maui, Hawaii, pp. 99- I10.
Katipamula, S., Reddy, T. A., and Claridge, D. E.. 1995b, "Effect of Time
Resolution on Statistical Modeling of Cooling Energy Use in Large Commercial
Buildings," ASHRAE Transactions, Vol. l0l, Pall.2.
Katipamula, S., Reddy, T. A., and Claridge, D.8., 1994, "Development and
Application of Regression Models to Predict Cooling Energy Consumption in
Large Commercial Buildings," Proceedings of the 1994 ASMAJSES/JSES Inter-
national Solar Energy Conference, San Francisco, Meuch 27 -3O, pp. 307 -322.
Katipamula, S., and Claridge, D.8., 1993, "Use of Simplified System Models
to Measure Retrofit Energy Savings," ASME JounNx- op Solnn ENEncv Excr-
NEERTNc, Vol. I16, No. 2, May, pp. 57-68.
Katipamula, S., and Claridge, D. E., 1992, "Monitored Air Handler Perfor-
mance and Comparison With A Simplified System Model." ASHP,/^E Transac-
rions, Vol. 98, Part 2. pp. 341-351.
Kissock, J. K., Haberl, J. S., Claridge, D. 8., and Reddy, T. A., 1992. "Measur-
ing Retrofit Savings for the Texas LoanSTAR Program: Preliminary Methodology .
and Results," Proceedings of the ASME/JSES/KSES Internationul Solar Energy
Conference, Maui, Hawaii, pp. 299-308.
Kissock, J. K., Reddy, T. A., Fletcher, D., and Claridge, D. E., 1993. "The
Effect of Short Data Periods on The Annual Prediction Accuracy of Temperatures-
Dependent Regression Models of Commercial Buildings Energy Use." Proceed-
ings of the 1993 ASME/ASES/SED International Solar Energ;" Conference, Wash-
ington, D.C., April, pp.455-463.
Knebel, D. E., 1983, "Simplified Energy Analysis Using the Modified Bin
Method." American Society of Heating, Refrigerating. and Air-Conditioning En-
gineers. Inc., Atlanta, GA.
Leslie. N. P., Aveta, G. A., and Sliwinski. B. J.. 1986, "Regression Based
Process Energy Analysis System," ASHRAE Transactions, Vol. 92, Part lA. pp.
93-tOz.
Liu, M., and Claridge, D. E., 1995, "Application of Calibrated HVAC System
Models to Identiff Component Malfunctions and to Optimize Operation and Control
Schedules." Proceedings of the 1995 ASMAJSMUJSES htenutional Solar Energt
Conference, Vol. I, Maui, Hawaii, March 19-24, 1995, pp. 209-217.
MacDonald, M., 1988, "Power Signatures as Characteristicsof Commercial
and Related Buildings," Fifth Annual Svmpositun on Improving Building Ettergv
Eficiencv in Hot and Humid Climates, Energy Systems Laboratory, Texas A&Nl
University, College Station, Texas. pp. 80-87.
Mazzucchi, R. P., 1986, "The Project ot RestaurantEnergy Perlormance End-
Use Monitoring and Analysis." ASHRAE Trattsttt'tiotts,Vol. 92. Pan 28. pp.
328- 349.
A U GU S T 1998. V o l. 120 | 183
 R a c h l i n . J . . F e l s . i \ ' 1 . .a n d S o c o l o u . R . . l 9 t i 6 . " T h e S t a b i l i t y o f P R I S M E s t i -
mates." Energt and Buildings. Vol. 9. No. I and 2.
Reddy. T. A.. Katipamula. S.. Kissock J. S.. and Claridge, D. E., 1995. "The
Functional Basis of Steady-StateThermal Energy Use in Air-Side HVAC Equip-
ment," ASME JounNnl oF SoLAR ENencv ENGINEenI^-G.
3l -39. February..
Reddy, T.A., and Claridge. D.E., 1994, "Using Synthetic Data to Evaluate
Multiple Regressionand Principal Component Analysis for Statistical Models of
Daily Energy Consumption." Energ,t'and Buildings, Vol. 21, pp. 35-44.
Ruch, D., and Claridge, D. E.. 1992, "A Four Parameter Change-Point
Model for Predicting Energy Consumption in Commercial Buildings"'
A S M E J o u n N n l o F S o L A R E N e n c v E N c t N e E n t N cV
8.1. Februarv.
S c h r o c k . D . W . , a n d C l a r i d g c . I ) . E . . l 9 t t 9 . " l ' r c d i c t i n g [ i n e r g v U s ; ' t g ci r t r t
Superrrrarket."Pnx'ceding.sol'tlrc Si.tth Atrttttul.5r'lt7ro.rittlt ttrt ltrtprttring Builtl-
lrrg Sls/t'nr.rin Hot and Humid Clirttate,s,Dallas. Texas. October 3-'1. 1989. pp.
Ft9-F21.
Vol. ll7, No. I' pp. Seem. J. E.. and Braun, J. E.. 199 l. "Adaptive Methods lbr Real-Time Fore-
casting of Building Electric Demand," ASHRAE Transactiotts,Vol. 97. Part l,
pp.7lo-721.
Sullivan. R. T., and Nozaki. S. A., 1984. "Multiple Regression Techniques
Applied to Fenestration Effects On Commercial Building Energy Performance,"
ASHRAE Transactions,Vol. 90. Part lA, pp. I l6- 123.
Vandon, M., Kreider, J. S., and Norford. L. K.. 1991, "Improvement of the
, ol. l14, No.2,pp.77- Solar Calculations in the Modified Bin Method." ASHRAE Transacriotts, Yol-
97. Part 2.

APPENDIX

Table A Building characteristics (all buildings are institutional)

Supply Operation
Temp.(oC) People/
T"lMax. Tn HVAC (h/dav)

15.6/40.6

L2124

12124

L2.8120.4

10,310

1U I V ol. 120, A U GU ST 1 9 9 8 Transactions of the ASME