A.

Dhar
Enron Corporation
Modeling Hourly Energy Use in
1400 Smith St., EB884
Houston, TX 77002 Commercial Buildings With
T. A. Reddy Fourier Series Functional Forms
Drexel University
Civil and Arctiitecturai Engineering Dept. Hourly energy use in commercial buildings shows periodic variation in daily and
Ptiiiadelpiiia, PA 19104 annual cycles. Moreover, the pattern of variation is such that Fourier series functional
forms provide one of the best approaches for modeling this behavior. Results from
numerous case studies have demonstrated the power of Fourier series as a modeling
D. E. Claridge tool to represent hourly energy use in commercial buildings. This paper describes
Texas A&M iJniversity the Fourier series approaches that have been developed for modeling energy use in
IVIeclianicai Engineering Oept. commercial buildings and presents the results of application to data collected from
Coiiege Station, TX 77843
many sites.

1 Introduction modeling procedure are described followed by the results of

these procedures applied to data collected from various build-
The pattern of hourly energy use in almost all commercial ings in Texas.
buildings is periodic. Some uses, such as lighting and equipment
loads, vary periodically in daily and annual cycles with respect
to time and do not depend on ambient temperature and other 2 Literature Review
weather variables. Other uses, such as heating and cooling, are
weather dependent showing variation due to weather variables Weather variables, such as outdoor temperature and solar
such as outdoor dry bulb temperature, outdoor specific humidity radiation, are periodic and generally show patterns similar to
and solar flux along with the time-dependent periodic variation. those exhibited by hourly energy use in commercial buildings,
This behavior has been observed in the data obtained from many normally accompanied by phase shifts due to thermal inertia in
sites in Texas under the Texas LoanSTAR program (Claridge the buildings. Such climatic data have been analyzed using
e t a l , 1991). Fourier series by several researchers (Philips, 1984; Hittle and
Pedersen, 1981). The variation of the weather variables has
The nonlinear behavior of the hourly energy use in commer-
been represented by using daily and annual Fourier frequencies.
cial buildings can be modeled using nonlinear techniques such
as Fourier series functional forms. Artificial Neural Networks The use of sine and cosine basis functions to describe nonlin-
(ANNs) and Wavelet Analysis. While all these nonlinear tech- ear, periodic behavior of hourly energy use with respect to
niques provide good approximations, the Fourier series func- time was proposed by Pandit and Wu (1983). Seem and Braun
tional forms are particularly suitable because the periodic pat- (1991) followed by using a trigonometric form assuming a
tern of variation in hourly energy use is such that the Fourier weekly periodic behavior of hourly energy use. The predictive
frequencies are more suitable than any other type of basis func- ability of this model was relatively poor. Dhar et al. (1994)
tions. This statement will be supported by comparing the results hypothesized that this was because commercial buildings un-
of application to actual data later in this paper. dergo major operational changes from weekdays to weekends
and proposed the separation of data into several groups using
The hourly energy data from commercial buildings are uni-
a simple, calendar-based day-typing technique and developing
formly distributed when plotted in the form of a time series. It
a separate model equation for each day-type. This method pro-
can be proved that if there are {2n + 1) uniformly distributed
vided very good approximations to weather independent and
data points within a fixed interval of 0 < .x < 2, then there is
weather dependent energy use in many buildings. Further devel-
always a unique trigonometric (Tolstov, 1962) polynomial of
opments in the use of Fourier series functional forms were
order n or less that takes prescribed values at these points.
(Dhar et al., 1995a and 1995b): (1) generaUzation of the Fou-
Such a trigonometric polynomial is also called the finite Fourier
rier series equation for modeling weather independent and
weather dependent energy use and (2) introduction of double
Fourier series functional form for a Fourier series model which
F{x) = ao/2 -I- X [«t sin {kx) + bk cos (kx)] (1) represents nonlinear time and temperature dependencies. Addi-
tional details are available in Dhar (1995). These Fourier series
The Fourier series model equations described in this paper approaches developed to model hourly energy use in commer-
for weather independent and weather dependent energy use are cial buildings will be described and illustrated in the following
essentially variants of Eq. (1). sections.
It is important to remove the effect of major operational
changes from the data by creating subsets, which are called
day-types, and then develop final model equations for each day- 3 General Procedures for Fourier Series Modeling
type. Creating the day-types and determining the final model The first step in Fourier series modeling is to perform day-
equations for the daytypes are all parts of the modeling proce- typing. The entire data set is segmented into several groups
dure. In the following sections, the model equations and the according to operational characteristics of the building systems.
For example, some of the lights and HVAC equipment that
operate during the weekdays may be shut off during the week-
Contributed by the Solar Energy Division of The American Society of Mechani- ends. Therefore, a simple grouping of the data set can be week-
cal Engineers for publication in the ASME JOURNAL OF SOLAR ENERGY ENGI-
NEERING. Manuscript received by the ASME Solar Energy Division, Jan. 1996; days and weekends. The procedure for segmenting the data into
final revision, May 1998. Associate Technical Editor: G. Vhet. multiple groups is called day-typing and the calendar as well

Journal of Solar Energy Engineering Copyright © 1998 by ASME AUGUST 1998, Vol. 1 2 0 / 2 1 7

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 11/29/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 as certain statistical criteria are used to accomplish this. A model and
equation for each day-type is then developed separately. 'muji Ami
, . 27r , , . 27r ,
3.1 Day-typing. The day-typing is started with a primary z{d,h) = E X a>: sm — d + Wi sin — d
i=0 J-0 P, ^ Pi
grouping of the entire data set based on the calendar. Weekdays,
weekends and holidays, Christmas and other periods such as
semester breaks in universities are initially considered. Dun- . 27r , J, 2TI ,
77, sin — h + L, cos — h
can's multiple range test (Ott, 1988) is then performed. The " P, ^' P,
day-groups with statistically insignificant differences in mean
energy consumption are aggregated together to arrive at primary
There are 24 hourly observations (subscript /?) in a daily cycle
day-types. A histogram of the amplitudes of each important
and 365 days (subscript d) in a year which means that the
frequency is developed and checked for multimodal behavior.
values of (max andy'max are (365/2 - 1) = 181 and (24/2 - 1)
The important frequencies are those which appear consistently
in a particular type of building. In general, first sine and cosine, = 11 respectively. The expressions for P, and P, are
second cosine and fourth sine frequencies are important for P. = 3 6 5 ^^^ ^^^24
lights and equipment energy use in institutional buildings (Dhar (3)
etal., 1994). i J
Only if (1) the histogram is multimodal and (2) a physical for daily and annual cycles respectively.
reason can be attributed to such a distribution is it recommended
that the particular day-type be divided into groups. The last step 4.2 A Detailed Case Study. Modeling "Whole Building
of day-typing is to repeat the Duncan test and aggregate the Electric" energy use (^wbe) in ZEC in 1992 is presented as an
day-types with statistically insignificant differences in mean example of a detailed case study. ZEC is a large institutional
energy consumption. building containing class rooms, labs, offices and computer
facilities. It should be noted that heating and cooling for this
3.2 Selection of Principal Frequencies. Fourier series building are provided by a central campus plant; hence, £WBE
functional forms are regressed for each day-type to determine for ZEC does not contain these major weather-dependent uses
the coefficients of the frequencies. However, the functional form as would be the case in most buildings.
contains a large number of frequencies and the important fre- One year of E^y,^ data was grouped as follows by using the
quencies need to be identified to obtain model equations with calendar: (i) weekdays, (ii) weekends, (iii) holidays (F* Jan.,
fewer regressor terms. There can be several ways to do this; Spring break: 16* Mar., to 20* Mar., 3"* Jul., Thanksgiving:
one popular statistical way is stepwise regression. Dhar et al. 26"' to 29'*' Nov.,), referred to as the holiday group and (iv)
(1994) proposed the use of stepwise regression with the forward 23"* to 31"' Dec, referred to as the Christmas group. Duncan's
selection procedure to identify the important frequencies. The multiple range test was then performed which suggested three
initial set of frequencies is decided by Mallow's Cp (coefficient primary day-types: (i) weekdays, (ii) weekends -f- holidays
of parameters) criteria. The value of Cp decreases with the (will be called weekend day-type) and (iii) Christmas. The
increase in number of terms (or parameters) and becomes al- Christmas period shows separate day-type behavior in this case
most equal to the number of terms (Cp = p) at a certain point because the entire university closes for this nine-day period and
which is used as the cutoff (Ott, 1988). The number of frequen- activity level is noticeably lower than the activity levels during
cies is reduced further by using an arbitrary but reasonable other holidays. The histogram of important frequencies (1"' sine
partial i?-square cutoff criteria (e.g., partial if-square < 0.005). and cosine, 2""^ cosine and 4 * sine) was developed and the 1"'
cosine frequency suggested that the weekdays be divided into
two groups (Dhar et al., 1995a). This was also physically mean-
4 Fourier Analysis of Weather Independent Data
ingful in the sense that the days that fell in the semester group
The weather independent data are modeled by a Fourier series were the semester break days. The four final day-types obtained
with hour of day and day of year as the independent variables. in this way were: (i) weekdays school-in-session, (ii) week-
The development of the model equation, a detailed case study ends, (iii) semester break weekdays and (iv) Christmas.
and the results of several case studies are described in the fol-
Equation (2) was then used to develop a model equation for
lowing subsections.
each day-type. Mallow's Cp statistic suggested a 29-parameter
4.1 Model Equation. The variation of hourly weather in- model (Table 1); however, the partial /?-square cutoff criteria
dependent energy use generally shows periodicity with respect brought the number of important terms down to seven for the
to time in the daily and annual cycles for the cases treated in weekday model equation. The 7?-square and Coefficient of Vari-
the literature. The variation due to other cycles is generally not ance (C.V.) of the fit for the four day-types are given in Table
significant for most commercial buildings. A model equation 2. The C.V. values are between 3.8 percent and 9.5 percent
that accounts for the variation in these two cycles is, therefore, indicating very good model fits. The /^-square for Christmas is
presented. If the pattern shows the existence of any other set 0.33 which is relatively low. The data for Christmas has less
of important cycles, a similar approach can be taken to obtain scatter around the mean resulting in a poor 7?-square value. The
the preliminary model equation with unknown coefficients. The C.V., however, is below 5 percent which is comparable to other
model equation that accounts for variation in daily and annual day-types. The time series plots of measured and residual (mea-
cycles and the interaction between these two cycles Dhar et al. sured-predicted) E^ba are shown in Fig. 1.
(1994) is as follows:
4.3 Results From Other Case Studies. Fourier series
Ea.,^X{d) + Y{h) + Z{d,h) + e,j, (2) modeling of weather independent data from eighteen sites was
performed. The weekday models for different periods at differ-
where ent sites were studied and the results showed significant consis-
max tency with a small set of frequency terms being important. The
. 27r , J. 27r ,
X(d) ---1 -y, sin — a + o, cos — a four important terms identified for the institutional buildings
i=0
. Pi Pi . weekday energy consumption are 1" sine and cosine, 2"'' cosine
and 4"' sine frequencies. These frequencies correspond to 24
-'max
hour, 12 hour and 6 hour periods, representing day-time vs.
• 27r , _ 27r , night-time variation of energy use and the drop during lunch
Y(h)-- a, sin — « + PI cos — n
P P breaks. For more details, refer to Dhar et al. (1994).

218 / Vol. 120, AUGUST 1998 Transactions of the ASME

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 11/29/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Table 1 Summary of the forward selection procedure for whole building electricity use during school-in-
session weel(days in ZEC. Data period covers the calendar year 1992. Standard errors of all variables are
statistically insignificant (less than F = 0.0001). CHi (& SHI) and CDI (& SDI) represent the /th frequency of
the cosine (and sine) terms of the diurnal cycle and the annual cycle respectively.

No. of parameters (p) Variable entered Partial R^ Model R^ Mallow's CD

2 CH1 0.6092 0.6092 43582.1
3 SHI 0.2670 0.8762 10497.9
4 CH2 0.0413 0.9175 5388.0
5 SH4 0.0123 0.9268 3863.5
6 SH3 0.0068 0.9366 3024.4
7 SD1 0.0061 0.9427 2274.2

The /f-square and C.V. values of twenty model equations 5.1 The Temperature-Dependent Model. Temperature-
developed using data from four sites are listed in Table 2. The dependent Unear (2-Parameter type) and segmented linear models
data for both pre-retrofit and post-retrofit periods have been (3-Parameter and 4-Parameter type) have been developed and
modeled. The C.V. values range from 3.8 percent to 9.6 percent used for modeling daily energy use in commercial buildings for
while the /?-square values range from 0.27 to 0.96. The results retrofit savings analysis (Claridge et al., 1991). However, model-
in general indicate extremely good model fit. The /{-square ing weather dependent energy use at the hourly level should con-
values are low for certain day-types because of less scatter of sider two important factors. (1) weather dependent energy use in
energy use around the mean for those groups. commercial buildings often exhibits a nonlinear relationship with
outdoor temperature. Three- and four-parameter model behavior
can be observed in the data obtained from different sites (Katipa-
5 Fourier Analysis of Weather Dependent Data mula et al, 1994, 1995) and a generic nonlinear functional form
may be more appropriate, (2) Interaction between outdoor temper-
Weather dependent energy uses, such as hourly cooling and ature variation and time (hour-of-day) due to factors such as build-
heating, vary in daily and annual cycles and also are affected ing thermal mass and HVAC systems characteristics is significant
by outdoor dry bulb temperature, outdoor specific humidity and at the hourly level. The final model equation described in this
horizontal solar flux. Two model equations, known as the Tem- section has been developed from these perspectives.
perature-based Fourier Series (TFS) and the Generalized Fou-
First, we describe a Fourier series functional form to represent
rier Series (GFS), have been developed for modeling hourly
the nonlinear relationship between weather dependent energy
weather dependent energy use. The GFS equation uses outdoor
use and outdoor temperature:
dry bulb temperature, outdoor specific humidity and horizontal
solar flux as the input weather variables assuming linear depen- (2irm{T - r^n)
dence, while the TFS equation uses outdoor dry bulb tempera- Er = a' + P'T + X a™ sm
ture as the weather variable, but treats nonlinear dependence. Ar
The TFS equation sacrifices little in terms of prediction accu-
racy when compared to the GFS equation and can be used when (lirmjT - r^iJ
humidity and solar data are unavailable. + Pm cos + £7 (4)
Ar

Table 2 Fourier model results of weather independent energy use at four educational buildings In Texas

Building Energy Period Data Period Day-type R* C.V.

Name Use (%)
Lighting Pre- 9/1/89 to Weekdays, school-in session 0.96 4.1
ZEC retrofit 12/31/89 Weekends 0.64 5.1
(4 months) Christmas 0.87 8.4
Whole Post- 1/1/92 to Weekdays, school-in session 0.94 3.8
ZEC building retrofit 12/31/92 Weekends 0.41 4.4
electricity (12 months) Semester-break weekdays 0.84 9.5
Christmas 0.33 4.7
MSB Whole Pre- 9/1/91 to Weekdays, school-in session 0.95 3.3
building retrofit 8/6/93 Weekends 0.36 4.5
electricity (36 months)
Whole Post- 8/7/93 to Weekdays, school-in session 0.93 4.5
MSB building retrofit 11/30/94 Weekends 0.61 4.6
electricity (16 months) Christmas 0.70 7.8
Whole Post- 1/1/93 to Weekdays, school-in session 0.90 5.0
WEL building retrofit 12/31/93 Weekends 0.68 5.3
electricity (12 months) Semester-break weekdays 0.73 7.1
Christmas 0.79 5.2
Whole Post- 1/1/93 to Weekdays, school-in session 0.82 6.9
PAI building retrofit 12/31/93 Weekends 0.27 6.8
electricity (12 months) Semester-break weekdays 0.44 9.6
Christmas 0.34 5.7

Journal of Solar Energy Engineering AUGUST 1998, Vol. 120 / 219

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 11/29/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Jo»-%. — Jujo.

mm M
GOO '•v^oosnj^od daia

3 OO
r osiduiol
O
• ugB<VVW^']*vi<V*Y^
JAM TEB MAR APR

(a)
1 SCO

1zoo -

9 0 0

rrtoosurocj eiola
6 OO -

3 0 0 -
r «»lcjwal
' . ^ K ^ ^ /•^./'V ^^V.^.^I ^ ^ . A ^^ ^'^J-^^J^ • ,^^^^/^^VM^y, .,n^.,^^_^^y.U/>A>lAX^n^ • .

2 0 2 5

(b)
Fig. 1 ( a ) - ( b ) Time series plots measured and residual whole building energy use in ZEC for (a) Jan.-Jun.
1992 and (b) May 1992

where T = (Tmax - T^m) is the range within which the outdoor would be (100/2 - 1) = 49 if there were data points at 100
temperature varies in a particular geographic location. One can discrete temperature values, but results have shown that includ-
observe that the right hand side of Eq. (4) adds a temperature ing only the first few frequencies (approximately five) in the
based Fourier series to a linear functional form ofa'+P'T.ln equation are needed for modeling weather dependent energy
this way, Eq. (7) fits both hnear and nonlinear cases efficiently. use.
The next step is to modify Eq. (7) to incorporate the effect
of the interaction between hour of day and outdoor temperature. 5.2 The Multivariate Model. The TFS equation is useful
This is accomplished by taking the product of the time series when only temperature data are available, the data of other
and the temperature series which, after a few rearrangements, weather variables (humidity and solar radiation) being unavail-
results in Eq. (5) described below. For further details on the able because either those were not measured or the measuring
derivation, see Dhar et al. (1995b). instruments failed. However, if all the weather data is available,
a more rational approach would be to use a model equation
. /2irnh\ , flirnh with all the weather variables as the regressors. The weather
ET.,. = r I <p„ sin + w„ cos dependent energy use at each individual hour can be represented
^ V 24 / ^ V 24 as a weighted linear combination of outdoor dry bulb tempera-
ture ( r ) , outdoor specific humidity difference QN^) and hori-
/iTrmjT - Tmin) zontal solar flux (/) (Katipamula et al., 1994, 1995):
ri,„ sin
AT
EH = a,, + bhTh + ChW t + d,Jh + e* (6)
2Trm(T - T^i„) W^ is expressed as W * = W - W^c where Wcc is the saturated
+ Cm cos
AT specific humidity at the cold coil surface temperature of the air-
conditioning system and can be assumed to be equal to 0.0092
pounds water per pound of dry air for normal air-conditioning
. (2'Knh\ r /2nnh\
purposes. The above equation assumes that the energy use varies
X X y„ sin I + o„ cos
V 24;
K-w) + tT.H (5)
linearly with outdoor temperature, humidity difference and solar
flux. This assertion has been supported by a detailed study
The number of maximum allowable frequencies n^ai and Wmax conducted by Kissock (1993), which indicated that higher order
depend upon the number of distinct data points within the time terms contribute insignificantly to the prediction and, therefore,
and temperature periods. There are 24 hours in a day, which can be neglected for practical purposes. The above equation
means n^ax = (24/12 - 1) = 11. For temperature series, Wmax can be generalized for all hours of the day by allowing the

220 / Vol. 120, AUGUST 1998 Transactions of the ASME

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 11/29/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 coefficients to vary from hour to hour. This can be efficiently of energy use during weekends is humidity (partial /{-Square
done by using Fourier series. The final equation takes the fol- is 0.77) with model /{-square and C.V. of 0.92 and 12.02 percent
lowing form (Dhar et al., 1995a): respectively, while the TFS model has /{-square = 0.84 and
C.V. = 14.18 percent. The nonlinear temperature terms of the
E,., = I k[X{d) + Y{h) + Z{d, h)] + £,,, (7) TFS model are able to represent much of the humidity effect
on energy use because of the strong multicollinearity between
where the summation includes terms for each of fc = 1, T, W *, outdoor humidity and dry bulb temperature. This multicollinear-
and / and ea,h is the error term. The functional forms ofX{d), ity also means that while the models provide a good representa-
Y(h) and Z{d, h) have been provided in section 4.1. It may tion of the combined effects of temperature and humidity, indi-
be noted that while the GFS approach is intuitively more com- vidual coefficients (e.g., the strong humidity dependence in the
plete than the TFS approach, results discussed in subsequent GFS weekend model) may not be physically significant. The
sections show that the nonlinear temperature terms included in model fit of the TFS equation to ZEC cooling energy use is
the TFS approach of Section 5.1 are often more important than shown in Fig. 2.
the nontemperature dependent terms added in Eq. 7.
5.4 Results from Other Case Studies. The GFS and TFS
5.3 A Detailed Case Study. We will present the details models were developed for weekdays and weekends for heating
of modeling cooling energy use in ZEC from Jan. through June, and cooling energy use and weather dependent whole building
1992. As stated earlier, the first step of Fourier series modeling electric energy use (includes chiller electricity consumption) at
is day-typing. Although an elaborate day-typing can be done five sites. Sixteen model equations were developed using the
on the residual of Eq. (6), fitted to the weather dependent data from these five sites, /{-squares and C.V.s of these model
data, a mere separation into weekdays and weekends has given equations are shown in Table 4. GFS cooling energy models
significantly improved C.V. and R^ of model fit in many cases have slightly lower C.V.s. However, TFS though unable to
(Dhar et al., 1995a). Weather dependent energy use depends model cooling energy use as accurately as GFS, nevertheless
on many factors such as building construction, HVAC plant captures most of the variations due to humidity and solar effects.
and systems, etc., in addition to the weather variables. The But the TFS provided better models of heating energy use than
pattern, therefore, is often less sensitive to minor operational the GFS models, as indicated by lower /{-squares and C.V.s in
changes than to weather changes. As a result, considering addi- all the cases. TFS model performance was similar to that of
tional day-types, as in the case of weather independent energy GFS models for whole building electric energy use, the only
use treated in Section 4.2, is not useful. exception being the weekends in MCC, where the TFS model
Stepwise regression was performed using TFS (Eq. (5)) and is significantly better than the GFS model. This is physically
GFS (Eq. (7)) equations to identify respective sets of important reasonable since heating energy use does not depend on outdoor
parameters. The important parameters and respective partial R- humidity in a significant way unless humidification is provided,
square values for weekday and weekend cooling energy use in and the TFS approach treats nonlinear temperature dependence,
ZEC are given in Table 3. For the weekdays group, final GFS which is not treated in the GFS approach presented here.
and TFS models each have five parameters while the final week-
end models have three and seven parameters respectively.
The /{-squares of the GFS and TFS weekday models are the 6 Comparison With Other Nonlinear Models
same (0.89) and the C.V. values, although not identical, are The ability of nonlinear models such as TFS, ANN with a
quite close (12.45 percent and 12.05 percent respectively). The backpropagation algorithm (BPN) and WaveNet (An Artificial
humidity term appears to be important in the GFS equation. Neural Network with wavelet basis functions) (Dhar et al,
Since /f-square and C.V. values of the two weekday models are 1995c) to predict heating and cooling energy use and chiller
very close, it can be seen how nonlinear temperature terms in electric energy use were compared using data from five sites.
the TFS equation can represent a significant portion of the varia- The BPNs had (i) three input nodes feeding hour of day, out-
tion due to humidity. This can also be observed in the weekend door temperature and day of week inputs, (ii) two hidden layers
models. The GFS model indicates that the most important driver each having ten hidden nodes, (ii) one output node predicting

Table 3 Partial f7-squares of the individual parameters of Generalized Fourier Series (GFS) and Temperature based Fourier
Series (TFS) models for cooling energy use In ZEC from Jan. through Jun. 1992. CH1 and SIH1 represent the 1st frequency of
the cosine and sine terms of the diurnal cycle while CT1 and CT2 represent the temperature based first two cosine terms in
Eq. 4

Day-type Variable Partial R-square Cumulative R-square

GFS Model TFS Model GFS Model TFS Model
T 0.8042 0.8149 0.8042 0.8149
W+ 0.0619 0.8661 0.8149
Weekdays T*CH1 0.1790 0.0070 0.8884 0.8219
l*CH1 0.0082 0.8922 0.8219
ST1 0.0537 0.8922 0.8758
ST1*SH1 0.0190 0.8922 0.8946
T 0.1420 0.6936 0.1420 0.6936
W+ 0.7780 0.9200 0.6936
Weekends CT1 0.0463 0.9200 0.7399
ST1*SH1 0.0661 0.9200 0.8080
T*SH1 0.0173 0.9200 0.8253
T*CH1 0.0107 0.9200 0.8360
CT2*SH1 0.0078 0.9200 0.8468

Journal of Solar Energy Engineering AUGUST 1998, Vol. 120 / 221

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 11/29/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 — 4 -<
JAN FEB MAR APR MAY JUN

Fig. 2 Time series piots of measured and residual cooling energy use in ZEC from Jan. through Jun. 1992
when the TFS model is used

the weather dependent energy use, (iv) gain = 0.9, learning chilled water lines. This probably accounts for the better values
rate = 0.1 and bias = 0, (v) sigmoid activation function and of ZEC chilled water CV for this period obtained for all of the
(vi) a normalization range of 0 to 0.9 for all the inputs and entries in Table 5. The heating data experienced only minor
output. WaveNet is a one hidden layer neural network with a impact when the chilled water lines broke, and so this compari-
variable number of hidden nodes. The number of hidden nodes son is probably more meaningful. It may be noted that TFS
of the WaveNet were optimized using the stepwise regression models generally have fewer parameters than other models,
approach. The wavelet basis functions were generated from suggesting that TFS will often be a more accurate approach.
cubic splines (Chui, 1992).
When the model fit of the TFS equation is compared with 7 Summary and Conclusions
BPN and WaveNet, the C.V. values of TFS models are found The Fourier series functional forms have provided excellent
to be the lowest in 4 out of 8 cases (Table 5). Comparison of model fits in several cases described in this paper. However, it
ZEC cooling and heating results for 9/1/89-12/20/89 with is important to investigate the strengths and weaknesses of these
the results of winners in the ASHRAE Predictor Shootout I models in order to be able to provide suitable guidelines as to
(Kreider and Haberi, 1994) competition is likewise favorable. when these equations should be used for modeling hourly en-
The CVs for the top five cooling predictors ranged from 12.78 ergy use in commercial buildings. Fourier series models are
percent - 14.32 percent compared to 7.30 percent and 8.03 very suitable for modeling weather independent data unless the
percent for the GFS and TFS approaches. The CVs of the top daily variation of energy use is very irregular, as is sometimes
six heating predictors ranged from 15.24 percent to 31.65 per- true in primary and secondary schools. If higher frequencies
cent with only one below 28.08 percent while the GFS and TFS are present, then a local basis function approach, such as wave-
values shown are 20.88 percent and 19.56 percent respectively. lets, will do significantly better than Fourier series. On the other
It should be noted that the Predictor Shootout I data set covered hand, Fourier series will provide a better prediction with fewer
9/1/89-12/31/89 and hence includes 11 additional days. The terms in the model if the variation is harmonic or close to
11 days at the end of the year were not included in the GFS harmonic.
and TFS tests since that period contained two days of chilled When modeling weather dependent energy use, the Fourier
water data that were abnormally low due to a freeze and broken series models can be recommended in general. The variation of

Table 4 Comparison of /7-square and C.V. values of Generalized Fourier Series (GFS) and Temperature
based Fourier Series (TFS) models of weather dependent energy use in five buildings

Site Energy Use Period Day-type R2 C.V'.(%)

GFS TFS GFS TFS
ZEC Cooling 01/07/92- Weekdays 0.89 0.89 12.45 12.05
(GJ/h) 06/30/92 Weekends 0.89 0.84 12.05 14.18
ZEC Cooling 09/01/89- Weekdays 0.87 0.83 7.70 8.42
(GJ/h) 12/20/89 Weekends 0.91 0.89 6.30 7.07
TDH (Lab Cooling 02/16/91- Weekdays 0.85 0.81 17.07 18.75
& Main) (GJ/h) 08/12/92 Weekends 0.82 0.80 17.85 18.02
TDH Heating 02/16/91- Weekdays 0.81 0.82 20.96 20.60
(GJ/h) 08/12/92 Weekends 0.73 0.80 24.55 21.32
ZEC Heating 09/01/89- Weekdays 0.90 0.91 20.80 19.37
(GJ/h) 12/20/89 Weekends 0.87 0.89 21.10 20.10
BUR Heating 01/01/92- Weekdays 0.87 0.90 17.03 14.94
(GJ/h) 06/30/92 Weekends 0.86 0.90 16.59 14.57
TCOM (Med. Whole building 06/01/93- Weekdays 0.85 0.83 6.17 6.70
bidg. 1 & 2) electricity 08/31/93 Weekends 0.46 0.47 8.68 8.64
MCC including chiller 04/07/92- Weekdays 0.89 0.90 13.88 13.02
(kWh/h) 05/15/92 Weekends 0.80 0.92 11.96 8.69

222 / Vol. 120, AUGUST 1998 Transactions of the ASME

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 11/29/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Table 5 Comparison of C.V. values of four different modeling approaches applied to eight data channels
from five sites In Texas. The shaded boxes Indicate the lowest C.V. values among the temperature based
modeling approaches (TFS, BPN and Wave-Net)

Site Energy Use Period C.V. RMSE (%)

GFS TFS BPN Wave-Net
ZEC Cooling 01/07/92- 12.32 12.66 1074 12.69
(GJ/h) 06/30/92
ZEC Cooling 09/01/89- 7.30 8 03 632 7.72
(GJ/h) 12/20/89
TDH (Lab Cooling 02/16/91- 17.29 18S4 19 12 18.78
& Main) (GJ/h) 08/12/92
TDH Heating 02/16/91- 21.98 20.80 14.16
(GJ/h) 08/12/92
ZEC Heating 09/01/89- 20.88 19.58 16.96 18.34
(GJ/h) 12/20/89
BUR Heating 01/01/92- 16.90 14 83 16.60 22.39
(GJ/h) 06/30/92
TCOM (Med. Whole building 06/01/93- 6.88 725 7.27 7.43
bidg. 1 & 2) electricity 08/31/93
MCC including chiller 04/07/92- 13.33 1178 14.79 12.40
05/15/92

weather dependent energy use is driven by a large number of the 1995 ASME/JSME/JSES International Solar Energy Conference, Mawuii,
Mar. 27-30, Vol. 1, pp. 125-134. Also accepted for publication in ASME JOUR-
parameters. However, results show that a nonlinear temperature NAL OF SOLAR ENERGY ENGINEERING.
dependent Fourier series model (TFS) gives very good predic- 6 Dhar, A., Reddy, T. A., and Claridge, D., 1995c, "An Artificial Neural
tion in most of the cases examined. TFS is a very efficient way Network With Wavelet Basis Functions to Predict Hourly Heating and Cooling
of capturing this nonlinear variation, since only a few terms of Energy Use in Commercial Buildings with Temperature as the Only Weather
the TFS equation are adequate to represent the variation and Variable," Proceedings of the 1995 ASME/JSME/JSES International Solar En-
ergy Conference, Mawuii, Mar, 27-30, Vol. 1, pp. 145-156.
often give better results than other nonlinear techniques. 7 Dhar A., 1995, "Development of Fourier Series and Artificial Neural
Network Approaches to Model Hourly Energy Use in Commercial Buildings,"
Ph.D. Dissertation, Mechanical Engineering Department, Texas A&M University,
8 Acknowledgments College Station, Texas, May.
We gratefully acknowledgefinancialsupport from the Texas 8 Hittle, D. C , and Pedersen, C, O., 1981, "Periodic and Stochastic Behav-
ior of Weather Data," ASHRAE Transactions, Vol. 87, Part 2, pp. 173-194.
State Energy Conservation Office of the Intergovernmental Di- 9 Katipamula, S., Reddy, T. A., and Claridge, D. E., 1994, "Development
vision of the General Services Commission (State Agencies and Application of Regression Models to Predict Cooling Energy Consumption
Program) as part of the LoanSTAR Monitoring and Analysis in Large Commercial Buildings," Proceedings of the 1994 ASME/JSME/JSES
Program. Discussions with Dr. Srinivas Katipamula, Dr. Kelly International Solar Energy Conference, San Francisco, Mar. 27-30, pp. 307-
Kissock, Dr. Jeff Haberl and Dr. Sabratnam Thamilseran have 322.
10 Katipamula, S., Reddy, T. A., and Claridge, D. E., 1995, "Effect of Time
been extremely useful. Resolution of Statistical Modeling of Cooling Energy Use in Large Commercial
Buildings," ASHRAE Transactions-Research, Vol. 101, Pt. 2.
11 Kissock, J. K., 1993, "A Methodology to Measure Retrofit Energy Savings
References in Commercial Buildings," Ph.D. Dissertation, Mechanical Engineering Depart-
1 Chui, C. K., 1992, An Introduction to Wavelets, Academic Press Inc., San ment, Texas A&M University, College Station, Texas, Dec.
Diego, California. 12 Kreider, J. F., and Haberl, J. S., 1994, "Predicting Hourly Building Energy
2 Claridge, D., Haberi, J., Turner, W., O'Neal, D., Heffington, W., Tomban, Usage: The Great Predictor Shootout—Overview and Discussion of Results,"
C , and Jeager, S., 1991, "Improving Energy Conservation Retrofits with Mea- ASHRAE Transactions, Vol. 100, Pt. 2, pp. 1104-1118.
sured Savings," ASHRAE Journal, Oct., pp. 14-22. 13 Ott, L., 1988, An Introduction to Statistical Methods and Data Analysis,
3 Dhar, A., Reddy, T. A., and Claridge, D., 1994, "Improved Fourier Series PWS-KENT Publishing Company, Boston.
Approach to Modeling Hourly Energy Use in Commercial Buildings," ASME/ 14 Pandit, S. M., and Wu, S. M., 1983, Time Series and System Analysis with
JSME/JSES International Solar Energy Conference, San Francisco, Mar. 27-30, Applications, John Wiley and Sons, New York.
pp. 455-468. 15 Philips, W. F., 1984, "Harmonic Analysis of Climatic Data," ASME
4 Dhar, A., Reddy, T. A., and Claridge, D., 1995a, "Generalization of the JOURNAL OF SOLAR ENERGY, Vol. 32, No. 3, pp. 319-328.
Fourier Series Approach to Model Hourly Energy Use in Commercial Buildings," 16 Seem, J. E., and Braun, J. E., 1991, "Adaptive Methods for Real-time
Accepted for publication in ASME JOURNAL OF SOLAR ENERGY ENGINEERING . Forecasting of Building Electrical Demand,'' ASHRAE Transactions, Vol. 97, Pt.
5 Dhar, A., Reddy, T. A., and Claridge, D., 1995b, " A Fourier Series Ap- 1, pp. 710-721.
proach to Predict Hourly Heating and Cooling Energy Use in Commercial Build- 17 Tolstov, G. P., 1962, Fourier Series, Translated by Richard A. Silverman,
ings with Outdoor Temperature as the Only Weather Variable," Proceedings of Dover Publications Inc., New York.

Journal of Solar Energy Engineering AUGUST 1998, Vol. 120 / 223

Downloaded From: http://solarenergyengineering.asmedigitalcollection.asme.org/ on 11/29/2017 Terms of Use: http://www.asme.org/about-asme/terms-of-use