Applied Energy 136 (2014) 186–196

Contents lists available at ScienceDirect

Applied Energy
journal homepage: www.elsevier.com/locate/apenergy

Modeling and forecasting of cooling and electricity load demand

A. Vaghefi a,⇑, M.A. Jafari a, Emmanuel Bisse b, Y. Lu b, J. Brouwer c
a
Department of Industrial and System Engineering, Rutgers University, Piscataway, NJ 08904, USA
b
Siemens Corporate Research, Princeton, NJ 08540, USA
c
National Fuel Cell Research Center, University of California, Irvine, Irvine, CA 92697, USA

h i g h l i g h t s

 We propose a model for forecasting cooling and electricity load demand.

 The model takes the advantage of both time series and regression methods.
 The model is able to accurately forecast the load demands of the CCHP system.

a r t i c l e i n f o a b s t r a c t

Article history: The objective of this paper is to extend a statistical approach to effectively provide look-ahead forecasts
Received 25 January 2014 for cooling and electricity demand load. Our proposed model is a generalized form of a Cochrane–Orcutt
Received in revised form 14 August 2014 estimation technique that combines a multiple linear regression model and a seasonal autoregressive
Accepted 2 September 2014
moving average model. The proposed model is adaptive so that it updates forecast values every time that
new information on cooling and electricity load is received. Therefore, the model can simultaneously take
advantage of two statistical methods, time series, and linear regression in an adaptive way. The effective-
Keywords:
ness of the proposed forecast model is shown through a use case. The example utilizes the proposed
Load forecasting
Time series-regression model
approach for economic dispatching of a combined cooling, heating and power (CCHP) plant at the Univer-
Cochrane–Orcutt estimation sity of California, Irvine. The results reveal the effectiveness of the proposed forecast model.
Thermal energy storage Ó 2014 Elsevier Ltd. All rights reserved.
Heat recovery steam generator

1. Introduction A wide range of optimal control strategies has been proposed to

improve the CCHP operation based on different objectives includ-
U.S. households and commercial buildings consume approxi- ing power flow, capacity, operation, energy-use and environmental
mately 40% of total energy use and account for 72% of total U.S. considerations [5–12]. A common element in almost all optimal
electricity consumption [1]. Commercial building energy demand, control strategies is to have an accurate estimation of cooling,
in particular, doubled between 1980 and 2000 and is predicted heating, and electricity load demands. Some researchers assume
to increase 50% over the next 15 years [2]. As a result, energy that load demands are known and available over a specific period
demand management has emerged as a key policy for both public [8,9]. However, cooling and electricity demands are typically sto-
and private organizations. CCHP systems can significantly contrib- chastic and unknown mainly because of the complex interaction
ute to reducing buildings energy use, curtail pollutant and carbon s between plant facilities and equipment, e.g. chillers and turbines
emission, and help to decrease risks of blackouts and brownouts in yields. Liu et al. [8] point out that in practical applications, the
the utility grid [3,4]. CCHP technology integrates processes of exact future load profile does not exist; and forecasting methods
production and simultaneous use of cooling, heating, and power should be taken into consideration by researchers. Therefore, a
at a single site. However, since most commercial and industrial forecasting mechanism should be applied by researchers to find
electrical loads are highly dynamic and typically not synchronized the future values of load demands.
with local heating and cooling demands, advanced control strate- A number of researchers employ building simulation platform
gies will be imperative to economic dispatch of CCHP resources. to generate building load demand based on its physical character-
istics and other dynamic input variables such as occupancy,
⇑ Corresponding author. weather, and time information. The cooling and electricity load
E-mail addresses: vaghefi@rutgers.edu (A. Vaghefi), [email protected] (M.A. demands are outputs of running the simulation and are then fed
Jafari), [email protected] (E. Bisse), [email protected] (Y. Lu), into the optimization model [10–12]. However, the quality of
[email protected] (J. Brouwer).

http://dx.doi.org/10.1016/j.apenergy.2014.09.004
0306-2619/Ó 2014 Elsevier Ltd. All rights reserved.
 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196 187

Nomenclature

CCHP combined cooling, heating and power Symbols

LSE least square estimate W electricity power
ARX autoregressive with exogenous variable Q cooling power
ARMAX autoregressive moving average with exogenous variable m_ mass flow
ARMA autoregressive moving average T temperature
AI artificial intelligence e random error term
ANN artificial neural network y dependent/output variable
R2 coefficient of determination x independent/input variable
R2adj adjusted coefficient of determination b coefficients of input
GT gas turbine U autoregressive operator
ST steam turbine H moving average operator
TES thermal energy storage B backward operator
HRSG heat recovery steam generator
COP coefficient of performance Subscripts
WkCHC cooling power generated by the kth chiller (kW) T time (h)
WkCHW power consumed by the kth chiller to cooling power J index for input variables
(kW) CHC cooling generated by chiller
COPk coefficient of performance for the kth chiller CHW power consumed by chiller
WCHW total power consumed by chillers to generate total cool- W water
ing power of the campus (kW) Chw chilled water
Qcooling cooling demand from the campus (kW) CHRw water returned to chiller
TCHRw returned water temperature to chillers (K) CHSw water supplied by chiller
TCHSw supply water temperature from chillers (K) GT gas turbine
m_ chw chilled water mass flow rate (kg/s) ST steam turbine
cw specific heat capacity of water (kJ/kg K) grid power grid
Welectricity electricity demand from the campus (kW)
Wgrid power purchased from grid (kW) Superscripts
WGT power produced by gas turbine (kW) k chiller number
WST power produced by steam turbine (kW)

results highly depends on quality of the simulation models and CCHP components, chiller types and generator nominal capacities
their inputs. In addition, for any CCHP optimization, a detailed are examples of such factors; (ii) Environmental variables extrinsic
building simulation model needs to be accordingly built and run to the building, such as climate and weather data; (iii) operational
repeatedly. Another way to deal with this problem is to consider variables, e.g. cooling/heating set point values, lighting, time sche-
uncertainty in CCHP optimization model. Hu and Cho [12] for dule to operate various equipment and system components within
instance, propose an optimization model with some probabilistic plant or building; and (iv) uncontrollable dynamical variables, such
constraints to guarantee that the model is reliable to satisfy the as number of occupants at any time, noise due to structural varia-
stochastic load demand. They assume load demands are indepen- tions etc. It is ideal to know all these factors and their impacts on
dent and follow normal distributions in which 95% of the area is energy dynamics in order to optimally forecast and control cooling
within the range of ±20% of the average load demands. Another and electricity demands for single building or a cluster of buildings.
approach to this problem is to develop a forecasting model and However, a complete forecast model is not practically attainable
embed it into the optimization model. This is the main motivation due to unknown significant dynamical variables, lack of tools to
of this work. In this paper, Cochrane–Orcutt estimation technique measure their effects, or that some of these variables are uncon-
is used as an effective linear model to provide look-ahead forecasts trollable. Therefore, a wide range of different methods has been
for cooling and electricity demand load. It simultaneously fits a proposed to model and forecast load dynamics. In overall, these
regression model and a time series to the data while maintaining methods can be categorized into three general approaches.
least square estimate (LSE) conditions. In addition, the forecast val- In the first approach, a linear or nonlinear statistical model is
ues are modified when a new data is received from the real system. used to explain the variability of response (load or energy dynam-
The proposed model is currently working as a part of an integrated ics) over time. The most popular example of such statistical models
optimal dispatch for CCHP plant at the University of California, is Box and Jenkins time series paradigm where load demands are
Irvine and providing accurate forecasts for the entire campus estimated based upon a linear combination of their past values
cooling and electricity load demand. [13,14]. There are a large family of different models in this category
that can deal with many special cases including seasonality, non-
2. Background study stationary, and non-homogeneity of variances (see e.g. [15,16]).
The major drawback of such models is that the future values are
In most real cases, cooling and electricity load demands are typically forecasted based upon the past and present values of
highly dynamic oscillating within a wide range of values during cooling and electricity load demands without considering any
course of a day. This is mainly because several physically explicit exogenous factors in the model. Another example of statistical
or latent factors can instantaneously influence cooling and electric- approach is using regression models (metamodel) where the vari-
ity demand patterns. These factors can be any one of the following ability within response is modeled via a number of exogenous fac-
types: (i) Static factors that are usually set at the design stage and tors [17–21]. The major problem of such models is that they often
only change due to aging wear and tear. Building characteristics, ignore the complex interactions between exogenous factors, which
188 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196

may result in less accurate forecast values. To overcome this prob- 3. Time series regression model
lem, a number of studies use a hybrid approach, which employs the
main components of both above-mentioned approaches [22]. ARX The common assumption of uncorrelated random error terms
and ARMAX are two examples of this approach. Although these (e’s) made in basic regression models is not appropriate to forecast
models perform effectively in many cases, they have many param- building energy consumption. Historical data shows that error
eters to be estimated since all input and output variables with their terms are frequently correlated (often positively) over time [33].
past and current values should appear in the forecast model. In particular, this typically happens when there are some uncon-
The second approach employs artificial intelligence to find the trollable, unknown, or non-measurable input variables. A special
k-step ahead forecasts for load demand. A broad range of numeri- case for the regression model with auto-correlated data can be
cal methods can be included in this category. Kalogirou [23] and shown as follows:
Mellit and Kalogirou [24] provide a comprehensive review of AI
X
k
techniques in some areas of energy. Although their techniques yt ¼ bj xtj þ et ; et ¼ nðet1 ; . . . ; etq Þ þ at ; ð1Þ
are not directly related to load forecasting, however, they can eas- j¼0
ily be used with minor changes. Artificial neural network (ANN) is
among most frequent AI techniques and has been widely used in where n(.) is a function of previous error terms e’s, yt is the power con-
load or energy forecasting. ANN’s have particularly evolved based sumed at time t and xtj is the jth input variable affecting the building
upon different settings of neuron arrangement, neuron connec- energy consumption at time t and at is a white noise. The error terms
tions, training techniques, and internal layers and become a pow- are typically modeled using Box and Jenkins model as a first order
erful competitor for statistical methods [25–29]. They can be auto-regressive model. A preliminary study of our historical data
designed to include both past observation of cooling and electricity on cooling and electricity load demands indicates a seasonal pattern
demands and associated exogenous factors. The main disadvantage with lag of 24 h. Therefore, the error terms in (1) is generalized to
of AI approach is that they are often black box and do not show any include seasonal patterns. To do this, assume that p, q, P and Q are
explicit relationship between response an input variables. For the order of non-seasonal and seasonal autoregressive and moving
example, the hidden layers of ANN’s are difficult to explain and range parts respectively, and s is the seasonal order. Then a general
cannot be appeared in an explicit forecasting equation [30]. ARMA model for error terms can be written as follows:
In addition, by developing computational methods, a third /p ðBÞUsP ðBÞet ¼ hq ðBÞHsQ ðBÞat ; ð2Þ
approach has recently been developed which is a combination of
s s
any abovementioned techniques. The main purpose of this hybrid where /p and U are autoregressive operators, hq and H are mov-
P Q
approach is to improve the accuracy of the forecast values by com- ing average operators and B is backward operator. s is set equal to
bining different numerical–analytical methods. Some hybrid meth- 24 showing the significance of autocorrelation between loads of
ods also partially include the physical aspects of the real system in same time in two consecutive days.
their computation and come up with a mixed physical–numerical Let
method, which is often referred to as gray models [30]. A few appli- /p ðBÞUsP ðBÞ ¼ 1  WðBÞ
cations of hybrid models in the area of energy can be found in
[31,32,29]. then
The proposed model can be classified in the statistical groups. It
et ¼ WðBÞet þ hq ðBÞHsQ ðBÞat ;
first fits a linear regression to find the correlation between the
cooling and electricity load demands and exogenous factors. Any Furthermore, Eq. (2) can be written as follows:
variability that cannot be explained by regression models can be
X
k X
p X
P
aggregated in residual terms. Then, a seasonal time series model yt ¼ bj xtj þ et ð1Þiþj1 ui /j Biþsj
is applied to the residuals to express the remaining variability. j¼0 i¼0 j¼0
Since, the regression parameters should be estimated using least
X
q X
Q
square error method, the process of parameters estimation is þ at ð1Þiþj hi Hj Biþsj ; ð3Þ
applied iteratively and simultaneously. Further details will be i¼0 j¼0
explained in the next section.

Fig. 1. Schematic framework of CCHP plant at University of California, Irvine.

 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196 189

Note that u0 ¼ /0 ¼ 0 and h0 ¼ H0 ¼ 1. For example, for the ARMA determination R2adj are employed as measures for model adequacy
(1,0)  (1,0)n=24, Eq. (3) is written as follows: checking. These measures can be calculated as follows:
X
k ^b0T X0T ðI  HÞX0 ^b0
yt ¼ bj xtj þ et ðu1 B þ /1 B24  u1  /1 B25 Þ R2 ¼ 2 2
; ð6Þ
j¼0
y0T
2 ðI  ð1=nÞJÞy 2
0

X
¼ bj xtj þ u1 et1 þ /1 et24  u1  /1 et25 ; ð4Þ and
j
^b0T X0T ðI  HÞX0 ^b0 =k  1
The main significance of Eq. (3) is that it includes seasonal error R2adj ¼ 2 2
; ð7Þ
y0T 0
2 ðI  ð1=nÞJÞy2 =n2  k
and tends to capture statistical similarities between two periods,
which are n hours apart. The major problem of multiple linear where k is number of exogenous variables, n2 is sample size for test-
regression with auto-correlated error terms is the estimation of ing dataset, I is identity matrix and H can be calculated by
1
coefficients. With auto-correlated error terms, the ordinary least H ¼ X2 ðXT2 X2 Þ XT2 as well. R2 and R2adj are both between 0 and 1
square (OLS) procedures can be misleading and does not guarantee and explain the percentage of variation that is explained by model.
estimation with the minimum variance [33]. To overcome this A closer value to 1 depicts a better model.
problem, Cochrane and Orcutt [34] proposed a transformation
when error terms follow a first order autoregressive process. 4. Case study and experimentation
According to Cochrane–Orcutt model, one should transform the
response values in such a way that In this section, the forecast model is employed as a part of opti-
Y 0t s 0 s
b00 s mal dispatching of a CCHP plant at the University of California,
¼ /p ðBÞU P ðBÞY t ; xt ¼ /p ðBÞU P ðBÞxt and ¼ /p ðBÞU P ðBÞb0 :
Irvine. Cooling and electricity forecast values are fed into an opti-
Therefore, Eq. (3) can be replaced by mal control strategy, which searches for optimal set points for
24 h ahead. The forecast model then is used to compute optimal
Y 0t ¼ b00 þ x0t b0t þ at ; ð5Þ
control values to minimize energy consumption during course of
Eq. (5) is an ordinal multiple linear regressions with independent a day in a building.
error terms and can be calculated via OLS estimation method. As The UC Irvine Central Plant consists of eight electric chillers,
b0 ¼ b
a result, the fitted linear function Y ^0 þ x0 ^0
t 0 t bt can eliminate the providing cold water, a 13.5 MW gas turbine (GT), a 5.7 MW steam
autocorrelation structure of the error terms. The following turbine (ST), thermal energy storage (TES) tank, and a heat recov-
algorithm summarizes our approach: ery steam generator (HRSG). It provides heating and cooling loads
for the entire campus as well as the majority of the campus electric
3.1. Algorithm loads. The chillers are able to supply as much as 14,500 tons
(51 MW) and the steam driven chiller can provide an additional
2000 tons (7 MW). The TES tank capacity is 60,000 ton-hour (211
{
Step 1: Divide the original dataset into two subsets:
training dataset and testing dataset, which are used for 12 500
model estimation and model verification respectively and
denoted by X1 and X2. Set i = 0. 12 000
Cooling demand (kW)

Step 2: Fit a multiple regression model to training subset 11 500

and estimate vector of b^i in y0i ¼ X1 b
^i , where y1 ; X1 2 X1 are
1
11 000
response (cooling or electricity load demand) and
independent variables (exogenous variables). Then 10 500
^i ¼ y1  y0i1 ¼ y1  X1 ^
calculate initial residual values by e bi . 10 000
^i ’s are correlated fit an ARMA model,
Step 3: If e
^ si ^ si ðBÞat , and find estimation values for 9 500
/ ðBÞU ðBÞei ¼ ^
^
p
i
P hi ðBÞH
q Q
^ si ðBÞ and H
^i ; U
/ ^ si ðBÞ using least square error technique or 9 000
p P Q S un Mon T ue W ed T hu Fri S at
other estimators. W ee kdays
i
Step 4: Apply following transformations y0i1 ¼ /ip ðBÞUsP ðBÞy1
b) Inter val Plot of Electr icity Demand
i
and X0i1 ¼ /ip ðBÞUsP ðBÞX1 on y1 ; X1 2 X1 . Then fit a new 14000

multiple regression model to transformed subset and

Electricity Demand (kW)

estimate vector of ^ b0i in y0i1 ¼ X0i1 ^

b0i . 13500
 0i 
Step 5: Check ^ b ^ b  < d. If the criterion is met then set
0i1
13000
^0 ¼ b
b ^0i and go to Step 6. Otherwise, calculate the residual
values by e^i ¼ y1  y0i1 ¼ y1  X0i1 ^
b0i
and go to Step 3.
12500
Step 6: Apply anti-transformations for b00 ¼ /p ðBÞUsP ðBÞb0
and ^
b¼^ b0 and use them in Eq. (1). 12000
}
11500
Sun Mon Tue W ed Thu Fri Sat
Weekdays
It is quite common to use the estimated parameters as well as
subset X2 to check the adequacy of the given model. In this study, Fig. 2. 95% Confidence interval plots categorized by weekday for (a) cooling load
coefficient of determination R2 and adjusted coefficient of demand and (b) electricity load demand (Day 1 is Sunday, Day 7 is Saturday).
190 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196

electrical needs on the campus with the balance being served by

20000
utility import (14%) and an 893 kW-fixed panel solar photovoltaic
(1%).
Cooling load demand (kW)

17500
As mentioned, the electricity produced by two generators are
either sent directly to the campus to satisfy electricity demand or
15000
supplied as the energy input to the electrical chiller (see [9] for
more details), which is mainly responsible to provide cold water.
12500
Cold water can be either directly supplied to the campus to meet
10000
campus cooling needs or stored in the TES tank for later use. Hence,
the chillers and the TES together are the main sources for the
7500 campus cooling demands. Any additional electricity demand is
provided from the grid.
5000 Such a CCHP system is able to produce thermal energy along
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 with electricity over time. The thermal energy storage (TES) is a
hour flexible component of the plant, which allows the campus to
reshape the cooling demand particularly in peak hours. There are
many examples of CCHP supervisory control systems in literature
14500 [5,6,10]. A key element for such optimal control is to have accurate
information about the power (electricity and cooling) demand over
Electricity Demand (kW)

14000
the course of a day, which is the central focus of this study.
13500 Suppose that W kCHC is the cooling load generated by the kth chil-
ler (kW), and that W kCHW is the power consumed by the kth chiller
13000 (kW) to generate W kCHC units of cooling load. Then W kCHC is propor-
tional with W kCHW as follows:
12500
W kCHW ¼ wkCHC =COPk ð8Þ
12000
k
where COP is the coefficient of performance for the kth chiller
11500 which is the ratio between efficient energy acquired by and sup-
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
plied to the chiller; this is typically determined by the chiller man-
hour ufacturer. In this study, COPk is fixed and given by the chillers’
manufacturer. However, in reality, it is a function of the real oper-
Fig. 3. 95% Confidence interval plots categorized by hours for (a) cooling load ating temperature and reliability of the absorption chiller. This
demand (kW) and (b) electricity load demand (kW). information is not often available. Therefore, any variation due to
change in COPk is appeared in error term of (1) and should be mod-
eled via time series part of the proposed model.
megawatt-hour) which is able to shift, on average, 65% of the W kCHW presents the actual power (electricity) consumed by the
cooling load during the day to the night when electricity prices kth chiller to produce W kCHC . The total power consumed by all
are lower and temperature is cooler. chillers is given by:
Fig. 1 provides a schematic of the plant, where GT is the primary
source of electric power providing electricity for the campus and X
8

for the chillers. As a byproduct, the gas turbine generates the W CHW ¼ W kCHW : ð9Þ
k¼1
exhaust gas, which can be source of extra thermal energy. Such
energy is then used to produce steam using HRSG unit. HRSG can Note that W kCHC values do not reflect the cooling power supplied to
supply 23,500 kg/hour and 54,000 kg/hour without and with duct the campus. A portion of cooling load produced by the chillers is
fire, respectively. The generated steam drives the steam turbine sent to the TES tank and stored for peak hours. Thus, W kCHC values
(ST). The steam can also be used to produce hot water for the cam- cannot be a good measure for determining the total cooling demand
pus needs. A portion of the produced steam is also transferred to of campus at any time. Instead, the amount of cooling supplied to
use in a steam chiller unit. GT and ST supply about 85% of the total the campus can be expressed as follows:

50000 20000
Cooling load demand (KW)

18000
Electricity Demand (KW)

40000

16000
30000

14000
20000
12000

10000
10000

0
40 50 60 70 80 90 100 40 50 60 70 80 90 100
Sit e t emper at ur e (F) Site temperature (F)

Fig. 4. Scatter plots of cooling and electricity load demand versus site temperature.
 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196 191

a
50000

Cooling Demand (kW)

40000
30000
20000
10000
0
0 2500 5000 7500
Time (hour)

b
Residuals (kW)

10000

-10000

0 2500 5000 7500

Time (hour)

c
1.00

0.75
ACF

0.50

0.25

0.00
0 10 20 30 40
lag

d
0.9

0.6
PACF

0.3

0.0

-0.3
0 10 20 30 40
lag

Fig. 5. (a) Time series plot for the cooling load demand, (b) the residuals for a preliminary linear model, (c) autocorrelation plot and (d) partial autocorrelation plot for
residual values.

_ chw  cw  ðT CHRw  T CHSw Þ;

Q cooling ¼ m ð10Þ for a relatively negligible portion of the power consumption
throughout the campus. The proposed forecast model is used to
where Qcooling is the total amount of cooling (kW) provided by the forecast both Qcooling and Welectricity using a set of weather and time
chillers and supplied to the campus to meet cooling demands, variables as well as historical cooling and electricity data.
T CHRw is the temperature of returned water to chillers (K), T CHSw is
the supply water temperature from chillers (K), m _ chw is the chilled
water mass flow rate (kg/s) and cw is the specific heat capacity of 5. Results for the CCHP plant data
water (kJ/kg K) [9]. All above parameters are known and available
in the plant. This allows us to accurately estimate the actual cooling In this section, the performance of the proposed method is dis-
load demands. cussed using the CCHP plant data collected from the UCI campus.
Similar to the cooling load, the direct values for the electricity In this example, one year (September 2009–2010) and 4 months
load demand are not available. However, this can be calculated data (September 2009–December 2009) are used for building the
from the hourly power consumption by the chillers, the total forecast models for the cooling and electricity load demands,
power generated by gas and steam turbines, and the power pro- respectively. Both datasets are provided by the UCI campus plant
vided by grid. The electricity load at time t is therefore: based on actual values of the cooling and electricity consumption.
W electricity ¼ W GT þ W ST þ W grid  W CHW ; ð11Þ Each dataset is divided into two subsets. The first set is used for
model building and estimation purposes (training dataset). The
where WGT and WST are the power produced by gas and steam tur- rest of the data is used for validation purposes (testing dataset).
bines, respectively, and Wgrid is the power purchased from grid at In this work, Matlab is employed for creating and testing the pro-
any time. WCHW is the total power consumed by all chillers, which posed forecast model and plotting and visualization is done by
is calculated in (9), and Welectricity is the electricity load demand at Minitab and R. In this phase, Eqs. (6) and (7) are used to investigate
time t. In this study, due to lack of data, we ignore the power the performance of the forecast models. The testing subset does
consumption by pumps and chiller compressors, which account not share any information with the training dataset.
192 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196

Electricity Demand (kW)

a
18000

16000

14000

12000

10000

0 1000 2000 3000

Time (hour)

b
Residuals (kW)

2500

-2500

-5000

0 1000 2000 3000

Time (hour)

c
1.00

0.75
ACF

0.50

0.25

0.00
0 10 20 30
lag

d
1.0

0.5
PACF

0.0

0 10 20 30
lag

Fig. 6. (a) Time series plot for electricity load demand, (b) the residuals for a preliminary linear model, (c) autocorrelation plot and (d) partial autocorrelation plot.

4
x 10
5
Cooling Load (kw)

aa
4

0
0 500 1000 1500 2000 2500 3000 3500 4000
Time
4
x 10
5
Cooling Load (kw)

bb
4

0
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
Time

Fig. 7. Comparison of actual and forecasted values for cooling load demand using (a) training dataset (above) and (b) testing dataset (below).

Before building the forecast model, an exploratory data analysis both the cooling and the electricity load demands are higher in
is performed to capture the behavior of data over time. Fig. 2 working days than weekends. This is particularly obvious for the
depicts the 95% confidence interval plots for the cooling and elec- electricity load demand that is less than 12,000 (kW) in weekends
tricity load demands categorized by weekdays. It is observed that and more than 13,000 (kW) for weekdays. This implies that mixing
 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196 193

Table 1
The estimates values for cooling and electricity forecast models.

Cooling Electricity
Weekdays Weekends
Estimate Standard error Estimate Standard error Estimate Standard error
b0 13441.85 1851.10 12,783 458.2 9825.557 704.74
b1 357.28 13.98 21.0180 6.1137 39.168 8.809
u1 0.9059 0.016 0.8775 0.0319 1.1882 0.0750
u2 0.0513 0.021 0.0018 0.0425 0.1848 0.1162
u3 0.0830 0.02134 0.1298 0.0423 0.0140 0.1168
u4 0.0314 0.02138 0.0606 0.0425 0.0594 0.1168
u5 0.0299 0.02138 0.1054 0.0425 0.2021 0.1169
u6 0.0405 0.02137 0.0073 0.0425 0.1319 0.1178
u7 0.0829 0.02138 0.1304 0.0422 0.1240 0.1160
u8 0.0264 0.02142 0.1095 0.0423 0.1624 0.1170
u9 0.0289 0.02142 0.0625 0.0425 0.1140 0.1190
u10 0.0094 0.02140 0.0484 0.0425 0.1592 0.1195
u11 0.0191 0.02136 0.0155 0.0426 0.1565 0.1210
u12 0.0017 0.02135 0.0777 0.0427 0.0218 0.1219
u13 0.0138 0.02135 0.0466 0.0427 0.1390 0.1204
u14 0.0112 0.02134 0.0047 0.0427 0.1450 0.1190
u15 0.0275 0.02134 0.0287 0.0427 0.1449 0.1189
u16 0.0021 0.02134 0.0276 0.0426 0.0813 0.1194
u17 0.0012 0.02134 0.0957 0.0425 0.2598 0.1193
u18 0.0074 0.02129 0.1241 0.0424 0.2818 0.1206
u19 0.0084 0.02128 0.0501 0.0426 0.0009 0.1222
u20 0.0191 0.02128 0.0550 0.0425 0.0622 0.1233
u21 0.0392 0.02128 0.0531 0.0425 0.0664 0.1228
u22 0.0728 0.02124 0.0966 0.0423 0.1109 0.1241
u23 0.0751 0.02125 0.0439 0.0424 0.0776 0.1239
u24 0.0322 0.01576 0.0797 0.0319 0.0285 0.0814
R2 0.884 0.708 0.430
R2adj 0.883 0.700 0.405

all data and building a global forecast model without considering related to such missing exogenous factor(s) should be explained
the factor of ‘‘day’’ may result in a less powerful model. Thus, in and modeled by time series part of the proposed method.
this work, two different models are constructed for weekdays Fig. 5a and b present the hourly cooling load of the campus and
and weekends. the residual values given by fitting a linear model of cooling versus
Fig. 3 presents the 95% confidence interval plots for the cooling ambient temperature. The residuals are highly autocorrelated over
and the electricity load demands categorized by 24 h of the day. time in different lags (Fig. 5c). Furthermore, Fig. 5d is the partial
For example, 17 in x-axis means the 95% confidence interval for autocorrelation function (PCAF) for residual values and can identify
the cooling and electricity load demands at time 17:00, which is the extent of lags in an autocorrelation model. In this figure, PACF
constructed by all data collected at this particular time slot. This illustrates a strong autocorrelation structure in the first lag and the
figure can easily represent peak time for the cooling and electricity 24th lag, which accounts for seasonality in the data. Therefore, a
load demands. seasonal ARMA(2,0,0)  (1,0,0)24 seems an appropriate candidate
For cooling, the load demand increases constantly from 6:00 for the electricity load dataset.
and reaches its maximum value at time14:00 then decreases until Similarly, Fig. 6a and b are the electricity load demand and its
end of the day. The peak hours for the cooling load demand are corresponding residual values when applying a linear model to
between 11:00 and 17:00. This also implies that the cooling the data. Again, ACF and PACF in Fig. 6c and d reveal a correlated
demand load is highly correlated with the ambient temperature. structure for the electricity load dataset. Particularly, PACF illus-
Similarly, the peak hours for electricity load demand are between trates a positive autocorrelation for the first lag and a remarkable
9:00 and 19:00 as well. negative correlation for the 24th lag. This means that a seasonal
Fig. 4 shows scatter plots of the cooling and electricity load ARMA(1,0,0)  (1,0,0)24 model would be enough for the electricity
demands versus the ambient temperature. The cooling load values load demand.
show higher correlation with ambient temperature than the elec- Fig. 7a and b depict the result of forecast modeling for the cool-
tricity load demand. The estimated correlations between cooling ing load demand using training and testing datasets. In Fig. 7a, the
and electricity load demands with ambient temperature are forecast values are very close to the corresponding actual values.
0.905 and 0.374, respectively. This means that to find an accurate This is because the training dataset is used for parameter estima-
model for the electricity load demand, it is required to add more tion of the forecast model. Therefore, the model includes the infor-
significant exogenous factors than ambient temperature. mation of actual data.
For example, the average number of people in the campus at Fig. 7b represents the performance of the model with testing
time t would be a potential exogenous factor for modeling the dataset, which does not share any information with the estimated
campus electricity load demand. As number of people in the cam- parameters. It is observed that the model adequately fits with the
pus increases, it is logical to presume that the electricity load actual data. In addition, Table 1 provides the estimate values of the
demand increases. However, in this example, since the number model parameter, their standard errors as well as coefficient of
of people in the campus at time t is not available, it is not possible determinations for both cooling and electricity load demands. For
to analyze its effect. As a result, those parts of variation that are the cooling demand, coefficient of determination R2 and adjusted
194 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196

4
x 10
1.8

Electricity Load (kw)

1.6

1.4

1.2

1
0 100 200 300 400 500 600 700 800 900 1000
Time
4
x 10
1.8
Electricity Load (KW)

1.6

1.4

1.2

0.8
0 100 200 300 400 500 600 700 800 900 1000
Time

Fig. 8. Comparison of actual and forecasted values for electricity load demand in weekdays using training dataset (above) and testing dataset (below).

coefficient of determination R2adj are 88.4% and 88.3%, respectively for electricity load demand in weekdays are namely 70.8% and
implying that the proposed model can explain more than 88% of 70% and for electricity load demand in weekends are namely 43%
the total variability within data. and 40%. Therefore, the electricity load model should be enhanced
Figs. 8 and 9 present the actual and forecast values of electricity by adding more exogenous factors e.g. occupancy into the forecast
load demand using both training and testing datasets for weekdays model in order to capture larger amount of variability over time.
and weekends, respectively. As shown in Fig. 2b, the electricity Another potential reason for lower performance of the electric-
demand patterns are significantly different in weekends and week- ity demand forecast model is shown in Fig. 10. In this figure, the
days, probably because of fewer numbers of people in the campus values of electricity load demand are plotted over time and are
in weekends. Therefore, to improve the performance of the pro- grouped by months. It is shown that the load demand in the last
posed method, two separate models should be built for weekdays month follows different pattern than the other months. This is
and weekends. because the last month is December and the campus is probably
In addition, It is observed from Figs. 8 and 9 that the perfor- less populated at the last days of December. Since, the model uses
mance of the proposed model for the electricity load demand is the first two months for training and estimation and the rest of
still less than the same model proposed for the cooling load data (including December data) for the testing purposes, it cannot
demand. This is mainly due to lack of other exogenous factors in fit the last part of December.
electricity demand model. As shown in Fig. 4, the correlation A solution for this problem is to add the occupancy as another
between electricity load demand and the ambient temperature is exogenous variable into the model and re-estimate the model
moderate. It means that the ambient temperature can only explain parameters accordingly. This way, the model can differentiate
a relatively small portion of variation in electricity demand. This between those days that more people are in campus from the days
can be confirmed by observing Table 1. In this table, R2 and R2adj that less people are in campus including weekends. Another idea is

4
x 10
1.5
Electricity Load (kw)

1.4

1.3

1.2

1.1

1
0 20 40 60 80 100 120 140 160 180 200
Time
4
x 10
1.6
Cooling Load (kw)

1.4

1.2

0.8
0 100 200 300 400 500 600
Time

Fig. 9. Comparison of actual and forecasted values for electricity load demand in weekends using training dataset (above) and testing dataset (below).
 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196 195

18000 Month
9

Electricity Demand (kW)

10
11
16000 12

14000

12000

10000

1 293 586 879 1172 1465 1758 2051 2344 2637

Time (hour)

Fig. 10. Time series plot for electricity load demand grouped by month.

to build a new model solely for December. In doing so, the model [8] Liu Mingxi, Shi Yang, Fang Fang. Optimal power flow and PGU capacity of CCHP
systems using a matrix modeling approach. Appl Energy 2013;102:794–802.
switch to a new model that is designed and built based on
ISSN 0306-2619.
December data as soon as December begins. [9] Chandan V, Do A, Jin B, Jabbari F, Brouwer J, Akrotirianakis A, et al. Modeling
and optimization of a combined cooling, heating and power plant system,
accepted for publication in proceedings of the American Control Conference;
6. Conclusions 2012.
[10] Mago PJ, Chamra LM. Analysis and optimization of CCHP systems based on
energy, economical, and environmental considerations. Energy Build
In this paper, a statistical method for forecasting cooling and
2009;41(10):1099–106.
electricity load demands was proposed in both campus and building [11] Wang Jiang-Jiang, Jing You-Yin, Zhang Chun-Fa. Optimization of capacity and
level. The performance of the proposed model was evaluated using operation for CCHP system by genetic algorithm. Appl Energy
the CCHP plant data collected from the UCI campus. The results 2010;87(4):1325–35.
[12] Hu Mengqi, Cho Heejin. A probability constrained multi-objective
revealed that the proposed model was able to provide high quality optimization model for CCHP system operation decision support. Appl
forecasts for both cooling and electricity load demands. Coefficient Energy 2014;116(1):230–42.
of determination R2 and adjusted coefficient of determination R2adj [13] Huang Shyh-Jier, Shih Kuang-Rong. Short-term load forecasting via ARMA
model identification including non-Gaussian process considerations. IEEE
for forecasting cooling demand were 88.4% and 88.3%, implying that Trans Power Syst 2003;18(2):673–9.
the proposed model could explain more than 88% of the total vari- [14] Muche Thomas. Optimal operation and forecasting policy for pump storage
ability within testing data. These indices were 70.8% and 70% for plants in day-ahead markets. Appl Energy 2014;113(January):1089–99.
[15] Taylor JW, McSharry PE, Buizza R. Wind power density forecasting using
electricity load demand in weekdays and 43% and 40% in weekends ensemble predictions and time series models. IEEE Trans Energy Convers
respectively. The weekend cooling demand forecasts could signifi- 2009;24(3). p. 775, 782.
cantly improve by using more stable weekend data. [16] García-Martos Carolina, Rodríguez Julio, Sánchez María Jesús. Modelling and
forecasting fossil fuels, CO2 and electricity prices and their volatilities. Appl
The proposed model is now running in the campus and is
Energy 2013;101(January):363–75.
forecasting both cooling and electricity load demands as a part of [17] Goia Aldo, May Caterina, Fusai Gianluca. Functional clustering and linear
an integrated CCHP optimization platform. Further information regression for peak load forecasting. Int J Forecast 2010;26(4):700–11. ISSN
0169–2070.
from exogenous factors such as occupancy can improve the perfor-
[18] Serinaldi Francesco. Distributional modeling and short-term forecasting of
mance of the proposed model. electricity prices by generalized additive models for location, scale and shape.
Energy Econom 2011;33(6):1216–26.
[19] Jain Rishee K, Smith Kevin M, Culligan Patricia J, Taylor John E. Forecasting
Acknowledgment energy consumption of multi-family residential buildings using support vector
regression: Investigating the impact of temporal and spatial monitoring
This work was supported by U.S. Department of Energy and granularity on performance accuracy. Appl Energy 2014;123(15 June):168–78.
ISSN 0306-2619.
National Energy Technology Laboratory via grants.gov in response
[20] Apadula Francesco, Bassini Alessandra, Elli Alberto, Scapin Simone.
to Clean Distributed Generation: Fuel/Feedstock Flexibility Relationships between meteorological variables and monthly electricity
and Combined Heat and Power (Solicitation No. DE-PS26- demand. Appl Energy 2012;98(October):346–56. ISSN 0306-2619.
08NT0004312-03). [21] Mavromatidis Lazaros Elias, Bykalyuk Anna, Lequay Hervé. Development of
polynomial regression models for composite dynamic envelopes’ thermal
performance forecasting. Appl Energy 2013;104(April):379–91.
References [22] Inman Rich H, Pedro Hugo TC, Coimbra Carlos FM. Solar forecasting methods
for renewable energy integration. Prog Energy Combust Sci
2013;39(6):535–76. ISSN 0360-1285.
[1] U.S. Department of Energy. Energy efficiency and renewable energy. Energy
[23] Kalogirou Soteris A. Artificial intelligence for the modeling and control of
efficiency trends in residential and commercial buildings; October 2008.
combustion processes: a review. Prog Energy Combust Sci 2003;29(6):515–66.
[2] Energy Information Administration. Annual energy review 2004. Washington,
ISSN 0360-1285.
DC: U.S. Department of Energy, Annual Report 061–003-01141-1; 2005.
[24] Mellit Adel, Kalogirou Soteris A. Artificial intelligence techniques for
[3] Wu DW, Wang RZ. Combined cooling, heating and power: a review. Prog
photovoltaic applications: a review. Prog Energy Combust Sci
Energy Combust Sci 2006;32(5–6):459–95. ISSN 0360-1285.
2008;34(5):574–632.
[4] San Martín JI, Zamora I, San Martín JJ, Aperribay V, Eguía P. Trigeneration _
[25] Kankal Murat, Akpınar Adem, Kömürcü Murat Ihsan, Özsßahin Talat S ß ükrü.
systems with fuel cells. In: International conference on renewable energy and
Modeling and forecasting of Turkey’s energy consumption using socio-
power quality (ICREPQ); 2008.
economic and demographic variables. Appl Energy 2011;88(5):1927–39.
[5] Fang Fang, Wang QH, Shi Yang. A novel optimal operational strategy for the
[26] Kulkarni Santosh, Simon Sishaj P, Sundareswaran K. A spiking neural network
CCHP system based on two operating modes. IEEE Trans Power Syst
(SNN) forecast engine for short-term electrical load forecasting. Appl Soft
2012;27(2):1032–41.
Comput 2013;13(8).
[6] Liu Ye, Luo Yi. Optimal design of distributed CCHP system. In: International
[27] Kandil Nahi, Wamkeue René, Saad Maarouf, Georges Semaan. An efficient
conference on transportation, mechanical, and electrical engineering (TMEE),
approach for short term load forecasting using artificial neural networks. Int J
2011, 16–18 December; 2011. p. 1961–65.
Elect Power Energy Syst 2006;28(8):525–30.
[7] Cho Heejin, Mago Pedro J, Luck Rogelio, Chamra Louay M. Evaluation of CCHP
[28] Deihimi Ali, Orang Omid, Showkati Hemen. Short-term electric load and
systems performance based on operational cost, primary energy consumption,
temperature forecasting using wavelet echo state networks with neural
and carbon dioxide emission by utilizing an optimal operation scheme. Appl
reconstruction. Energy 2013;57(1):382–401.
Energy 2009;86(12):2540–9.
196 A. Vaghefi et al. / Applied Energy 136 (2014) 186–196

[29] Liu Hui, Tian Hong-qi, Pan Di-fu, Li Yan-fei. Forecasting models for wind speed [32] Lin Cheng-Ting, Chou Li-Der, Chen Yi-Ming, Tseng Li-Ming. A hybrid economic
using wavelet, wavelet packet, time series and Artificial Neural Networks. Appl indices based short-term load forecasting system. Int J Elect Power Energy Syst
Energy 2013;107(July):191–208. 2014;54(January):293–305. ISSN 0142-0615.
[30] Lee Yi-Shian, Tong Lee-Ing. Forecasting nonlinear time series of energy [33] Kutner M, Nachtsheim C, Neter J, Li W. Applied linear statistical models. Irwin,
consumption using hybrid dynamic model. Appl Energy 2012;94(June):251–6. Chicago, IL, US: McGraw-Hill; 2004.
[31] Li Kangji, Su Hongye. Forecasting building energy consumption with hybrid [34] Cochrane D, Orcutt GH. Application of least squares regression to relationships
genetic algorithm-hierarchical adaptive network-based fuzzy inference containing auto-correlated error. J Am Statistical Assoc 1949;44(245):32–61.
system. Energy Build 2010;42(11):2070–6.