Energy and Buildings 37 (2005) 1250–1259

www.elsevier.com/locate/enbuild

On-line building energy prediction using adaptive

artificial neural networks
Jin Yang a, Hugues Rivard b,*, Radu Zmeureanu a
a
Department of Building, Civil and Environmental Engineering, Centre for Building Studies Concordia University,
1455 Maisonneuve Blvd. West, Montreal, Canada H3G 1M8
b
Department of Construction Engineering, ETS, 1100 Notre-Dame Street West, Montreal, Canada H3C 1K3
Received 9 December 2004; received in revised form 28 January 2005; accepted 10 February 2005

Abstract

While most of the existing artificial neural networks (ANN) models for building energy prediction are static in nature, this paper evaluates
the performance of adaptive ANN models that are capable of adapting themselves to unexpected pattern changes in the incoming data, and
therefore can be used for the real-time on-line building energy prediction. Two adaptive ANN models are proposed and tested: accumulative
training and sliding window training. The computational experiments presented in the paper use both simulated (synthetic) data and measured
data. In the case of synthetic data, the accumulative training technique appears to have an almost equal performance with the sliding window
training approach, in terms of training time and accuracy. In the case of real measurements, the sliding window technique gives better results,
compared with the accumulative training, if the coefficient of variance is used as an indicator.
# 2005 Elsevier B.V. All rights reserved.

Keywords: On-line prediction; Electric demand; Energy demand prediction; Building cooling; Artificial neural networks; Adaptive models

1. Introduction
The prediction of building energy consumption can play
an important role in building management since it can help
optimize the building daily operation and select better
control strategies [1]. An automated energy prediction
system is often built on top of a mathematical prediction
model consisting of several parameters. The model
parameters are estimated using existing data that typically
include energy demand or consumption and temperature
measurements recorded in the past. A variety of prediction
models have been proposed in the literature that include
time-series models, Fourier series models, regression
models, artificial neural network (ANN) models and fuzzy
logic models. Each model type has its own features,
advantages and disadvantages, and in addition, its perfor-
mance varies from one application to another.
With the exception of a few ANN models, most of the
surveyed literature focus on static prediction, a prediction
scheme that involves a single prediction model that does not
evolve over time: when the estimation of the model
parameters is completed, the model is fixed; the most
recently collected data is not used to update the model
parameters. To obtain an accurate static model, a large
volume of historical data is required to estimate the model
parameters. The alternative approach is the use of a dynamic
(adaptive) model that constantly updates model parameters
based on newly available data. As the energy data collection
process is automated, the entire process of retrieving new
measurements, updating the model and making short-term
energy prediction can be performed in ‘real time’ on-line.
The objective of this paper is to evaluate the performance
of several adaptive neural network models for the on-line
prediction of building energy demand using both simulated
(synthetic) and measured data.
2. Literature survey

* Corresponding author. Tel.: +1 514 396 8667; fax: +1 514 396 8584. Regression models have been demonstrated to be
E-mail address: hugues.rivard@etsmtl.ca (H. Rivard). effective for building energy predictions in a number of

0378-7788/$ – see front matter # 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.enbuild.2005.02.005
 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259
experiments (e.g., [2–4]). Relatively few parameters must be
identified, thus reducing the time required for the model
development. The regression models do not, however,
accurately reflect the hourly or sub-hourly energy demand.
They are best suited for predicting the average consumption
over longer periods such as days or months. For different
buildings with different environment and weather condi-
tions, much effort and time must be spent on selecting time
scales and regressors to find a best fit model. Also, auto-
correlation or multicollinearity problems must be consid-
ered when evaluating the performance of prediction because
they tend to lead to model uncertainty.
The use of time-series analysis techniques to forecast
energy use is logical because the history of energy use can be
represented by a time series. Kimbara et al. [5] experimented
with the autoregressive integrated moving average
(ARIMA) model and found the performance of ARIMA
to be better than a two-dimensional autoregressive (AR)
model. Several models and applications have been
implemented based on the autoregressive moving average
with exogenous input model (ARMAX) model (e.g., [6]).
On one hand, time-series models can capture the relationship
between the hourly energy use and time variation given a set
of time-series data. On the other hand, both ARMA and AR
models work under the assumption that the present value is a
linear combination of the previous ones. In most cases, this
assumption is invalid. The ARIMA and ARMAX models
can handle the changes in an unstationary process, but
require the estimation of many parameters. Also, the auto-
correlation between variables must be considered because it
strongly impacts the accuracy of the prediction. Dhar et al.
[7] used a Fourier series model to predict the energy demand
in an institutional building. Fourier series models provide
better performance compared to the above time-series
models; however, they are based on the assumption that
energy use in most buildings is periodic. If dramatic changes
happen, high-frequency Fourier components must be
included in the model, thus dramatically increasing the
computational cost.
Artificial neural networks (ANN) is a type of artificial
intelligence technique that mimics the behavior of the
human brain. It can approximate a nonlinear relationship
between the input variables and the output of a complicated
system. The main advantage of an ANN model is its self-
learning capability. The use of ANN in building energy
prediction has been investigated by many researchers (e.g.,
[8,9]). Their models share some similarities, but each differs
significantly in implementation details because each is
tailored toward a specific type of energy prediction under a
specific building environment. On one hand, ANN models
estimate parameters faster by learning from examples
automatically. On the other hand, because it is hard to
distinguish structure from noise in the data, an ANN tends to
memorize noise. Also, an ANN might not be able to adapt to
dramatic changes such as unstable behavior in the power
load–temperature relationship.
1251
Most of the literature focuses primarily on static
prediction. In a static prediction, the prediction model is
set up in advance using historical data and does not change
afterward, when new information become available. It is
highly possible that such a model becomes invalid when new
patterns emerge and more recent data becomes available. In
this case, a dynamic prediction model that can adapt itself to
such changes in the energy consumption pattern is desirable.
This is especially true for short-term energy prediction. Only
a few dynamic prediction systems were found in the
literature and all in the field of electric load forecasting for
power system. They all used a sliding window approach in
which the size of data set used for training is kept constant;
however, the data set is periodically updated as the window
moves forward in time. Djukanovic et al. [10] used an
adaptive system for short-term load forecasting with a
moving window consisting of data associated with the 4
previous weeks as well as with 8 weeks at the same time in
the previous year. Khotanzad et al. [11] used a combination
of three separate models (i.e., weekly, daily and hourly
models) for short-term load forecasts. Each model is updated
at the end of each period with the data associated with that
period (e.g., at the end of each week, the weekly model is
updated using the last week data). Mohammed et al. [12]
used three ANN adaptive models and a two-stage training
algorithm. The first training stage produces a set of initial
ANN weights that capture the general, day by day trend of
the electric load. During the second-stage training, the ANN
is refined and enhanced to capture special features of that
particular day for which the forecast is made by using a
subset of data consisting of those that share similar
temperature conditions with the day being forecasted as
well as data from the previous 5 days. Charytoniuk and Chen
[13] used an ANN model with an adaptive scheme that
involves a moving window, which consists of training data
from the 3 last weeks as well as 2 weeks around the same day
in the previous year. These approaches have potentials in the
area of on-line building energy prediction.
The regression and time-series models are based on
classical mathematical theory. Thus, the behavior of these
models is well understood and the model parameters are
straightforward to estimate. However, these models tend to
work well only for energy systems that are well behaved.
The ANN model generally works better for buildings that
exhibit highly nonlinear energy consumption patterns.
However, the success of using ANN depends on a number
of design issues such as the choice of input and output data,
the number of hidden layers, the number of neurons used in
each layer and the training algorithms used. A few studies
have compared these forecasting methods. Kawashima et al.
[14] found that ANN models provide the best performance.
Dhar et al. [7] compared Fourier series models, ANN models
and the winners of the Great Energy Prediction Shootout
[15]. The performance of Fourier series was found to be
comparable to the performance of ANN models and that of
the winners of the shootout competition. Fourier series
1252 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259
models and artificial neural networks models are the two
most recommended methods found in the literature.
An ANN model has the unique advantage that no clear
relationship between the input variable and output needs to
be defined before the model is used in the prediction process,
since this relationship is identified through a self-learning
process. Because of this unique feature, the time and effort
that are normally required to establish a proper mathematical
model in a conventional prediction methodology can be
significantly reduced. The ANN-based models appeared to
perform well in the two Great Energy Shootout competitions
organized by ASHRAE. Therefore, this research adopts
ANN and explores ways to develop adaptive ANN models
for the on-line prediction of buildings energy demand.
3. Proposed artificial neural network models for
building energy prediction
This section presents two ANN models used for building
energy prediction.
3.1. Architecture issues for the ANN models
The electric, gas or chilled water demand or consumption
is usually the desired output of an ANN model used for
building energy prediction. Typical input elements include
the environmental data (e.g., outdoor dry-bulb temperature,
wet-bulb temperature, horizontal solar flux), time (e.g., hour
of the day) and operating parameters (e.g., status of chillers
or temperature of hot water leaving the boiler). For some
applications, not all environmental variables that contribute
to variations of energy usage may be available. In this case,
the accuracy of the ANN prediction will be limited by the
incompleteness of the data. On the other hand, if some
measurements are irrelevant to the energy usage to be
predicted, they carry no useful information and contribute
only noise to the ANN input. Removing these types of
measurements from the list of inputs can improve the
accuracy of the prediction and reduce the training time.
Feuston and Thurtell [16] used the principal component
analysis (PCA), a multivariate statistical analysis technique,
to assemble, synthesize and select relevant input variables
among a large number of measurements. As a consequence,
the neural network input vector does not consist of the
original input variables, but of linear combinations of these
variables. The PCA technique reduces the dimensionality of
data and removes redundancy by seeking clusters of data
points that can be used to represent the main features of the
data. This method is adopted in this study.
Typically one or two hidden layers are sufficient in an
ANN designed for building energy prediction. The number
of neurons in each layer may vary depending on the
quantities to be predicted. The presence of redundant
neurons does not pose a significant problem. A good
learning algorithm tends to ignore the excessive parameters
in the system. To ensure the accuracy of the ANN prediction,
a subset of the data is reserved for cross-validation. The
magnitude of the cross-validation error can be used as a
stopping criterion for the ANN training.
3.2. Adaptive ANN models
In a static prediction model, one tries to establish the
model in advance and estimate the model parameters using
historical data. Once the model is built, it is rarely changed
even though the presence of new data may indicate that the
model is no longer valid. To address this problem, this paper
evaluates the performance of adaptive ANN models that can
be constantly updated as new environmental and operational
data becomes available. The adaptive ANN prediction
models have an inherent self-revision capability to adapt to
weather and other condition changes. Two adaptive ANN
models are proposed: (1) the accumulative training and (2)
the sliding window training.
An ANN can be retrained periodically by a set of
augmented data infused with newly collected measurements.
This type of training strategy is referred to as accumulative
training. Accumulative training has the obvious advantage of
being able to identify both the local (for example, daily) and
the global (seasonal) trends of energy variation. Its main
disadvantage lies in the fact that the volume of data
accumulated continuously increases and may become too
large to be manageable. The larger the volume of the training
data, the longer it takes for training the ANN. It is also likely
that the latest changes in the accumulative training data set
have smaller impact on the model training because their
quantity is less compared to the older data.
Alternatively, the size of the training data set can be kept
constant and new measurements are added while some of the
oldest data are dropped from the training set. This approach
can be graphically viewed as periodically sliding a time
window across a time series of measurements to select the
training data. An adaptive ANN-based model on this
training technique will be referred to as a sliding window
ANN. The relative small and constant size of the training
data makes it possible to perform fast ANN re-training. The
drawback of this approach is that the training data may only
contain recent information, and the prediction result may not
accurately reflect the annual or seasonal change in energy
usage pattern. Also, determining the optimum window size
ahead of time is difficult. The window cannot be too large;
otherwise it defeats the purpose of limiting the amount of
training data. If the window size is too small, the resulting
ANN model maybe not completely capture the trend of the
energy demand.
4. Computational experiments using synthetic data
This section contains the computational results obtained
from applying the ANN models to a data set that was
 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259
obtained from the simulation of an office building using the
DOE 2.1E software. This data set is called synthetic data.
The Laval office building, located in Montreal was built in
1972 [17]. The building has a total floor area of 10,400 m2
spread over a seven-floor office tower, an underground
garage and a ground floor. There is a central variable air
volume system, which provides cooling in the summer and
ventilation all year to the office spaces. Direct expansion
cooling coils are connected to four condensing units, each
equipped with two compressors with a refrigeration capacity
of about 90 kW. The supply air temperature is controlled in
terms of the outdoor air temperature. The system is also
equipped with a dry-bulb temperature economizer system.
The simulated data associated with the Laval building is
noise free: the building is assumed to operate under normal
conditions, which do not change from week to week, and
from season to season. In addition, there are no measuring
errors, operation mistakes, faults or degradation of energy
performance of equipments in time. Because the simulated
data is generated from a well-behaved energy model defined
in the simulation software (DOE 2.1E), it provides an ideal
scenario under which the variation of the energy consump-
tion is more ‘‘predictable’’. Performing experiments on this
data set serves as the first step towards the understanding,
developing and testing a realistic ANN model.
Several experiments are performed on this data set.
Because a static ANN model serves as the building block for
a dynamic ANN, its construction and testing are described in
Section 4.3, before the experiments associated with the
dynamic models are shown in Section 4.4. Prior to present
these results, Section 4.1 describes the architecture of the
ANN models used followed by Section 4.2 that describes
data processing.
4.1. ANN architecture
The quantity of interest here is the hourly electric demand
of the chiller installed in the Laval building. The hourly data
set consists of: outdoor dry-bulb temperature, Td(t); outdoor
wet-bulb temperature, Tw(t); temperature of water leaving
the chiller, Tl(t); chiller electric demand (compressor and
fans of the condensing unit), E(t). It is assumed that the
electric demand of the chiller, at time t, is a function of some
of the above mentioned environmental and operation
variables collected at hours t i, for i = 1, 2, . . ., n. Those
variables will become inputs to an ANN to predict the
electric demand of the chiller at time t.
The ANN model used in the experiments consists of one
hidden layer in addition to the input and output layers. The
input layer contains n neurons used to feed n different inputs
into the network (n is a variable because it varies with the
experiments). The output layer contains one neuron from
which the predicted chiller electric demand is extracted. The
hidden layer consists of 2n + 1 neurons [18]. This three-
layer ANN model was found to be sufficient for making a
reasonably accurate prediction of the chiller electric demand
1253
for the Laval building. Adding more layers and/or neurons
can potentially improve the prediction accuracy, but also
adds complexity to the ANN training time.
The MATLAB ANN toolbox is used to build, configure
and train the network. The toolbox offers several algorithms
for training the network. All algorithms were tested in this
study. When the data size is small, the Levenberg–Marquart
(LM) algorithm appeared to be the fastest training algorithm
(the mean square error of the ANN output approaches to zero
at a quadratic rate). However, because the LM method must
solve a linear system of equations in order to obtain the
search direction, the computation becomes expensive when
the number of input elements and the volume of the training
data increase. Therefore, when the volume of the data is
large, the standard gradient descent algorithm is used for
training the ANN. Logistic sigmoid functions are used as the
activation function in each neuron. In the output layer, a
linear transfer function is used to allow the network to
produce values outside of the range [ 1, 1]. The training
process is terminated when the mean square error (MSE)
between the ANN output and the target values becomes less
than 10 5, or when a maximum of 500 epochs is reached.
The initial weights and biases of the ANN are generated
randomly.
4.2. Data processing
Besides defining the architecture parameters of the ANN
as discussed above, the data needs to be processed prior to
the training of the ANN. It has been recognized in the
literature that it is important to adopt a day-typing procedure
to separate energy data with distinct load patterns into
different prediction groups. Energy prediction can be made
within each group instead of on the entire data set. An
obvious separation can be drawn between weekdays and
weekends (holidays). The Laval building operates from 7:30
a.m. to 11:00 p.m., Monday–Friday. Since the chiller is not
used outside this interval, the non-working hours are
removed from the data set since prediction is only to be
made for the working hours. Without day-typing, the
implementation is easier. However, the prediction errors
increase if all data is included, as demonstrated below.
Two experiments were conducted by using data collected
in June to predict the electric demand of the chiller in July. The
prediction accuracy is measured by the coefficient of variation
(CV) and the root mean square error (RMSE). In the first test,
when the training process is carried out using only the
working-hour data, the statistics are the following: CV = 0.20
and RMSE = 27.0 kW. In the second test, the results for the
entire data set are: CV = 0.67 and RMSE = 34.4 kW. Since
the use of working-hour data improves significantly the
accuracy of prediction, all non-working hour data were
removed for the experiments presented below.
Because the numerical range of the input and output
variables may be quite different for some applications, it is
often useful to normalize the input and output variables so that
1254 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259

the training process does not suffer from severe numerical Table 1
round-off effects. In the following experiments, all input Statistical indices, CV and RMSE, used for comparing the difference
between the predicted and measured energy demand
variables are normalized to values between 1 and 1.
Experiment CV ( ) RMSE (kW)

4.3. Training and testing of static prediction models Static models for chiller electric demand using synthetic data
No. 1 0.04 6.10
No. 2 0.16 25.5
Three experiments were carried out with static prediction No. 3 0.15 22.0
models. The first experiment assesses the accuracy of ANN No. 3 (PCA) 0.07 11.4
to predict the chiller electric demand at time t based on input On-line models with accumulative training for chiller electric demand using
data at time t. The second experiment predicts the demand synthetic data
using time-lagged data by an hour (t 1). Finally, the third No. 4 (PCA) 0.15 28.3
experiment tests various combinations of time-lagged data No. 5 (PCA) 0.17 28.9
to improve the prediction. Seventy-five percent of data is On-line models with sliding window training for chiller electric demand
used to train the ANN models, and 25% of data are set aside using synthetic data
for testing. The second data set corresponds to the second No. 6 0.40 8.05
No. 6 (PCA) 0.15 27.7
week of each month.
No. 7 (PCA) 0.16 27.8

4.3.1. Experiment no. 1—Without time-lagged
measurements as input
The chiller electric demand at time t is predicted in terms
of Td(t), Tw(t) and Tl(t) at time t. Because the number of input
elements is small, the Levenberg–Marquardt algorithm was
used to train the ANN. Training takes 4.9 s on a personal
computer equipped with a Pentium III running at 933 MHz,
with 128 MB of RAM and Windows 98 as the operating
system. The mean square error (MSE) between the output of
the ANN and the target electric demand decreases below
10 3 in less than 10 epochs. Once the training process is
completed, the quality of the ANN prediction model is
assessed by predicting the chiller electric demand associated
with the second week of each month (the test data). The
accuracy is measured by using the CV and RMSE values of
the prediction error associated with the test: CV = 0.04,
RMSE = 6.10 kW (Table 1). Experiment no. 1 indicates that
the ANN model constructed in that experiment can
effectively capture the direct nonlinear relationship between
the chiller electric demand of the Laval building and various
temperature measurements at a given time. However, this
model cannot be used for the on-line prediction of the
electric demand at time t, since the input variables Td(t),
Tw(t) and Tl(t) are not available until the end of that hour.
4.3.2. Experiment no. 2—Using time-lagged
measurements as input
One possibility is to use h(t), Td(t 1), Tw(t 1) and
Tl(t 1) as input variables to train the ANN to predict E(t).
The variable h(t) indicates the hour for which to do the
prediction. It is observed that this change of input vectors to
the ANN leads to a longer training period of 11.4 s. Even
when the Levenberg–Marquardt algorithm is used, the mean
square error between the ANN output and the target
decreases slowly to zero. This difficulty is even more
pronounced when the ANN model is applied to the test data
for prediction. The error between the predicted electric
demand and the measured electric demand is defined by
CV = 0.16 and RMSE = 25.46 kW (Table 1).
Static models for chiller electric demand using measured data
No. 8 0.23 3.73
No. 9 (PCA) 0.26 4.28
On-line models with accumulative training for chiller electric demand using
measured data
No. 10 2.53 13.29
On-line models with sliding window training for chiller electric demand
using measured data
No. 11 0.26 12.88
4.3.3. Experiment no. 3—Using additional time-lagged
measurements as input and PCA to reduce the
dimension of the input
To improve the prediction accuracy, longer past history is
included as additional input data to train the ANN model.
Several combinations of time-lagged input data, up to 6 h
delay, are tried and tested (Table 2). It is assumed that h(t)
and Tl(t 1) are always chosen as input variables. The
inclusion of all temperature measurements collected at time
t i, for i = 1, 2, . . ., 6, (case D), does not necessarily gives
the best performance: CV = 0.22 and RMSE = 34.4 kW. The
best prediction result is obtained for case A, where the CV
and RMSE values are 0.15 and 22.0 kW, respectively.
The use of all available time-lagged measurements is not
effective because of undesirable redundancy. The redun-
dancy in the ANN input makes it difficult for the back-
propagation algorithm to capture the optimal weights and
biases for the desired ANN model. Thus, a meticulous
selection of time-lagged data must be used to train the ANN
model. However, in practice, it is not possible to try all
possible combinations of lagged timed temperature mea-
surements before a prediction is made. This problem can be
eliminated by using the principal component analysis (PCA)
technique to select appropriate input data from h(t),
Tl(k 1), Td(t k), Tw(t k), for k = 1, 2, . . ., 6. Six
principal components that contribute to more than 1% of the
variance of all past history data are retained. The training
time is 9.9 s, and the CV and RMSE values are reduced to
0.07 and 11.4 kW, respectively (Table 1). The CVand RMSE
 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259
Table 2
Prediction of the chiller electric demand at time t using different combinations of time-lagged temperature measurements (Experiment no. 3)
Outdoor dry-bulb temperature, Td
t 1 t 2 t 3 t 4 t 5 t 6
A 
B   
C     
D       
values of the prediction obtained in this experiment are
much better than the ones obtained in Experiment nos. 2 or 3
(without PCA).
4.4. Training and testing of on-line prediction models
The previous static ANN model is modified to take
advantage of new measurements that become available on a
continuous basis. This research evaluates two different
approaches to implement adaptive ANN. The first approach
simply accumulates all the measurements collected up to
time t, and retrain the ANN periodically using the entire set
of measurements. This is referred to as accumulative
training (presented in sub-Section 4.4.1). The second
approach maintains a fixed amount of training data by
discarding old measurements while adding new measure-
ments. This is referred to as sliding window training
(presented in sub-Section 4.4.2).
4.4.1. Accumulative training
Data about temperatures and chiller electric demand from
the month of June are set aside, and this portion of the data
file is used to establish, through training, what is called a
baseline ANN model for the chiller electric demand. The
baseline ANN model is then used to predict the chiller
electric demand associated with the first day of July. Once
the prediction has been performed, the hourly temperature
and electric demand measurements associated with the day
being predicted is added to the initial data set allocated for
the baseline training. This updated data set is used to retrain
the ANN model for carrying out subsequent predictions. The
weights and biases that emerge from the baseline model are
used as the initial weights and biases during the retraining
process. Since these initial weights and biases are likely to
capture some features of the nonlinear mapping between the
independent temperature variables and the electric energy to
be predicted, it is conceivable that retraining will not take as
long as the baseline training. This behavior is confirmed in
these experiments. The results presented below compares
the impact of different choices of input variables on the
accuracy of the accumulative on-line prediction model. In
particular, the fourth experiment uses only past temperature
measurements as input while the fifth experiment adds past
electric energy measurements as input.
4.4.1.1. Experiment no. 4—Accumulative training with
time-lagged temperature measurements as input. The input
1255
Outdoor wet-bulb temperature, Tw CV ( ) RMSE (kW)
t 1 t 2 t 3 t 4 t 5 t 6
      0.15 22.0
 0.17 26.5
      0.17 26.7
     0.22 34.4
to the ANN model consists of a combination of the past
history temperature measurements Td(t k), Tw(t k) and
Tl(t 1), where 1  k  6. The output of the ANN is the
chiller electric demand E(t). The PCA is applied to remove
the redundancy in the input and only components that
contribute to more than 1% to the variance are retained. Six
principal components emerged as the input to the ANN after
PCA has been applied.
The baseline model is modified and retrained daily by
including the electric demand and temperature measure-
ments collected on the previous day in the training data set.
The value of MSE between the target and ANN training
output converges to zero rapidly. The predicted chiller
energy demand matches the actual usage reasonably well
(Fig. 1). The CV and RMSE values obtained from testing are
0.15 and 28.3 kW, respectively.
4.4.1.2. Experiment no. 5—Accumulative training with
time-lagged chiller energy usage as input. In this experi-
ment, E(t - k) is added to the set of input variables. The
ANN model is constructed and trained in a way similar
to that carried out in Experiment no. 4. The predicted
electric demand matches the measurements reasonably
well (CV = 0.17 and RMSE = 28.9 kW) (Table 1). Since
the CV and RMSE values obtained in Experiment nos. 4
and 5 are comparable to those obtained with static ANN
models without PCA (Experiment nos. 2 and 3), one can
conclude that the on-line model with accumulative training
provides reasonable accuracy, when compared with static
models.
4.4.2. Sliding window training
The training data consists of the temperature and electric
demand measurements enclosed within the sliding window.
The ANN model used in the sliding window approach has
the same architecture as the one used in the accumulative
prediction model. The same parameter settings (such as the
learning algorithm, training convergence tolerance and the
maximum number of epochs allowed) are used for the
following experiments. The sixth experiment uses only past
temperature measurements as input while the seventh
experiment adds past electric demand measurements as
input.
4.4.2.1. Experiment no. 6—Training with sliding window
using temperature data collected in previous hours. Ex-
periments show that a window size of 20 working days
1256 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259
Fig. 1. Comparison between the predicted chiller electric demand and synthetic data using an on-line model with accumulative training and time-lagged
temperature measurements (Experiment no. 4).
Table 3
Comparison between the predicted and synthetic data using an on-line
model with sliding window training (Experiment no. 6)
Window size (days) CV ( ) RMSE (kW)
10 0.25 45.5
20 0.40 8.05
30 0.20 3.00
40 0.46 82.0
provides a reasonable balance between accuracy and
computational complexity per online prediction cycle
(Table 3). Hence, all subsequent experiments use this
window size. Measurements of temperature and electric
demand associated with the first 20 days of June were
selected as the initial set of training data. The prediction is
made on a daily basis. Thus, once the initial training is
completed and a prediction has been made for the electric
demand on the 21 working days in June, the hourly
temperature and electric demand measurements correspond-
ing to the first working day of June are removed from the
training data. The measurements of temperatures and
electric demand associated with the 21 days are added into
the training data and retraining is carried out. Consequently,
the volume of training data does not change, and the
selection window is shifted forward in time by 1 day. The
overall CV and RMSE values of the prediction are 0.40 and
8.05 kW, respectively.
After PCA is applied to the initial set of time-lag
temperature measurements to select principal components,
six principal components emerged as the ANN input.
However, the number of principal components may change
when the sliding window is updated. More or fewer principal
components may appear as daily training and prediction
move forward. When the number of principal components
associated with the new training data set is different from the
one associated with the previous training data, one cannot
restart from the ANN model obtained from previous training
cycle. Weights and biases must be reinitialized randomly,
and the training may take more time.
Fig. 2 shows that overall the training of sliding window
produces a qualitatively good prediction. The predicted
electric demand matches the actual demand curve reason-
ably well except at some hours where the actual load shows
some unexpected fluctuation. The overall CV and RMSE
values obtained in this experiment are 0.15 and 27.7 kW,
respectively (Table 1).
4.4.2.2. Experiment no. 7—Training with sliding window
using temperature and electric demand measured in the
previous hours as input. In this experiment, we investigate
whether adding E(t k) to the list of input variables: h(t),
Tl(t k), Td(t - k), Tw(t - k), k = 1, 2, . . ., 6, can improve
prediction accuracy. The PCA technique is used to remove
the potential redundancy in the data. There is no clear
improvement in prediction accuracy when E(t k) is
added. The CV and RMSE values of the prediction are 0.16
and 27.78 kW (Table 1). Results from the on-line models
using the sliding window training, applied to synthetic data,
are as accurate as those obtained from the accumulative
training.
 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259 1257

Fig. 2. Comparison between the predicted chiller electric demand and synthetic data using an on-line model with sliding window training and time-lagged
temperature measurements (Experiment no. 6).

5. Computational experiments using measured data Table 4 are initially identified to be the independent
variables that can potentially affect the variation of E(t). A
To evaluate the effectiveness of the proposed ANN closer examination of data reveals that not all variables listed
models some computational experiments are performed with in the data file are measured at every hour. Furthermore, the
measured data from a real environment. Measurements number of available measurements is different from one
contain a number of anomalies that make it more variable to another. The number of hourly measurements for
challenging to produce highly accurate prediction results. some variables is roughly 50% of the total number of hours
The computational experiments present the typical difficul- between the beginning and ending period of the measure-
ties encountered in developing an ANN model to predict the ments due to problems related to data collection. The low
energy demand in a real building. quality of the raw data makes it difficult to design, train and
This section contains the results of using the proposed test an accurate ANN model. However, this situation could
ANN techniques to predict the chiller electric demand of the occur in a monitoring system installed in buildings.
building housing the CANMET Energy Technology Center Therefore, the challenge is to develop, in the absence of
located in Varennes, Que., Canada. Because the original data
provided in this experiment is not prepared in a format that Table 4
can be directly used by the MATLAB code developed in this Input and output variables at time t related to the chiller electric demand
research, the raw data was first preprocessed and converted Variable Description
into the desired format. During the process of conversion, SC1(t) On/off status of compressor 1
several problems associated with the completeness and SC2(t) On/off status of compressor 2
fidelity of the data was discovered. Problems encountered SC3(t) On/off status of compressor 3
SC4(t) On/off status of compressor 4
and methodologies for addressing these problems are
SC5(t) On/off status of compressor 5
described below, before discussing the ANN architecture, SC6(t) On/off status of compressor 6
experiments and results. Te(t) Temperature of the water entering the ice tank
Tev(t) Temperature of the water entering the evaporator
5.1. Data Tlv(t) Temperature of the water leaving the evaporator
Hum(t) Outdoor relative humidity
TOD(t) Outdoor temperature
The data set contains measurements between 12:00 p.m. SV1(t) Is the chilled water prepared in the ice tanks? (yes/no)
June 21, 2002 and 12:00 a.m. March 27, 2003 and between SV2(t) Percentage of chilled water prepared in the ice tanks
11:00 a.m. May 8, 2003 and 0:00 a.m. July 10, 2003. The HD(t) Holiday indicator
objective of this study is to predict the chiller electric WS(t) Weekday schedule
CR(t) Electric current used by the chiller
demand E(t), in kW, at a particular time t. Variables listed in
1258 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259
complete information, an ANN model able to provide
reasonably accurate on-line predictions of electric demand.
The prediction of chiller electric demand is made only
when at least one compressor is turned on (indicated by
SCi(t) = 1 for i = 1, 2, . . ., 6). As a result, both training and
testing are only performed using measurements associated
with non-zero SCi(t) values.
5.2. ANN architecture
Similar to the ANN models used to predict the chiller
electric demand using synthetic data (Section 4), the ANN
models used in this section consists of one hidden layer in
addition to the input and output layers. The input layer
contains n neurons used to feed n different inputs into the
network. The output layer contains one neuron from which the
chiller electric demand is extracted. The hidden layer consists
of 2n + 1 neurons [18]. Logistic sigmoid functions are used as
the activation function in each neuron. In the output layer, a
linear transfer function is used to allow the network to produce
values outside of the range [ 1, 1]. The training process is
terminated when the mean square error (MSE) between the
ANN output and the target values becomes less than 10 3, or
when a maximum of 500 epochs is reached. The initial
weights and biases of the ANN are generated randomly.
Whenever possible, the Levenberg–Marquardt (LM) algo-
rithm is specified to train the network. When the number of
input elements or the volume of the data is large, the standard
gradient descent algorithm is used for training the ANN.
5.3. Training and testing for static prediction of chiller
electric demand
Two experiments were first carried out with static
prediction models. The eighth experiment assesses the
accuracy of ANN to predict the chiller electric demand at
time t based on input data at time t. The ninth experiment
predicts the demand using time-lagged data as input.
5.3.1. Experiment no. 8—Without time-lagged
measurements as input
In this experiment, the chiller electric demand at time t is
predicted in terms of all the variables listed in Table 4, at
time t, except the status variables (SCi(t), i = 1, 2, . . ., 6).
About 80% of the non-zero measurements data between
September 2002 and May 2003 are reserved for training.
After training the network for a maximum of 500 epochs, the
MSE becomes less than 10 3. The predicted chiller energy
usage matches well the actual chiller electric demand. The
CV and RMSE values are 0.23 and 3.73 kW, respectively
(Table 1). Training takes 2.4 s only.
5.3.2. Experiment no. 9—Using time-lagged
measurements as input
In this experiment, an ANN model is developed
that predicts E(t) based on measurements collected in
previous hours. The PCA is used to reduce the dimension
of the input and to remove redundancy in the data.
Six principal components that contribute to more than 1%
of the total variation are retained. Training time is 5.7 s.
The MSE of the ANN output at the end of the training
process is around 10 2. The overall accuracy of this ANN
model is satisfactory: CV is 0.26 and RMSE is 4.28 kW
(Table 1).
5.4. On-line prediction of electric demand with
accumulative training
The Experiment no. 10 was carried out to test the
accumulative training ANN model that uses the chiller
operating status at time t 1.
5.4.1. Experiment no. 10—Prediction using SCi(t 1)
In this experiment, the chiller related variables, measured
between September 2002 and May 2003, are selected for
baseline training. The volume of the training data is small (it
corresponds to 130 h). After the baseline training, the initial
ANN model is used to predict the chiller electric usage for
the next 24 h. In this accumulatively trained on-line model,
the ANN is updated daily by adding measurements that
become available on the day chiller energy is to be predicted
into the training data set.
The CV and RMSE values are 2.53 and 13.29 kW,
respectively (Table 1).
5.5. On-line prediction of electric demand with sliding
window training
The Experiment no. 11 was carried out to test the sliding
window training ANN model that uses the chiller operating
status at time t 1 as input.
Like in Experiment no. 10, the chiller related
variables measured between September 2002 and May
2003, are set aside for baseline training. Once the baseline
training is completed, the initial ANN model is used to
predict the chiller electric usage for the next 24 h. The
ANN is updated daily by adding new measurements into
the training data set and deleting some previous measure-
ments from the training set. Only data recorded during the
hours during which a chiller is on are added to the training
data set.
5.5.1. Experiment 11—Prediction using SCi(t 1)
The difference between the predicted and measured data
becomes large: CV and RMSE values are 0.26 and
12.88 kW, respectively (Table 1). The drawback with this
approach is that the prediction assumes that if in the last
hour the chiller was ON, it will be ON in the next hour. So
when the chiller is turned OFF, the system will miss it by
1 h since it assumes it is ON based on the previous hour.
Only in the following hour will the system finally ‘‘sees’’ it
as OFF and then works properly. The prediction of the ON/
 J. Yang et al. / Energy and Buildings 37 (2005) 1250–1259
OFF status of the chiller is always lagged by 1 h. One way
to address this would be to develop another ANN that
would predict when the chiller will be turned ON or OFF.
This could not be achieved at the moment due to the lack of
data.
6. Conclusions
Most of the surveyed literature focused on using static
ANN models to predict energy demand at time t, when all
independent parameters are known at the same time t.
Although this modeling approach has serious drawbacks for
the on-line prediction, it is presented in this paper for the
sake of comparison. With synthetic, noise-free data the
training time is between 10 and 11 s, and the coefficient of
variance (CV) is between 0.07 and 0.16. The static models
applied to real measurements lead to lower accuracy
(CV = 0.23–0.26) than in the case of synthetic data
(CV = 0.07–0.16). This is due to the lower quality of
available input data, the smaller amount of data and the
complex operation strategies of chillers in the existing
building, compared to the simulated one. Therefore, one
can expect that the adaptive models would have an even
lower accuracy. This is not always the case, as presented
below.
Two types of adaptive ANN training schemes are
presented in this paper. In the case of synthetic data, the
accumulative training technique appears to have an equal
performance with the sliding window training approach in
terms of CV (0.15–0.17). In the case of real measurements,
the sliding window technique gives better results, compared
with the accumulative training, if the coefficient of variance
is used as an indicator: CV of 0.26, compared with 2.53 (for
accumulative training).
Future research work will use another institutional
building, having a larger volume of operational data over
several years, to estimate the optimal window size for the
sliding window training approach. Other architectures and
types of neural networks, such as recurrent neural networks,
will also be considered.
Acknowledgement
The authors acknowledge the technical assistance and
financial support received from CANMET Energy Technol-
ogy Centre of Natural Resources Canada, in Varennes,
Quebec.
1259
References
[1] A. Dhar, T.A. Reddy, D.E. Claridge, A Fourier series model to predict
hourly heating and cooling energy use in commercial buildings with
outdoor temperature as the only weather variable, ASME Journal of
Solar Energy Engineering 121 (1) (1999) 47–53.
[2] M. Fels, Special Issues devoted to measuring energy savings: the
scorekeeping approach, Energy and Buildings 9 (1–2) (1986) 5–18.
[3] D. Ruch, D.E. Claridge, A four–parameter change-point model for
predicting energy consumption in commercial buildings, ASME
Journal of Solar Energy Engineering 114 (2) (1992) 77–83.
[4] S. Katipamula, T.A. Reddy, D.E. Claridge, Multivariate regression
modeling, ASME Journal of Solar Energy Engineering 120 (1998)
177–184.
[5] A. Kimbara, S. Kurosu, R. Endo, K. Kamimura, T. Matsuba, A.
Yamada, On-line prediction for load profile of an air-conditioning
system, ASHRAE Transactions 101 (2) (1995) 198–207.
[6] Y. Hong-Tzer, H. Chao-Ming, H. Ching-Lien, Identification of
ARMAX model for short term load forecasting: an evolutionary
programming approach, IEEE Transaction on Power Systems 11
(1) (1996) 403–408.
[7] A. Dhar, T.A. Reddy, D.E. Claridge, Modeling hourly energy use in
commercial buildings with Fourier series functional form, ASME
Journal of Solar Energy Engineering 120 (3) (1998) 217–223.
[8] M. Anstett, J.F. Kreider, Application of neural networking models to
predict energy use, ASHRAE Transactions 99 (1) (1993) 505–517.
[9] P.S. Curtiss, J.F. Kreider, M.J. Brandemuehl, Energy management in
central HVAC plants using neural networks, ASHRAE Transactions
100 (1) (1994) 476–493.
[10] M.R.S. Djukanovic, B. Babic, D.J. Sobajic, Y.H. Pao, A neural-based
short term load forecasting using moving window procedure, Electric
Power and Energy System 17 (6) (1995) 391–397.
[11] A. Khotanzad, R.C. Hwang, A. Abaye, A.D. Maratukulam, An
adaptive modular artificial neural network hourly load forecaster
and its implementation at electric utilities, IEEE Transactions on
Power System 10 (3) (1995) 1716–1722.
[12] O. Mohammed, D. Park, R. Merchant, T. Dinh, C. Tong, A. Azeem, J.
Farah, C. Drake, Practice experience with an adaptive neural networks
short-term load forecasting system, IEEE Transactions on Power
Systems 10 (1) (1995) 254–265.
[13] W. Charytoniuk, M.S. Chen, Very short-term load forecasting using
artificial neural networks, IEEE Transactions on Power Systems 15 (1)
(2000) 263–268.
[14] M. Kawashima, C.E. Dorgan, J. Mitchell, Hourly thermal load pre-
diction for the next 24 h by ARIMA, EWMA, LR, and an artificial
neural network, ASHRAE Transactions 101 (1) (1995) 186–200.
[15] J.S. Haberl, S. Thamilseran, The great energy predictor shootout II-:
measuring retrofit savings-overview and discussion of results, ASH-
RAE Transactions 102 (1) (1996) 419–435.
[16] B. Feuston, J. Thurtell, Generalized non-linear regression with ensem-
ble of neural nets, ASHRAE Transactions 100 (2) (1994) 1075–1080.
[17] R. Zmeureanu, L. Pasqualetto, F. Bilas, Comparison of cost and energy
savings in an existing large building as predicted by three simulation
programs, in: J.W. Mitchell, W.A. Beckman (Eds.), Proceedings of the
IBPSA Building Simulation’95 Conference, Madison, 1995, pp. 485–
492.
[18] R. Hecht-Nielson, Theory of the backpropagation neural network,
Proceedings of IEEE International Conference on Neural Networks 1
(1989) 593–605.