RESEARCH

A STUDY ON THE APPLICATION OF ARTIFICIAL

INTELLIGENCE TECHNIQUES FOR PREDICTING THE
HEATING AND COOLING LOADS OF BUILDINGS

Sushmita Das,1 Aleena Swetapadma,2* and Chinmoy Panigrahi3

ABSTRACT

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

The prediction of the heating and cooling loads of a building is an essential aspect in
studies involving the analysis of energy consumption in buildings. An accurate esti-
mation of heating and cooling load leads to better management of energy related tasks
and progressing towards an energy efficient building. With increasing global energy
demands and buildings being major energy consuming entities, there is renewed
interest in studying the energy performance of buildings. Alternative technologies
like Artificial Intelligence (AI) techniques are being widely used in energy studies
involving buildings. This paper presents a review of research in the area of forecasting
the heating and cooling load of buildings using AI techniques. The results discussed
in this paper demonstrate the use of AI techniques in the estimation of the thermal
loads of buildings. An accurate prediction of the heating and cooling loads of build-
ings is necessary for forecasting the energy expenditure in buildings. It can also help
in the design and construction of energy efficient buildings.

KEYWORDS
building energy performance, heating and cooling load, Artificial Neural Network,
Support Vector Machine, Iteratively Reweighted Least Squares, Random Forest

I. INTRODUCTION
Rapid advances in building technology and growing urbanization has led to a vast increase in
the number of buildings worldwide. This increase has led the building sector to become a major
energy consuming entity. Energy consumption by buildings is as high as 32% of the overall
energy consumption worldwide [1]. Thus, it has become a key area for researchers to devise
methods to make buildings more energy efficient and environmentally friendly. The energy
consumption in buildings can be attributed to their lighting system, heating, ventilation and
air conditioning (HVAC) system, pumping system and various other electrical equipment used
by building occupants. Among these subsystems in a building, the HVAC systems consume the
largest amount of energy. Almost 30% of the energy consumed by buildings can be attributed to

1. Department of Electrical Engineering, Trident Academy of Technology, Bhubaneswar, India

2. School of Computer Engineering, KIIT University, Bhubaneswar, India;
*Corresponding author: [email protected]
3. School of Electrical Engineering, KIIT University, Bhubaneswar, India

Journal of Green Building 115

their HVAC systems [2, 3, 5, 6]. Thus, there is ample scope to make buildings energy efficient
by making HVAC systems more efficient. In order to design better HVAC systems for build-
ings with regard to reduced energy consumption, it is important to forecast the heating and
cooling loads of buildings as accurately as possible. Conventional methods of calculating the
heating and cooling loads of buildings is very time consuming. An alternative approach to the
prediction of heating and cooling loads of buildings is by using AI techniques. There may be
two approaches, namely blackbox modelling and gray box modelling. In blackbox modelling,
usually the sensor data from buildings is used for modelling and a prediction of the heating
and cooling loads; whereas, in the gray box modelling approach, building characteristics such
as surface area, glazing, wall area, relative compactness, orientation etc. are also taken into

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

account in addition to data collected through sensors and weather data. Several AI techniques
like Artificial Neural Networks (ANNs), Support Vector Regression (SVR), Classification and
Regression Tree (CART), Iteratively Reweighted Least Squares (IRLS), General Regression
Neural Networks (GRNN), Random Forest (RF), Chi-Squared automatic interaction detec-
tor, ensemble models, deep learning algorithm and linear regression models have been used for
the purpose of predicting the heating and cooling loads of buildings. Table 1 displays the key
traits of each AI technique mentioned. This paper presents a review of research that has used
AI techniques for the estimation of the heating and cooling loads in buildings.

2. DISCUSSIONS OF MODELS PROPOSED FOR HL AND CL ESTIMATION

OF BUILDINGS
This section presents an overview of the AI techniques used for the estimation of HL and CL
of buildings, the building types considered, the software used to calculate the HL and CL and
set the reference values and the input variables considered for each model. Table 2 presents the
overview of the models discussed in the next section:

3. DISCUSSION OF THE MODELS

A. Tsanas and A. Xifara in [12] have proposed statistical machine learning models for accu-
rately predicting the heating load and cooling load of residential buildings. The authors have
selected eight input variables for prediction of the heating load and cooling load which are the
two output variables. The input variables are selected based on previous literature which sug-
gests these variables to be the factors primarily affecting a building’s energy requirements. The
variables chosen as inputs are relative compactness, surface area, wall area, roof area, overall
height, orientation, glazing area and glazing area distribution. The building dataset was gener-
ated using Ecotect. The volume of all the buildings was the same (771.75 m3 ). However, the
surface areas and dimensions of the buildings were different. Low U-Value building materials
were considered. A total of 768 buildings were studied. These building samples were gener-
ated using Ecotect considering 12 building forms, 3 glazing area variations with 5 glazing area
distributions each for 4 orientations and 12 building forms for 4 orientations without glazing
((12x3x5x4=720) + (12x4=48) = 768). The machine learning techniques used to analyse the
data were IRLS (Iteratively Reweighted least Squares) and RF (Random Forests). The values of
HL and CL obtained from Ecotect were assumed to be the actual values and the results obtained
using the proposed models were compared against these actual values. The results showed that
RF greatly outperformed IRLS. The performance of the two models was evaluated in terms of
MAE, MSE and MRE. A limitation of the study is that climate and occupancy were assumed

116 Volume 14, Number 3

TABLE 1. Various methods used.

AI Technique Advantages Disadvantages

ANN Capable of learning and modelling Computational time
nonlinearities and complexities in data. taken for deep learning networks is long.
Ability to generalize and predict new Huge data is required for ANN
data presented to model. architectures with multiple layers.
Does not impose any limitations on the Are not probabilistic.
input variables and their distribution.

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

SVR Has powerful regularization Satisfactory performance largely depends
characteristics which are helpful in on the proper selection of kernel.
generalizing the model to new data. Computation time taken is longer.
Ability to perform satisfactorily on Complex algorithms are used and large
datasets with several attributes even if memory required for programming in
number of cases available for training big tasks.
the model are less.
CART Ability to perform feature selection. Even a small variation in data can result
Data preparation tasks like tackling in a large change in the design of the
differences in parameter scales and optimal decision tree.
accounting for missing values in the data Often gives less accurate results in
are not required. comparison to other techniques.
IRLS Can be used with Gauss-Newton and Computational time taken is longer.
Levenberg-Marquart algorithms.
GRNN Applicable to regression, prediction and Huge data involved can make
classification problems. computation costlier.
Uses single pass learning approach so no Not possible to improve network by
back propagation is required. optimisation.
High accuracy of prediction.
RF Highly flexible and accurate. Highly complex and takes a lot of
Does not require data preparation. time to construct as compared to
decision trees.
Helps overcome the problem of over
fitting faced while using a single Less intuitive.
decision tree. Computational time is very large.
Performs satisfactorily even if a large
amount of data is missing.

to be constant. Including these as input variables will improve the accuracy of prediction for
HL and CL.
The authors in [13] have used different data analytic methods for the prediction of heating
and cooling loads of buildings. The methods used are Support vector regression (SVR), ANN,
Classification and Regression Tree (CART), General Linear Regression, Chi-Squared Automatic
Interaction Detector and Ensemble/Combined models of two, three, four and five single
models. The building data was the same as the data used in [12]. Ecotect was used to simulate
twelve building types with the same volume and same materials used for construction but their

Journal of Green Building 117

TABLE 2. Overview of models described.

Authors and No. of Software used AI

Reference input for calculation Techniques Building
no. variables Input variables of HL and CL Outputs used Typology
A. Tsanas 8 1. relative compactness Ecotect HL and RF and Residential
and A. Xifara 2. surface area CL IRLS
[12] 3. wall area
4. roof area
5. overall height
6. orientation

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

7. glazing area
8. glazing area
distribution
J. Chou and 8 Same as in [12] Same as in HL and SVR, ANN, Residential
D. Bui [13] [12] CL CART,
Ensemble
Models
M. Cheng 8 Same as in [12] Same as in HL and EMARS, Residential
and M. Cao [12] Cl MARS,
[14] BPNN,
RBFNN,
CART,
SVM
D. 9 1. atmo- Energy Plus HL Multilayer Commercial
Kapetanakis spheric temperature Perceptron
et al [15] 2. ambient rela- ANN,
tive humidity RBFNN
3. wind speed
4. solar radiation
5. clearness of the sky
6. indoor air temperature
7. indoor rela-
tive humidity
8. occupancy
9. CO2 levels
Y. Sonmez et 8 Same as in [12] Same as in HL and k-NN, Residential
al [16] [12] CL abc-kNN,
ga-kNN,
classical
ANN,
ga-ANN,
abc-ANN
C. Deb and 5 1. air conditioned load ---- CL Feed Commercial
et al [17] 2. non-aircondi- forward (Educational)
tined load ANN
3. air temperature
4. relative humidity
5. solar radiation

118 Volume 14, Number 3

TABLE 2. (Continued)

Authors and No. of Software used AI

Reference input for calculation Techniques Building
no. variables Input variables of HL and CL Outputs used Typology
C. Fan et al 7 1. Month ------ CL Deep Commercial
[18] 2. Day Learning (Educational)
3. Hour
4. Minute
5. Day type
6. Outdoor temperature

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

7. Outdoor relative
humidity
B. Gunay et 6 1. Outdoor Energy Plus HL and ANN Commercial
al [19] air temperature CL (Office)
2. Horizontal
solar irradiance
3. Wind speed
4. Moisture content of
outdoor air
5. Electrical loads
6. Work hour
H. Pombeiro 10 1. Weekday ------ HL and Fuzzy Commercial
et al [20] 2. Hour CL networks, (Institutional)
3. Minute Neural
4. Occupancy networks
5. Temperature
6. Relative humidity
7. Wind speed
8. Atmospheric pressure
9. Precipitation
10. Solar radiation
A.Jihad and 6 1. Height DesignBuilder HL and ANN Residential
M.Tahiri [21] 2. Relative CL
Compactness
3. Wall surface
4. Building surface
5. Orientation
6. Window ratio

dimensions and surface areas were different. Building activities were assumed to be sedentary in
nature. Eight input variables, namely Relative compactness, Surface area, Wall area, Roof area,
Overall height, Orientation, Glazing area, Glazing area distribution were considered and the
two output variables were the heating load and cooling load. The performances of the various
models used for estimation of heating and cooling load were compared using evaluation indi-
cators like Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute
Percentage Error (MAPE), Linear Correlation Coefficient R and a Synthesis Index (SI) which
is a combination of RMSE, MAE, MAPE AND 1-R. The performance outputs of the different
models showed that the cooling load was predicted with the highest accuracy using the ensemble

Journal of Green Building 119

model of SVR and ANN with a SI of 0.11; whereas, the model best suited for estimation of the
heating load was the SVR model with a SI of 0.00. It was also seen that the computation time
taken by the suggested models was significantly less and within a few seconds. The limitation of
this work is the use of default settings in the models, both for the single and ensemble models.
The models proposed in the paper are only applicable to the twelve building types which are
specified. The authors suggest that higher accuracy in prediction of HL and CL of the buildings
can be achieved by optimization of parameters in the models (single and ensemble) by using
evolutionary computing and swarm intelligence techniques.
The authors M. Cheng and M. Cao in [14] have proposed an AI model EMARS
(Evolutionary Multivariate Adaptive Regression Splines) to accurately forecast the energy per-

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

formance of buildings. The EMARS model is a hybrid of MARS (Multiple Adaptive Regression
Splines) and Artificial Bee Colony. The proposed model was developed using data from [12].
The performance of the proposed EMARS model was compared with that of five other AI
models namely, MARS, Back Propagation Neural Network (BPNN), Radial Basis Function
Neural Network (RBFNN), CART and Support Vector Machine (SVM). The performance of
the EMARS model was compared with that of other models on the basis of RMSE (Root Mean
Squared Error), MAPE (Mean Absolute Percentage Error), MAE (Mean Absolute Error) and
R2 (Coefficient of Determination). The results showed that among all models, the proposed
EMARS model had the lowest RMSE values of 1.00 for cooling load and 0.47 for heating load.
The EMARS model gave superior results even in terms of MAPE, MAE and R2. The accuracy
of predicting the heating load was better than that of the cooling load. The parameter analysis
showed that the two factors primarily affecting the heating load are surface area and roof area
whereas the cooling load is equally affected by six out of the seven factors used for mapping.
The limitation of the work is the use of simulation data. The model should be tested taking
actual datasets.
Data driven or blackbox models have been developed for the estimation of heating loads
in commercial buildings in the work proposed in [15]. The data is collected from Building
Energy Management Systems (BEMS) installed in the buildings and in addition weather data
is also included. To resolve the problem of lack of some data in the data collected from BEMS,
a reference building developed using Energy Plus was set as a benchmark. Sensors are located in
the buildings and these sensor data are the BEMS variables. Statistical techniques were used for
analyzing the correlation between the input variables and output variable. The output variable
is the heating load of the building which is to be predicted and the input variables chosen for
analysis were atmospheric temperature, ambient relative humidity, wind speed, solar radiation,
clearness of the sky, indoor air temperature, indoor relative humidity, occupancy and CO 2
levels. The predictive models developed were Regression models- Multiple Linear, Multiple
Nonlinear and Generalized Linear and ANN models- Multilayer Perceptron and Radial Basis
Function. The performance accuracy of each model was evaluated using an equation involv-
ing Mean Absolute Error (MAE) and maximum and minimum predicted values. Correlation
studies done by calculating Pearson coefficient and Spearman analysis showed that there is a
very strong correlation- between a. Sky clearness and Solar Radiation and b. Heating Load and
Gas Consumption. A combination of different situations led to the development of 90 regres-
sion models and 60 ANN models. The regression models had an average accuracy of 74.52 %
whereas the ANN models had an average accuracy of 83.31 %. The most accurate regression
model was obtained using generalized linear regression technique and the accuracy was 92.1%.
The most accurate ANN model was generated using the multilayer perceptron approach and the

120 Volume 14, Number 3

accuracy was 97%. The common inputs to both these models were atmospheric temperature,
ambient relative humidity, solar radiation and indoor air temperature. It was concluded that
models based on ANN predicted the heating load of a building better than models based on
regression techniques.
The authors in [16] have used hybrid Artificial Intelligence (AI) techniques to predict the
heating load and cooling load of buildings. The data required for analysis is obtained from [12].
In this work six machine learning algorithms have been used for predicting the heating load and
cooling load. These are classical estimator knn (k-nearest neighbor), artificial-bee-colony based
k- nearest neighbor (abc-knn), genetic algorithm based knn (ga-knn), classical ANN, ga-ANN
and abc-ANN. Out of the 768 datasets, 576 were used for training and 192 for testing and

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

verification. Results showed that among the knn models, the abc-knn gave the best performance
with a MAE of 1.10. It was observed that the input variables most affecting the cooling load
are surface area, overall height and glazing area whereas the input least affecting the cooling
load of a building is glazing area distribution. Among the ANN models, the abc-ANN model
performed best with a MAE of 0.52. When compared with the abc-knn, both the adaptive
ANN models performed better. While MAE for abc-knn is 1.10, it is 0.61 for ga-ANN and
0.52 for abc-ANN. It has been concluded that the accuracy of predicting the cooling load and
heating load of buildings can be significantly improved with the use of heuristic based hybrid
knn and ANN methods.
A model using ANN was developed to forecast the daily cooling load energy consumption
for three institutional buildings in a university campus in Singapore [17]. The three buildings
were part of the same school but the purpose and use of each building varies. Energy consump-
tion data of the three buildings was collected over two years. The forecasting model developed
was made more relevant by the inclusion of weather data namely air temperature, relative
humidity and solar radiation. The air conditioning load was correlated with these three climatic
factors and the results of correlation showed that none of the climatic factors considered had
any significant impact on the air conditioning load of the buildings. Rather it was concluded
that the key factors affecting cooling load energy used in these buildings are building occupancy
and indoor conditions. A feedforward ANN network consisting of an input layer, hidden layer
and output layer was used. Using a trial and error method, it was seen that when the number of
hidden neurons is 20, the model gave accurate results in less time. Thus, the number of neurons
in the hidden layer was fixed to 20. Different ANN architectures were obtained by adding more
hidden layers and tested for prediction accuracy; it was observed that the computing time for
the 5-20 network was the least. A 5-20 network implies five input neurons and one hidden
layer of 20 neurons. A Bayesian regularization algorithm was used as the training algorithm
and the training was done in MATLAB. The energy consumption data was equally divided into
five classes ranging from very low, low, medium, high and very high. This was done to reduce
the high variation in data. The data was also distributed into 25 class numbers. This helps in
differentiating between the different week days, and there is no overlapping of data between
class levels of different days. The inputs to the ANN were the previous five days’ measured
energy consumption data in the form of energy classes. With this approach of energy classes
and class numbers, the model predicted results with an accuracy as high as 0.9794. The model
is able to predict the daily cooling load for the next 20 days with good prediction accuracy
by taking only the previous five days measured data values as inputs. The study suggests that
in the future a forecasting model can be developed to make it applicable for a wider range of
institutional buildings.

Journal of Green Building 121

 In [18], the authors have proposed a method to forecast the cooling load for the next 24
hours by use of deep learning algorithms. Deep learning can be defined as a method which
involves the combination of various machine learning algorithms. Deep learning algorithms
are capable of displaying nonlinear and complicated formations in big data. The building
whose data has been used in this study is an educational building in Hong Kong. The building
mainly consists of classrooms, offices and a data center. The total floor area is 11,000 m2 out of
which the conditioned area is 8500 m2. Annual data from 2015 was collected and the collected
data had a time interval of 30 minutes. The set of collected data had the mentioned variables:
time variables such as month, day, hour, minute and day type ( i.e. weekday or weekend),
external temperature, external relative humidity, supply chilled water temperature, returned

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

chilled water temperature and flow rate of chilled water temperature. The total number of
observations in the dataset is 15,792. Deep learning technique can be utilized in a supervised
way to form a forecasting model or in an unsupervised way for retrieving important attri-
butes from unanalyzed data. In this work, first feature extraction was done to retrieve relevant
attributes from the building data which were then given as inputs to the model. Four types of
feature extraction methods were used to extract important engineering, statistical, structural
and deep learning features from the building data. This step of feature extraction shows the
unsupervised aspect of deep learning. Supervised deep learning is highlighted by the seven
forecasting techniques used to form prediction models using various feature sets. The predic-
tion techniques used are Multiple Linear Regression, Elastic Net, Random Forests, Gradient
Boosting Trees, Support Vector Regression, Extreme Gradient Boosting and Deep Learning.
Seventy percent of the dataset was used for training, 15% for validation and the remaining
15% for testing. The performance of the prediction models is evaluated on the basis of indi-
cators like Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Coefficient
of Variation of the Root Mean Square Error (CV-RMSE). Results show that the models
developed using Multiple Linear Regression and Elastic Net give the poorest performances
among all models proposed. Among the remaining five techniques, the model developed using
Extreme Gradient Boosting gives the best performance. The prediction outputs using linear
techniques is found to be very poor as compared to the nonlinear prediction techniques. In
other words, it is the nonlinear prediction techniques that are preferable over linear techniques
in such situations. The model using Extreme Gradient Boosting gave the best performance
in the prediction of cooling load for the next 24 hours. It can been concluded that using the
features of the extraction step can greatly enhance the prediction capability of the models
using nonlinear forecasting techniques.
The authors in [19] have proposed inverse blackbox modelling for estimation of heating
and cooling load of buildings. The buildings chosen for the study are five office buildings in
Ottawa, Canada. The data collected for the study were from five buildings during three years’
hourly data of heating and cooling load, hourly electricity consumption data for plug-in equip-
ment, lighting, fans and pumps for the same period for each building, simultaneous weather
data such as external temperature, solar irradiance, relative humidity and wind speed recorded
at 15 minute intervals at a nearby weather station. Two types of inverse blackbox models were
developed with one based on Linear Regression and the other based on ANN. The ANN
models were tested for two configurations- a. one hidden layer and one output layer and b. two
hidden layers and one output layer. Sigmoid activation functions were used in hidden layers
whereas linear activation functions were used in the output layer. The Levenberg- Marquardt

122 Volume 14, Number 3

back propagation algorithm was used for parameter estimation. In the Linear Regression based
model, the method of least squares was adopted for estimating unknown parameters. In both
the models, the inputs were added one at a time to study their effect on the performance of the
model. Before training the models, a correlation study was done between the inputs and the
heating and cooling loads by calculating the average of the Pearson correlation coefficients for
all the five buildings under study. The evaluation criteria used for assessing the performance of
the developed models are Root Mean Square Error (RMSE) and R2 values. The results showed
lower RMSE values for ANN based models as compared to the linear regression-based model.
This can be attributed to ANN’s ability to handle nonlinearities in data better. There was no
significant difference in performance of the two ANN models which implies that increasing

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

the complexity of the model will not necessarily lead to better performance. An analysis of
performance results of the three models led to the selection of a single layer ANN model with
six inputs and one hour input record as the final inverse blackbox model.
Pombeiro and et al. in [20] have done a comparative study of models for predicting the
energy usage in an institutional building. The models used are based on linear regression, fuzzy
networks and neural networks. The work shows that fuzzy and neural network models have
much better prediction accuracy than linear regression models. The name of the building chosen
for the case study is IST-T located in Oeiras, Portugal, with a Mediterranean climate, which can
be classified as mild in nature. The datasets used for the models were electricity consumption
data having energy readings for the month of May, from 01/05/2014 to 31/05/2014, with data
being recorded at an interval of every 15 minutes, weather data constituting temperature, rela-
tive humidity, wind speed, precipitation, atmospheric pressure and solar radiation, occupancy
data and time data pertaining to type of day i.e. weekday or weekend and time of day. The energy
usage data was collected every 15 minutes by a smart meter installed in the electricity board of
the building which was then converted into kWh. Since the building does not have any sensors,
an indirect method using local area connected WiFi users was adopted. The weather data was
collected from a weather station located on the roof of the building. The correlation between
the input variables was studied, and it was observed that temperature has a high correlation
with relative humidity (54.3%) followed by temperature with solar radiation (29.9 %). Another
significant correlation was between relative humidity and solar radiation being 29.6%. A correla-
tion of 21.5% was observed between occupancy and time of day. For each of the three models,
the datasets were divided into input and output data with a division of 60% training data and
40% test data. The MATLAB Statistics Toolbox was used for developing the linear regression
model. The function LinearModel.fit was used for the model. The fuzzy model proposed in
the work was formed using the function genfis2 from the MATLAB Fuzzy Logic Toolbox.
The genfis2 function applies the Sugeno type FIS with Subtractive Clustering algorithms. A
feed-forward back propagation network with one hidden layer using fitnet function from the
MATLAB Neural Network Toolbox was used for the neural network model developed. In the
fuzzy and neural models, a low number of parameters were varied to train the models, the aim
being to develop low complexity models. In the fuzzy model, the radii parameter was varied
whereas in the neural model, the number of neurons in the hidden layer was varied. The results
show that the linear regression model was not effective enough to justify the load profile. The
VAF was 35.9 % and the MAE was 11.8 kWh. The fuzzy model showed considerable improve-
ment with a VAF of 70.3% and MAE of 6.0 kWh. The neural network modelled also showed
better performance when compared to the linear regression model with a VAF of 61.6% and

Journal of Green Building 123

MAE of 8.5 kWh. Results showed that occupancy data when used improves the performance
of models. Weather data did not provide any significant improvement in performance of the
models, the reason being the chiller which is the main energy consuming load in the building
was operated everyday at the same period, regardless of the temperature variation.
Authors A. Jihad and M. Tahiri in [21] have predicted the heating load and cooling load
of residential buildings in the climatic zone of Agadir, Morocco, by using ANN as a learn-
ing Algorithm. The database of buildings required for the study was generated using Design
Builder. The heating load and cooling load requirements for different building scenarios was
then simulated with Energy Plus. The methodology used for dataset generation in the study is
similar to the work done by Tsanas and Xifara in [13]. The learning algorithm invovled a gradi-

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

ent descent optimization. The ReLU (Rectified Linear Unit) was used as an activation function
at the output of each neuron to eliminate all negative input values and increase the convergence
speed. The ANN model trains the data of the 5625 buildings generated using Design Builder for
prediction of heating and cooling loads. The highest prediction accuracy of 98.6% was reached
after 25 million iterations. The validity of the proposed model was tested using data from three
residential buildings in Morocco. The heating and cooling loads for the three buildings predicted
using the proposed neural network model were compared with the simulated HL and CL data
obtained using Design Builder. The proposed model showed an accuracy of 97.6% for the test
data and an accuracy of 98.7% for the training data. The ANN model when compared to the
model developed using Design Builder showed an improved prediction accuracy of 94.8% and
98.5% for training and test data respectively. It was concluded that the model can easily predict
the energy requirements of a new or an existing building without having to calculate thermody-
namic balance or use the simulation software. The limitation of the work is the use of only one
climatic zone; therefore, the model cannot be used for situations involving other climatic zones.

4. CONCLUSION
In most of the buildings, whether residential or commercial, the important objectives of thermal
comfort and reduced energy consumption are not achieved. Recently developed models for
predicting energy requirements of buildings are using AI techniques and results show that
these strategies have improved the situation. A difficult and time consuming aspect in this area
of study is the collection of actual building energy datasets. However, this limitation has been
greatly reduced recently with many studies using real data collected from buildings through
smart energy meters, sensors and other means. Another limitation observed in most of the
research is that the models developed will not be applicable if the climatic zone in which the
building is situated changes. Accurate prediction of a building’s energy requirements needs to
consider building occupancy as an important factor. However, in some of the studies mentioned,
the occupancy has been assumed to be constant or considered less significant. It is important
to keep in mind that in most buildings the cooling equipment is turned on everyday for a fixed
duration irrespective of building occupancy. The building energy manager has an important
role to play in this aspect.
Various techniques like ANN, SVM, Chi Squared Automatic Detection Error, linear
regression, ensemble models, IRLS, RF etc. have been used for the purpose of prediction of
heating and cooling loads of buildings. But results show that models using ANN are able to
address the issue of accurate forecasting of building energy requirements more effectively. It can

124 Volume 14, Number 3

be concluded that ANN is indeed a powerful tool used in the study of energy performance of
buildings. The literature study in this area also signifies the usefulness of ANN in the energy
analysis studies of buildings. Table 3 summarises the results obtained in each paper, their merits
and demerits, and a scope for improvement in future work. This work does not intend to
undermine the importance of statistical learning techniques, but based on results, suggests
ANN to be a superior method to address nonlinearities present in building energy data and
thus be able to provide a better platform for evaluation, estimation and prediction of building
energy requirements.

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

TABLE 3. Results of methods.

Research
Reference no. Outputs Merits Demerits Future improvement
[12] HL and CL The input variables Glazing area, Climate and
chosen are same although an occupancy are
as in previous important variable assumed to
literature and helps is not the most be constant.
in comparing results correlated with Use of simulated data
presented in other neither HL nor CL
papers considering
the same variables.
Benefits of RF in
energy studies is
demonstrated
[13] HL and CL Computation 1. Use of default 1. Use of
time taken by the settings in the evolutionary
suggested models is models, both for computing and
very less the single and swarm intelligence
ensemble models techniques for
2. Models proposed optimization of
in the paper are parameters in the
only applicable models (single
to the twelve and ensemble) for
building types achieving 1. 1.
which are Higher accuracy
simulated in prediction of
HL and CL of
the buildings
2. Testing of model
using actual
datasets
[14] HL and CL Accuracy of proposed Use of simulation Testing of model
EMARS model data using actual datasets
higher than BPNN,
RBFNN, CART and
SVM

(conitnues)

Journal of Green Building 125

TABLE 3. (Continued)

Research
Reference no. Outputs Merits Demerits Future improvement
[15] CL Inclusion of zone Effect of occupancy Application of model
air temperature for on HL prediction not to various types of
prediction of HL significant commercial buildings
in different climatic
regions
[16] HL and CL Performance of Use of simulated data Model can be tested

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

classical ANN is using actual datasets
enhanced by using
adaptive ANN
[17] CL High variability in Only three Generalize the
data is tackled by institutional model for a wider
division of data into buildings considered range of institutional
classes and class buildings
numbers
[18] CL Computation time is Computation time Control strategies to
reduced increases with feature regulate operations in
numbers and is response to demand
maximum when the can be developed
RAW feature set is
used
[19] HL and CL 1.Use of Only data from five Can be extended
actual datasets large office buildings for analysing energy
2.Weather used usage of different
conditions taken building typologies
into consideration
3.Useful where no or
limited sensor data
available for greybox
modelling
[20] HL and CL Significance of 1. Other occupancy 1. Comparison
occupancy in data like lecture of the models
prediction of HL and schedules for a different
CL not considered test dataset
2. Energy tariff data 2. Use of higher
not included range of historical
data
[21] HL and CL Actual data used Only one climatic Can be extended to
as test data for the region is considered other climatic regions
neural network
model

126 Volume 14, Number 3

ABBREVIATIONS

AI Artificial Intelligence
ANN Artificial Neural Network
CART Classification and Regression Tree
CL Cooling Load
GRNN General Regression Neural Networks
HL Heating Load

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

HVAC Heating Ventilation and Air Conditioning
IRLF Iteratively Reweighted Least Squares
MAE Mean Absolute Error
MAPE Mean Absolute Percentage Error
MRE Mean Relative Error
MSE Mean Square Error
RBFNN Radial Basis Function Neural Network
RF Random Forest
RMSE Root Mean Square Error
SVM Support Vector Machine
SVR Support Vector Regression

5. REFERENCES
[1] Urge-Vorsatz D, Cabeza LF, Serrano S, Barreneche C, Petrichenko K. Heating and cooling energy trends
and drivers in buildings. Renew Sust Energy Rev 2015;41:85–98.
[2] Perez-Lombard L, Ortiz J, Pout C. A review on buildings energy consumption information. Energy Build
2008;40:394–8.
[3] Li XW, Wen J. Review of building energy modeling for control and operation, Renew Sustain Energy Rev
2014;37:517–37.
[4] W.G. Cai, Y. Wu, Y. Zhong, H. Ren, China building energy consumption: situation, challenges and cor-
responding measures, Energy Policy 37 (6) (2009) 2054–2059.
[5] G. Platt, J. Li, R. Li, G. Poulton, G. James, J. Wall, Adaptive HVAC zone modelling for sustainable build-
ings, Energy and Buildings 42 (2010) 412–421.
[6] R. Yao, B. Li, K. Steemers, Energy policy and standard for built environment in China, Renewable Energy
30 (2005) 1973–1988.
[7] A. Yezioro, B. Dong, F. Leite, An applied artificial intelligence approach towards assessing building perfor-
mance simulation tools, Energy and Buildings 40 (2008) 612–620.
[8] S. Sholahudin, Hwataik Han, Heating load predictions using the static neural networks method, Int. J.
Technol. 6 (2015) 946–953.
[9] Hamid R. Khosravani, María Del Mar Castilla, Manuel Berenguel, Antonio E. Ruano, Pedro M. Ferreira, A
comparison of energy consumption prediction models based on neural networks of a bioclimatic building,
Energies 9 (2016) 57.

Journal of Green Building 127

[10] Chengdong Li, Zixiang Ding, Dongbin Zhao, Jianqiang Yi, Guiqing Zhang, Building energy consumption
prediction: an extreme deep learning approach, Energies 10 (2017) 1525.
[11] S. Aman, Improving Energy Use Forecast for Campus Micro-grids using Indirect Indicators, in: 2011 IEEE
11th International Conference Data MiningWorkshops (ICDMW), Vancouver, BC, Canada, 2010, pp.
389–397.
[12] A. Tsanas, A. Xifara “Accurate quantitative estimation of energy performance of residential buildings using
statistical machine learning tools. Energy and Buildings. Volume 49, 2012, pp. 560–567.
[13] Jui-Sheng Chou, Dac-Khuong Bui, Modeling heating and cooling loads by artificial intelligence for energy-
efficient building design, Energy and Buildings 82 (2014) 437–446.
[14] M. Cheng, M. Cao, “Accurately predicting building energy performance using evolutionary multivariate
adaptive regression splines,” Applied Soft Computing 22 (2014) 178–188.
[15] Dimitrios-Stavros Kapetanakis, Eleni Mangina, El Hassan Ridouane, Konstantinos Kouramas, Donal Finn,

Downloaded from http://meridian.allenpress.com/jgb/article-pdf/14/3/115/2439379/i1943-4618-14-3-115.pdf by Vietnam user on 25 June 2024

COMPARISON OF PREDICTIVE MODELS FOR FORECASTING BUILDING HEATING LOADS,
14th Conference of International Building Performance Simulation Association, Hyderabad, India, Dec.
7–9, 2015.
[16] Yusuf Sonmez; Uğur Guvenc; H. Tolga Kahraman; Cemal Yilmaz, “A comperative study on novel machine
learning algorithms for estimation of energy performance of residential buildings,” Smart Grid Congress and
Fair (ICSG), 2015 3rd International Istanbul.
[17] Chirag Deb, Lee Siew Eang, Junjing Yang, Mattheos Santamouris, Forecasting diurnal cooling energy load
for institutional buildings using Artificial Neural Networks, Energy and Buildings 121 (2016) 284–297.
[18] Cheng Fan, Fu Xiao, Yang Zhao, A short-term building cooling load prediction method using deep learning
algorithms, Applied Energy 195 (2017) 222–233.
[19] Burak Gunay, Weiming Shen, Guy Newsham, Inverse black box modeling of the heating and cooling load
in office buildings, Energy and Buildings 142 (2017) 200–210.
[20] H. Pombeiro, R. Santos, P. Careirra, C. Silva, J. Soussa,” Comparative assessment of low-complexity models
to predict electricity consumption in an institutional building: Linear regression vs. fuzzy modeling vs. neural
networks,” Energy and Buildings 146 (2017) 141–151.
[21] Alaoui Sosse Jihad, Mohamed Tahiri, “Forecasting the heating and cooling load of residential buildings by
using a learning algorithm “gradient descent,” Morocco,” Case studies in Thermal Engineering 12 (2018)
85–93.
[22] S. Sholahudin, Hwataik Han, Heating load predictions using the static neural networks method, Int. J.
Technol. 6 (2015) 946–953.
[23] Hamid R. Khosravani, María Del Mar Castilla, Manuel Berenguel, Antonio E. Ruano, Pedro M. Ferreira, A
comparison of energy consumption prediction models based on neural networks of a bioclimatic building,
Energies 9 (2016) 57.
[23] Chengdong Li, Zixiang Ding, Dongbin Zhao, Jianqiang Yi, Guiqing Zhang, Building energy consumption
prediction: an extreme deep learning approach, Energies 10 (2017) 1525.
[24] J.W. Taylor, Short-term load forecasting with exponentially weighted methods, IEEE Transactions on Power
Systems 27 (1) (2012) 458–464.
[25] H.B. Gunay, W. O’Brien, I. Beausoleil-Morrison, Control-oriented inverse modeling of the thermal charac-
teristics in an office, Science and Technology for the Built Environment (2016) 1–20.
[26] P. Bacher, H. Madsen, Identifying suitable models for the heat dynamics of buildings, Energy and Buildings
43 (7) (2011) 1511–1522.
[27] D. Lazos, A.B. Sproul, M. Kay, Optimisation of energy management in commercial buildings with weather
forecasting inputs: a review, Renewable and Sustainable Energy Reviews 39 (2014) 587–603.
[28] C. Martani, D. Lee, P. Robinson, R. Britter, C. Ratti, ENERNET: studying the dynamic relationship between
building occupancy and energy consumption, Energy Build. 47 (2012) 584L 591.
[29] I. Berry, J. Conner, A. Braun, I. Berry, T. Roman, L. Adams, J.P. Pasnak, J. Conner, R. Scheck, A. Braun,
The Cacti Manual, The Cacti Group, 2012.

128 Volume 14, Number 3