ISSN (Online) : 2454 -7190 ISSN (Print) 0973-8975

J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

Prediction of Heating and Cooling Load to improve

Energy Efficiency of Buildings Using Machine Learning
Techniques
1
Srihari J, 2Santhi B

Sastra Deemed University, India

E-mail : 1srihari2761@gmail.com, 2shanthi@cse.sastra.edu

https://doi.org/10.26782/jmcms.2018.12.00008

Abstract
Global warming has been a severe threat to humanityand greenhouse gases
emitted from power plants is one of the major causes of global warming. In this
paper, we use machine learning to incorporate energy efficiency techniques to
buildings by predicting the Heating and Cooling Load using eight input
features.Heating load is the amount of heat per unit time that a building needs to
maintain the temperature at an established level whereas Cooling load is the amount
of heat per unit time that must be removed. Heating, cooling, and ventilation systems
are used to handle heating and cooling load. We train four regression (linear
regression, Lasso, Ridge, and Elastic-Net) and three gradient boosting models
(GBM, XGBoost, and LightGBM) and test them to compare their performance using
768 rows of data of residential buildings. We observe that the gradient boosting
models perform significantly better than the standard regression models for both
Heating Load and Cooling Load. XGBoost achieves the highest R-squared score of
0.99 for Heating Load and 0.99 for Cooling Load. From the results of this study, we
conclude that machine learning techniques can predict Heating Load and Cooling
Load with high accuracy. The obtained Heating load and cooling load values can be
used to install efficient heating, cooling and ventilation systems and thus reduce both
energy consumption and money.

Keywords: Energy efficiency · Heating Load · Cooling Load · Machine Learning

I. Introduction
Over the past 50 years, the global temperature has been rapidly increasing.
This phenomenon, called global warming, occurs when emission of CO2 and other
air pollutants known as greenhouse gases collect in the atmosphere and absorb the
sunlight bouncing off from earth’s surface, thus trapping the heat which otherwise
would escape into space. Majority of the greenhouse gases are released when fossil
fuels are burnt in power plants to produce electricity. Thus a significant step towards

97
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

reducing the greenhouse gases is to reduce the energy consumption, ultimately

decreasing the energy produced at power plants and thus reducing the emission of
greenhouse gases.
An average household spends more than $2,200, with nearly 50% on going to
heating and cooling costs. By using less energy at home, we not only reduce
emissions that contribute to global warming but also save much money, thus
improving the country’s economy. Energy efficiency is a strategy with an objective
of reducing the energy required for products and services. Energy efficiency policies
are becoming a crucial part of the global energy market. Many countries and
institutions consider energy efficiency with high priority since it significantly reduces
the energy consumption and thus slows down the energy consumption rate. It is the
key to ensure a safe and reliable energy system for the future. Reducing power,
although lacking an upfront capital, eventually minimise energy costs and may result
in saving financial value over a long run.
International Energy Agency is one such institution which has been working with
countries around the world to implement energy efficiency measure. Energy star is
also a program backed by the government that helps to protect the environment and
save money with energy efficient practices and products. Smart decisions about
heating, ventilating and air conditioning system (HVAC) are made by the U.S.
Environmental Protection Agency (EPA), helping us to save energy costs.
Buildings are one of the major fields where energy efficiency techniques can be
used. Energy-efficient typically consumes very little energy to provide services such
as heating, cooling, and ventilation. Buildings can be made energy-efficient in winter
by using high quality and necessary building materials to prevent heat loss. The
capacity of energy efficiency for buildings is relative to the location and the situation
of when it is applied. For example, countries with low temperatures require draught
proofing, installation of insulation, installation of double glazing windows while
countries with high temperatures require the installation of reflective and radiative
roofs. In this paper, we use machine learning techniques to predict the Heating and
Cooling Load of buildings. Heating load is the amount of heat per unit time that a
building needs to maintain the temperature at an established level whereas Cooling
load is the amount of heat per unit time that must be removed. Heating, cooling, and
ventilation systems are used to handle heating and cooling load.
With the predicted Heating and Cooling Load, we can install efficient heating,
cooling and ventilation systems accordingly and save money on energy spent. Also,
less energy spent implies fewer energy requirements from the power plants.
Therefore power plants consume fewer fossil fuels, emit fewer greenhouse gases and
ultimately help prevent global warming.
The rest of the paper is organized as follows:
1) Section II discusses previous attempts and existing techniques to estimate
energy consumption and energy efficiency of buildings. This extensive survey helps

98
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

to gain knowledge of how previous researches were performed and the results
obtained by them.
2) Section III describes the methodology of the experiment conducted. Different
features used in the dataset are listed and explained. Then, the algorithms used for
prediction of heating and cooling load are briefly explained.
3) Section IV contains information on how the experiment was performed. An
exploratory data analysis is first performed on the dataset. Then the algorithms are
implemented and evaluated using various evaluation metrics.
4) Section V analyzes the results obtained from the algorithms. All the
algorithms used are compared in this section, and their performance is evaluated.
II. Background
In the past, several attempts have been made to analyse and estimate the
energy efficiency of buildings. State of the art energy estimation methods uses
statistical data and default values. It utilises a generic approach to predict the energy
performance of buildings with an assumption of average reference values which may
vary according to different National laws and uses. [II] present a novel method for
estimation and optimisation of thermal load of buildings using LiDAR data. They
observed growth in buildings heat gain and reduction in thermal load. [XXIV] use the
prospected operation time and three derating factors to suggest an improvement in the
procedure for evaluation of lighting energy. [XV] use energy modelling technique to
analyse the highest energy consumption factor of medium rise commercial buildings.
They were able to come with designs that contribute towards energy savings with the
help of the study. These energy conservation measures contribute to energy saving
without affecting the thermal comfort. [XIV] investigate the techniques used to
reduce the energy consumption of commercial buildings by incorporating energy
conservation techniques during the design phase of the building. [XXVII] provide a
review of the energy performance of buildings by analysing the phase change
materials using to optimise building envelope and equipment.
Improvement in the energy performance of buildings results in improved energy
efficiency. [XXIII] implement a prediction system for indoor comfort conditions.
They use behaviour performance to improve energy efficiency and the lifetime of the
equipment, resulting in a reduction of system inefficiencies. [XVIII] created a
stochastic building energy model to provide energy consumption measures for
remodelling of existing building. [XVI] present an analysis of leading software
applications for sensor infrastructures concerning how metadata is handled. They
present their vision on the discovery, maintenance, and validation of metadata in
buildings to improve the energy performance of buildings. [XXXIII] analyse the
impact of the atrium on hotel buildings energy demands for space heating and
cooling. They create an imaginary model of an atrium type hotel building, with which
they perform numerical stimulation of energy performance and analyse the obtained
results. [XX] present a novel modelling approach to predict an individual’s thermal

99
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

preference based on learning from thermal control behaviour with PCS chairs. They
compare six machine learning algorithms for consistent data collection and prediction
accuracy. [XVII] used the assertion that each element of the building can impact on
energy performance to build a generic metamodel. The metamodel represents a fast
way to perform calculations with high accuracy and can be used as a basis for future
building regulations and to perform parametric studies.
Analyzing the amount of energy consumed also plays a significant role in
improving energy efficiency as an abnormal consumption implies inefficiency.
Building energy use can be predicted by engineering modelling approaches or using
Artificial Intelligence. [V] provide energy usage profiles of household appliances.
They observe from that average appearance load profiles have equal daily
distributions compared to those obtained from current energy stimulation software for
buildings. [VII] presents a novel model called Domestic Energy and Carbon Model
for prediction of energy consumption of english housing stock. The DECM tool is
used for abetment of creation of energy policies at both local and national levels. [III]
identify 333 failures in energy and environmental performance of buildings and
categorise them in 51 separate modes. [I] design a tariff structure to deliver optimum
value and cost saving which results in huge savings, both economically and
environmentally. [XI] use an event-based approach calibrated from measured data to
identify residential occupancy and electrical consumption. The method proposed by
[XI] performs better than models designed from time-use survey activity diaries and
per-timestep probability models.
While engineering approach requires solving physical equations which represent
the thermal performance of components of buildings, the AI-based machine learning
approach uses historical data to predict future performance. With the growing
accretion of publicly available energy data, they are more accurate and overcome the
limitations of conventional techniques. [XXII] use an artificial neural network model
to estimate the cooling load of a building. They explain the importance of building
occupancy rate in cooling load prediction and discuss the difficulty in obtaining
building-related data. [XXI] use real-time feedback collected using novel mobile
sensing technology to generate data-driven predictive models to evaluate the thermal
comfort of occupants. [VIII] Propose an SVM model to predict thermal demands
using different skin temperatures. [XIII] propose a strategy to determine the time
ahead of the prediction mode. The predictive models are developed and analysed
using machine learning techniques such as multiple linear regression, support vector
regression, BP neural network, and Extreme Learning Machine. [IX] test different
machine learning approaches for predicting Energy Use Intensity for individual
energy end-uses and US commercial office buildings. They observe that Random
Forest and Support Vector Machine establish excellent accuracy and stability.
[XXVI] propose a stimulation model for prediction of energy consumption in
residential areas. They use the proposed model to predict energy consumption in
Osaka City and compare it with results obtained from statistical methods. [XXXII]
present a tool called REMA for prediction of energy consumption and determine

100
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

patterns in energy consumption after recent changes in the policy. [IV] propose a
uncomplicated and simple methodology for definition and evaluation of energy
strategies in different areas. The proposed methodology takes into account the input
from different stakeholders for development of low carbon settlements. [XII]
introduce a model for analysis of efficient energy strategies in a region. The proposed
model was found useful in rural, urban and educational planning proceedings.
[XXXIV] use an ensemble technique called Ensemble Bagging Trees to predict
hourly electricity demand of the test building using data from meteorological
systems. The ensemble method achieves high accuracy and alleviates instability
issues of other models. [XXXI] establish a statistical machine learning framework to
analyse the effect of eight input features on Heating and Cooling load. Their study
supports the usefulness of machine learning techniques to estimate heating and
cooling load. [XXV] use machine learning models to estimate commercial building
energy consumption from a small number of features. They observe that gradient
boosting regression models perform the best at commercial building energy
consumption. [XXX] uses gradient boosting for energy consumption of 410
commercial buildings. They observed that gradient boosting significantly improved
the R-squared and RMSE score compared to piecewise linear regression and random
forest algorithm. [XXIX] uses multiple non-linear regression for prediction CDHS’s
heating load and validating it using data obtained from continuous monitoring over 2
years. They use 4-step procedure to predict the energy demand and analyze the
results. [XXVIII] propose a simplified model by using the thermal mass, impact
shading and internal heat generation to predict the heating demand of a district. The
simplified model was validated against stimulations from a complex stimulation tool.

III. Methodology
III. a. Loading the data

The dataset obtained from UCI machine learning repository consists of 8

features (Relative Compactness (F1), Surface Area (F2), Wall Area (F3), Roof Area
(F4), Overall Height (F5), Orientation (F6), Glazing Area (F7) and Glazing Area
Distribution (F8) and 768 instances of data aiming to predict two realvalued output
variables - Heating Load (Y1) and Cooling Load (Y2). The dataset containing 768
instances of data are loaded into a dataframe of 10 columns and 768 rows.

III.b. Train-Test split

The dataframe comprising of 768 instances of data are randomly sampled
into train and test set with a ratio of 80% and 20% respectively. The train set contains
614 rows while the test set contains 154 rows. The train set is used to train the
algorithms mentioned below while the test set is used to evaluate the model.

101
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

III.c. Methods
We use train data to analyse seven models and then evaluate with test data.
The seven algorithms fall into two categories- regression and gradient boosting
technique.
Linear regression is one of the widely used techniques to predict continuous
variables. It assigns one scale factor to each input column called a coefficient. Apart
from coefficients for each input feature, one additional coefficient called bias
coefficient is included, giving the line/plane an additional degree of freedom. When a
coefficient of an input feature is high, then the input feature contributes more, that
has a high influence on the output variable. When it is low, say close to zero, it has
profound influence or does not contribute to the output variable. The following
equation represents the linear regression
(1)
Where Y is the output variable, X is the input features in the form of a vector and is
the coefficients corresponding to the input features in the form of a vector. The
coefficients are randomly initialised at the start and are optimised using gradient
descent.
LASSO regression is another technique similar to linear regression. The
difference between LASSO and linear regression is that LASSO penalises the
absolute size of regression coefficients. By penalising the coefficients, some of the
coefficients may be precisely zero, nullifying the effect of the corresponding feature.
This helps to perform automatic feature selection when dealing with highly correlated
features.
Ridge regression is similar to lasso, except it cannot dilute the value of
coefficients and hence automatic feature selection cannot be performed in Ridge
regression. Lasso uses L1 regularisation which imposes sparsity among the
coefficients, making the model more interpretable. Ridge uses L2 regularisation,
which limits the size of the coefficient vector.
Elastic Net or E-Net regression linearly combines the L1 regularisation of
Lasso and L2 regularisation of Ridge regression. E-Net overcomes the limitations of
Lasso by adding the quadratic part from E-Net, making the loss function convex.
Apart from regression techniques, gradient boosting is also used to predict
the output variables. Gradient boosting uses an ensemble of weak prediction models
such as decision trees in a stage-wise fashion. It allows optimisation of arbitrary loss
functions. A regression tree is fit on the negative gradient of the given loss function
each stage.
XGBoost [VI] and LightGBM [XX] are two variations of standard Gradient
Boosting Model (GBM). XGBoost uses regularisation to prevent overfitting, unlike
GBM. XGBoost implements parallel processing making it faster than GBM.
XGBoost splits the nodes in the tree up to max depth specified and then prunes the

102
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

tree backwards to remove splits beyond which there is no positive gain. XGBoost
contains an inbuilt crossvalidation mechanism to get an exact optimum number of
boosting iterations in a single run.
LightGBM is similar to XGBoost, but it is faster and efficient since it
buckets continuous feature values into discrete bins. It uses leaf-wise split approach
instead of level-wise approach to produce more complex trees which help to achieve
an excellent accuracy but also making it prone to overfitting.

IV. Experimental setup

IV.a. Dataset
The dataset is obtained from [X] provided by Tsanas and
Xifara[XXXII]comprises 768 samples of 12 different residential building shapes
made of 18 elements stimulated in Ecotect, aiming to predict two output variables.
The 768 samples are characterised by eight input features and two output features
listed in Table 1 .The mathematical notation and statistical measures of the eight
features and two output variables are given in Table 1. Using the calculated statistical
values, we plot boxplot for each of the variables in Figure 1.

Table 1. Description of input and output features

X1 X2 X3 X4 X5 X6 X7 X8 Y1 Y2

Count 768 768 768 768 768 768 768 768 768 768
Mean 0.76 671.71 318.50 176.60 5.25 3.50 0.23 2.81 22.31 24.59
STD 0.11 88.09 43.63 45.17 1.75 1.12 0.13 1.55 10.09 9.51
Min 0.62 514.50 245.00 110.25 3.50 2.00 0.00 0.00 6.01 10.90
25% 0.68 606.38 294.00 140.88 3.50 2.75 0.10 1.75 12.99 15.62
50% 0.75 673.75 318.50 183.75 5.25 3.50 0.25 3.00 18.95 22.08
75% 0.83 741.12 343.00 220.50 7.00 4.25 0.40 4.00 31.67 33.13
Max 0.98 808.50 416.50 220.50 7.00 5.00 0.40 5.00 43.10 48.03

IV.b. Exploratory Data Analysis

Exploring the data is an essential step before we work on data, as some
important inferences can be captured from the data in this step. First, we perform
numerical computations on the data which is given in Table 1. The count column
specifies the number of values for each attribute. Since all attributes contain 768
values, we can assume there are no missing values. After further analysis, we can see

103
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

there are no NaN values. Hence there is no need for cleaning the data. Mean,
standard deviation, minimum value, maximum value, 25th percentile, 50th percentile,
75th percentile are also calculated and recorded in Table 1.

Fig.1. Boxplot for input and output features

We then plot scatter plots between the input and output variables in Figure 4.
From the scatter plots, we can observe that some features are linearly related. To
further analyse the correlation between different variables, we plot the Correlational
Matrix in Figure 3 which contains the correlation coefficients between the input and
output variables. We can observe that some variables are highly correlational, but we
do not remove any features since there are only eight features and removing features
will make the models underfit.
The univariate distribution of input and output variables with kernel density
estimate and the normal distribution is plotted in Figure 2. From the distribution
plots, we can observe that the data is non-Gaussian since it does not have a unimodal
peak. Also, there is not much skewness in the data. Thus we do not perform data
transformation.

IV.c. Evaluation Metrics

To evaluate the model and check how well the model performs, we use four
evaluation metrics specified in Table 2. In Table 2, ypr is the vector containingthe
output variable predicted by the model, is the vector containing theoriginal output
variables, represents the mean of ytr, ∑ represents the numberof samples and n
represents the sum of elements in the vector.

104
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

Table 2. Performance metrics

Evaluation Metric Name Mathematical Mathematical Formulae

Representation
Coeffcient of determination

Mean Absolute Error

Mean Squared Error

Root Mean Squared Error

Fig.2. Univariate distribution graph for input and output features

105
J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

Fig.3. Correlation matrix for input and output features

106
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

Fig.4. Scatter plot for input and output features

V. Results and Discussion

The seven models are trained using the training set, and then test set is used
to predict output values. The models are evaluated with the original output values and
predicted output values using the evaluation metrics explained above and recorded in
Table 3 and 4.
The R2 value is a statistical measure of the response variable variation that is
explained by a linear model. R2 lies between 0 and 1. A value close to 1 demonstrates
that the model explains all the variability of the predicted data around its mean. From
Table 3 and 4, we can observe that XGBoost has the highest R2 value followed by
GBM and LightGBM. For heating load, GBM has a slightly better R2 score than
LightGBM. For cooling load, LightGBM has a slightly better R2 score than GBM. All
three boosting algorithms show superior R2 score over the four standard regression

107
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

algorithms. Among the four regression algorithms, E-Net has the maximum R2 score.
However, we cannot conclude with just the R2 value as a high R2 value does not
guarantee that the model has agood fit, so we also use other evaluation metrics to
evaluate the model.
ErrorMAE, ErrorMSE, and ErrorRMSEare used to measure the performance of the
model. They measure the average magnitude of the error between the predicted and
original output variables. They can range from 0 to ∞ , and they consider the
direction of the errors, which means that exchanging the vector of predicted and
actual values do not yield different results for these metrics. Since the evaluation
metrics mentioned in Table 2 measure the errors, if the model has a lower error, then
the model performs better.
For the training set, XGBoost has the lowest ErrorMAEfor heating load
followed by GBM and LightGBM. Among the regression models, E-Net has the
lowest ErrorMAEfollowed by Linear Regression, Ridge and Lasso regression. The
cooling load variable also obtains similar results.
However, on the test set, GBM has lower ErrorMAEon the training set instead
of XGBoost. To explain this irregularity, we use ErrorMSEand ErrorRMSEsince they
give relatively high weight to significant errors, which is undesirable in our problem.
The more significant importance of large errors is due to the square term in
ErrorMSEand ErrorRMSE.
For heating load, XGB has lowest ErrorMSE and ErrorRMSE values
followed by GBM and LightGBM. Among the regression algorithms, E-Net has
lowest ErrorMSE and ErrorRMSE values followed by Linear Regression, Lasso and
Ridge Regression. Cooling Load also obtains similar results except GBM and
LightGBM where LightGBM has better ErrorMSE and ErrorRMSE values than
GBM. For both output variables, the Boosting Algorithms perform significantly
better than standard regression algorithms in both training and test sets.
We compare and contrast our techniques with previous related work to
analyze the performance of the proposed technique compared to previous state of the
art techniques. The Error value and R2 score of the test set recorded in Table 3 and 4
is used to benchmark our result.[XXXII] use Iteratively Reweighted Least Squares
(IRLS) and Random Forests to predict
Heating and Cooling Load using eight input features. We compare the results
of Random Forests which outperforms IRLS from [XXXII], with the results of
XGBoost from Table 3 and 4 and present it in Table 5.From Table 5, it is inferred
that the proposed XGBoost model outperforms Random Forest used in [XXXII] with
a minimum margin of 0.13 for ErrorMAEscore and 0.46 for ErrorMSEscore.

108
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

Table 3. Evaluation metric results for Heating Load

Evaluation Linear Lasso Ridge E-Net GBM XGBoost Light

Metric regression regressio regressio regression GBM
n n
R2(train) 0.92 0.92 0.92 0.93 0.99 0.99 0.99
ErrorMAE(train) 2.06 2.03 2.03 1.93 0.16 0.15 0.23
ErrorMSE(train) 7.81 8.10 8.11 7.42 0.07 0.05 0.12
ErrorRMSE(train) 2.80 2.85 2.85 2.72 0.27 0.21 0.34
R2(test) 0.92 0.91 0.91 0.92 0.99 0.99 0.99
ErrorMAE(test) 2.19 2.18 2.18 2.02 0.27 0.28 0.37
ErrorMSE(test) 8.82 8.95 8.96 8.10 0.14 0.13 0.37
ErrorRMSE(test) 2.97 2.99 2.99 2.85 0.38 0.36 0.61

Table 4. Evaluation metric results for Cooling Load

Evaluation Metric Linear Lasso Ridge E-Net GBM XGBoost LightG

regression regression regressio regression BM
n
R2(train) 0.89 0.89 0.89 0.89 0.99 0.99 0.99
ErrorMAE(train) 2.25 2.27 2.27 2.22 0.41 0.27 0.46
ErrorMSE(train) 10.19 10.29 10.29 9.94 0.48 0.13 0.40
ErrorRMSE(train) 3.19 3.21 3.21 3.15 0.69 0.36 0.63
R2(test) 0.89 0.89 0.89 0.89 0.98 0.99 0.99
ErrorMAE(test) 2.16 2.20 2.20 2.14 0.73 0.64 0.73
ErrorMSE(test) 9.91 9.98 9.98 9.69 1.55 1.10 1.26
ErrorRMSE(test) 3.15 3.16 3.16 3.11 1.24 1.05 1.12

Table 5. Comparison of results of Tsanas and Xifara’s statistical machine learning

technique [XXXII] with proposed XGBoost regressor

Output Variable ErrorMAE ErrorMSE

Tsanas and Xifara[XXXII] Y1 0.51 ± 0.11 1.03 ± 0.54

Proposed XGBoost regressor Y1 0.28 0.13
Tsanas and Xifara[XXXII] Y2 1.42 ± 0.25 6.59 ± 1.56
Proposed XGBoost regressor Y2 0.64 1.10

109
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

VI. Conclusion
In this paper, we used machine learning techniques to predict the Heating and
Cooling Load with the help of 8 input variables. We trained four regression (linear
regression, Lasso, Ridge, and Elastic-Net) and three gradient boosting models (GBM,
XGBoost, and LightGBM) and tested them to compare their performance using 768
rows of data of residential buildings. We compared the models using four evaluation
metrics. We observed that the gradient boosting models perform notably better than
the standard regression models. Out of the seven models we used, XGBoost
accurately predicts the Heating and Cooling Load with a R-squared score of 0.99 for
Heating Load and 0.99 for Cooling Load.
The future scope of this paper includes collection and aggregation of more
data to train accurate models. The proposed machine learning models can easily be
implemented in real life since it predicts the Heating Load and Cooling Load with
high accuracy. The predicted Heating Load and Cooling Load values can be used to
design efficient heating, cooling and ventilation equipments to save energy and thus
eventually prevent global warming.

References

I. Al Fardan, A. S., Al Gahtani, K. S., and Asif, M. (2017). Demand side

management solution through new tariff structure to minimize excessive load
growth and improve system load factor by improving commercial buildings
energy performance in Saudi Arabia. 2017 5th IEEE International
Conference on Smart Energy Grid Engineering, SEGE 2017, pages 302–308.
II. Bizjak, M., Zalik, B.,ˇ Stumberger, G., and Lukaˇc, N. (2018). Estimation
andˇ optimisation of buildings’ thermal load using LiDAR data. Building and
Environment, 128:12–21.
III. Borgstein, E. H., Lamberts, R., and Hensen, J. L. (2018). Mapping failures in
energy and environmental performance of buildings. Energy and Buildings,
158:476–485.
IV. Caputo, P., Costa, G., and Ferrari, S. (2013). A supporting method for
defining energy strategies in the building sector at urban scale. Energy
Policy, 55:261 – 270. Special section: Long Run Transitions to Sustainable
Economic Structures in the European Union and Beyond.
V. Cetin, K. S., Tabares-Velasco, P. C., and Novoselac, A. (2014). Appliance
daily energy use in new residential buildings: Use profiles and variation in
time-ofuse. Energy and Buildings, 84:716–726.

110
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

VI. Chen, T. and Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting

System.
VII. Cheng, V. and Steemers, K. (2011). Modelling domestic energy consumption
at district scale: A tool to support national and local energy policies.
Environmental Modelling Software, 26(10):1186 – 1198.
VIII. Dai, C., Zhang, H., Arens, E., and Lian, Z. (2017). Machine learning
approaches to predict thermal demands using skin temperatures: Steady-state
conditions. Building and Environment, 114:1–10.
IX. Deng, H., Fannon, D., and Eckelman, M. J. (2018). Predictive modeling for
US commercial building energy use: A comparison of existing statistical and
machine learning algorithms using CBECS microdata. Energy and Buildings,
163:34–43.
X. Dheeru, D. and KarraTaniskidou, E. (2017). UCI machine learning
repository.
XI. Flett, G. and Kelly, N. (2017). A disaggregated, probabilistic, high resolution
method for assessment of domestic occupancy and electrical demand. Energy
and Buildings, 140:171–187.
XII. Fonseca, J. A. and Schlueter, A. (2015). Integrated model for
characterization of spatiotemporal building energy consumption patterns in
neighborhoods and city districts. Applied Energy, 142:247 – 265.
XIII. Guo, Y., Li, G., Chen, H., Wang, J., and Huang, Y. (2017). A thermal
response time ahead energy demand prediction strategy for building heating
system using machine learning methods. Energy Procedia, 142:1003–1008.
XIV. Gupta, N. and Shet, H. N. (2016). Analysis of Measures to Improve Energy
XV. Performance of a Commercial Building by Energy Modeling. 2016
Online International Conference on Green Engineering and Technologies
(IC-GET) Analysis, pages 1–4.
XVI. Hamid, M. F. A., Ramli, N. A., and Syawal Nik Mohd Kamal, N. M. F.
(2017). An analysis of energy performance of a commercial building using
energy modeling. In 2017 IEEE Conference on Energy Conversion
(CENCON), pages 105–110. IEEE.
XVII. Holmegaard, E., Johansen, A., and Kjærgaard, M. B. (2016). Towards a
metadata discovery, maintenance and validation process to support
applications that improve the energy performance of buildings. 2016 IEEE
International Conference on Pervasive Computing and Communication
Workshops, PerCom Workshops 2016.
XVIII. Jaffal, I. and Inard, C. (2017). A metamodel for building energy
performance. Energy and Buildings, 151:501–510.

111
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

XIX. Jeong, Y.-k., Kim, T., Nam, H.-S., and Lee, I.-w. (2016). Implementation of
energy performance assessment system for existing building. 2016
International Conference on Information and Communication Technology
Convergence (ICTC), (20142010102370):393–395.
XX. Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., and Liu,
T.-Y. (2017). Lightgbm: A highly efficient gradient boosting decision tree. In
Guyon, I., Luxburg, U. V., Bengio, S., Wallach, H., Fergus, R.,
Vishwanathan, S., and Garnett, R., editors, Advances in Neural Information
Processing Systems 30, pages 3146–3154. Curran Associates, Inc.
XXI. Kim, J., Zhou, Y., Schiavon, S., Raftery, P., and Brager, G. (2018). Personal
comfort models: Predicting individuals’ thermal preference using occupant
heating and cooling behavior and machine learning. Building and
Environment, 129:96–106.
XXII. Konis, K. and Annavaram, M. (2017). The Occupant Mobile Gateway: A
participatory sensing and machine-learning approach for occupant-aware
energy management. Building and Environment, 118:1–13.
XXIII. Kwok, S. S. K., Yuen, R. K. K., and Lee, E. W. M. (2011). An intelligent
approach to assessing the effect of building occupancy on building cooling
load prediction. Building and Environment, 46(8):1681–1690.
XXIV. Onose, B.-a. (2016). Control optimization for increasing energy performance
of existing buildings. 2016 Eleventh International Conference on Ecological
Vehicles and Renewable Energies (EVER), pages 1–4.
XXV. Parise, G., Martirano, L., and Parise, L. (2014). Energy performance of
buildings: An useful procedure to estimate the impact of the lighting control
systems. Conference Record - Industrial and Commercial Power Systems
Technical Conference, pages 1–7.
XXVI. Robinson, C., Dilkina, B., Hubbs, J., Zhang, W., Guhathakurta, S., Brown,
M. A., and Pendyala, R. M. (2017). Machine learning approaches for
estimating commercial building energy consumption. Applied Energy,
208(May):889– 904.
XXVII. Shimoda, Y., Fujii, T., Morikawa, T., and Mizuno, M. (2004). Residential
enduse energy simulation at city scale. Building and Environment, 39(8):959
– 967. Building Simulation for Better Building Design.
XXVIII. Song, M., Niu, F., Mao, N., Hu, Y., and Deng, S. (2018). Review on building
energy performance improvement using phase change materials. Energy and
Buildings, 158:776–793.

112
 J.Mech.Cont.& Math. Sci., Vol.-13, No.-5, November-December (2018) Pages 97-113

XXIX. Talebi, B., Haghighat, F., and Mirzaei, P. A. (2017). Simplified model to
predict the thermal demand profile of districts. Energy and Buildings,
145:213 – 225.
XXX. Talebi, B., Haghighat, F., Tuohy, P., and Mirzaei, P. A. (2018). Validation of
a community district energy system model using field measured data.
Energy, 144:694 – 706.
XXXI. Touzani, S., Granderson, J., and Fernandes, S. (2018). Gradient boosting
machine for modeling the energy consumption of commercial buildings.
Energy and Buildings, 158:1533–1543.
XXXII. Tsanas, A. and Xifara, A. (2012). Accurate quantitative estimation of energy
performance of residential buildings using statistical machine learning tools.
Energy and Buildings, 49:560–567.
XXXIII. Tuominen, P., Holopainen, R., Eskola, L., Jokisalo, J., and Airaksinen, M.
(2014). Calculation method and tool for assessing energy consumption in the
building stock. Building and Environment, 75:153 – 160.
XXXIV. Vujoˇsevi´c, M. and Krsti´c-Furundˇzi´c, A. (2017). The influence of atrium
on energy performance of hotel building. Energy and Buildings, 156:140–
150.
XXXV. Wang, Z., Wang, Y., and Srinivasan, R. S. (2018). A novel ensemble
learning approach to support building energy use prediction. Energy and
Buildings, 159:109–122.

113