Hindawi

Computational Intelligence and Neuroscience

Volume 2021, Article ID 6028573, 12 pages
https://doi.org/10.1155/2021/6028573

Research Article
Forecasting Time-Series Energy Data in Buildings Using an
Additive Artificial Intelligence Model for Improving
Energy Efficiency

Ngoc-Son Truong, Ngoc-Tri Ngo , and Anh-Duc Pham

Faculty of Project Management, The University of Danang-University of Science and Technology, 54 Nguyen Luong Bang,
Danang, Vietnam

Correspondence should be addressed to Ngoc-Tri Ngo; [email protected]

Received 20 May 2021; Revised 6 July 2021; Accepted 20 July 2021; Published 27 July 2021

Academic Editor: Yu-Ting Bai

Copyright © 2021 Ngoc-Son Truong et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
Building energy efficiency is important because buildings consume a significant energy amount. The study proposed additive
artificial neural networks (AANNs) for predicting energy use in residential buildings. A dataset in hourly resolution was used to
evaluate the AANNs model, which was collected from a residential building with a solar photovoltaic system. The proposed
AANNs model achieved good predictive accuracy with 14.04% in mean absolute percentage error (MAPE) and 111.98 Watt-hour
in the mean absolute error (MAE). Compared to the support vector regression (SVR), the AANNs model can significantly
improve the accuracy which was 103.75% in MAPE. Compared to the ANNs model, accuracy improvement percentage by the
AANNs model was 4.6% in MAPE. The AANNs model was the most effective forecasting model among the investigated models in
predicting energy consumption, which provides building managers with a useful tool to improve energy efficiency in buildings.

1. Introduction Various studies have been conducted to improve

building energy performance [2–5]. An accurate forecast of
National development, urbanization, and population growth the building energy use is a vital issue in smart building
require a growing energy demand. Buildings account for re- applications. Building energy prediction is typically per-
markable energy consumption during their operational stages formed by using engineering-based methods and artificial
and are responsible for carbon emissions and global warming. intelligence (AI) approaches. Because the engineering ap-
Energy performance in buildings is of prime importance all proach applies thermodynamic equations to predict energy
over the world. Buildings should be designed for occupant’s consumption in buildings, they are time-consuming and
comfort while consuming less energy. Thus, energy efficiency is require a high level of expertise to customize and set thermal
one of the most concerning topics among academic researchers parameters for energy performance analysis. To perform
and decision-makers in the energy sector. It plays a remarkable energy prediction, the engineering method requires detailed
role in targeting a low-carbon economy [1]. National gov- information on the building envelope, thermal properties of
ernments have also recognized the benefits of efficient uses of construction layers and windows, and the heating, venti-
energy in the building sector. The efficient use of energy in lation, air-conditioning system.
buildings strongly affects the building’s capability to meet the The AI-based method infers future energy consumption
building green certificates in the green building rating system to profiles in buildings using historical data [6]. The advantage
reduce carbon emission and greenhouse effects. Thus, energy of the AI-based approach lies in its learning capability to
usage prediction in buildings is necessary for energy planning, model the relationship between historical data and future
management, and conservation. data. The AI-based prediction model requires historical data
2 Computational Intelligence and Neuroscience

instead of much detailed building information like the paying special attention to the interpretability. Recently,
engineering methods. It does not require users to poses a expert systems, mainly developed by means of linguistic
deep knowledge of the thermodynamic behavior of build- fuzzy rule-based systems, allow us to deal with the system
ings. Some studies have proposed AI models for solving the modeling with good interpretability [14]. However, these
prediction of building energy performance. For example, models have strong dependency on an expert and often
Song et al. (2020) presented the evolutionary model con- cannot generate good accuracy. Therefore, combination
struction for predicting electricity data in smart buildings models, based on the popular methods, expert systems, and
[7]. Wang and Srinivasan (2017) presented an in-deep re- other techniques, are proposed to satisfy both high accurate
view of AI-based prediction models for predicting energy level and interpretability.
use in buildings with special attention on ensemble [6]. Although ML models have been applied in the building
Huang et al. (2021) applied a deep learning method in energy domain and yielded good forecasting accuracy, the
developing an energy management system [8]. Jahani et al. improvement of their performance in energy prediction is
(2020) developed a prediction model by integrating a genetic still necessary. In addition, the energy consumption in the
algorithm and numerical moment matching method to residential buildings equipped with the solar photovoltaic
predict energy consumption in residential buildings [9]. system has not been investigated in the literature. Therefore,
Analyzing energy structure and electricity use behavior this study proposed the additive artificial neural networks
is important to propose an energy-efficient policy in nations (AANNs) that can accurately predict energy consumption in
[10]. Forecasting results of energy consumption in buildings residential buildings with the renewable energy system. The
is the basis for optimizing building performance and re- contributions of this study include (1) collection of building
ducing energy costs [11]. AI and machine learning (ML) energy consumption profiles in hourly resolution and their
models have been used in the building energy domain associated weather data, (2) investigation of the potential
[11–15]. A model can predict future data or generate new power of artificial intelligence techniques in predicting fu-
insights based on learning from historical data. ML models ture building energy consumption, and (3) development of
have been used to forecast thermal demands using skin the effectiveness and capability of the AANNs in the pre-
temperatures [12]. Chou and Ngo (2016) analyzed time- diction of building energy consumption.
series energy data by integrating an ML and an optimization The rest of the study is organized as follows. Section 2
algorithm to identify building energy data patterns [4]. contains the literature review. Section 3 presents the
Among various ML models, artificial neural networks methodology. Section 4 contains the results and discussion.
(ANNs), support vector regression (SVR), and multiple Section 5 shows the conclusions.
linear regression are popular [6]. Ganguly et al. (2020)
applied the ANNs model to forecast building energy use in a 2. Literature Review
historical art gallery [16]. Saleh et al. (2019) evaluated the
performance of ML models in predicting heating and Buildings are responsible for about 30% of the total energy
cooling loads in buildings [17]. Their investigated ML consumption. Energy consumption prediction in buildings
models included ANNs, support vector regression (SVR), is imperative in energy management and conservation be-
Gaussian process (GP), random forest, and gradient boosted cause it facilitates a process to assess energy efficiency,
regression trees (GBRT). Their experimental result revealed perform commissioning, and detect and diagnose building
that the GBRT obtained the best performance in terms of the system anomalies [6]. Energy performance in buildings is
root-mean-square error (RMSE) value. They also concluded affected by various uncertainty factors such as heating
that the ANNs model was the best fit for complex datasets. ventilation air-conditioning (HVAC) systems, building
The computing time of the ANNs model is faster than other envelops characteristics, and building operating schedules.
investigated ML modes in their study [17]. Najafzadeh and Uncertainty analysis has been used widely in assessing
Oliveto (2020) applied the support vector machine (SVM), building energy performance [22–24] since the inherent
multivariate adaptive regression splines (MARS), and ran- uncertainty of occupant behavior, the building thermal
dom forest (RF) to predict the approach densimetric Froude property, HVAC system, and weather conditions. Four
number at the incipient motion of riprap stones that can perspectives for assessing building performance have been
protect rivers from erosion problems [18]. presented in [22] including uncertainty data sources, for-
The various ML models were used in [19] to infer data of ward and inverse methods, application of uncertainty
CO2, TVOC, and HCHO in buildings, which include the analysis, and available software.
SVM, GP, M5P, and backpropagation neural network. The Forward and inverse uncertainty analyses are common
ANNs have been widely used in the energy domain. Sharifa ways in energy assessment in buildings [22]. The former
and Hammad (2019) applied ANNs models to select energy analysis mainly purposes on quantifying the variation of
renovation methods in buildings concerning energy usage, outputs propagated from the uncertainty from inputs via
life cycle cost, and life cycle assessment [20]. The ANNs mathematical models as visualized in Figure 1. This ap-
model was applied to forecast short-term load in buildings proach can be used to predict energy consumption or en-
[21]. ergy-efficient design using building energy models (e.g.,
Though these methods can yield a significant proven EnergyPlus or DOE). In contrast, the later analysis aims to
forecasting accuracy improvement in some cases, they have determine unknown variables through mathematical models
usually focused on the improvement of the accuracy without from the measurement and verification process. This
Computational Intelligence and Neuroscience 3

Building information Building energy simulation Building energy

performance results
300k

200k

100k

Building envelope EnergyPlus

Thermal properties Green building studio 0k
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

Occupant schedule Integrated environmental Energy and carbon results

Operational schedule solutions (IES) Cooling and heating load
DOE-2

Figure 1: Forward uncertainty analysis for building energy assessment.

approach is also called model calibration that is often used Amasyali and El-Gohary (2018) reviewed data-driven
for retrofitting and maintaining buildings. Literature reveals approaches for predicting building energy consumption
that the forward uncertainty analysis has been used in [29]. Their study focused on investigating the prediction
assessing the building energy performance more than the scopes, the data preprocessing methods, the ML prediction
inverse uncertainty analysis [22]. model, and the performance measures used for evaluation.
Building energy efficiency is extremely important in the In terms of the prediction scopes, there were two types of
sustainability of energy and the environment. Chou et al. buildings which are residential and nonresidential building;
(2016) presented a smart decision system based on big data five data resolutions are subhourly, hourly, daily, monthly,
analytics and cloud computing for energy efficiency in and yearly. Regarding data size, most of the reviewed studies
buildings [4]. Hartono et al. (2020) analyzed modern energy used a one-month to a one-year dataset. The review indi-
projects in Indonesia and confirmed that modern energy cated that ANNs (47%) and SVM (25%) were the two most
access is one of the main factors of energy spending, especially popular ML models used for building energy prediction.
for low-income households and rural areas [25]. Shaikh et al. There is no study that applied ML models for dealing with
(2017) presented a comprehensive review of building energy energy consumption in a residential building that uses a
scenarios, the policy perspectives, and building energy effi- solar photovoltaic system. The solar photovoltaic system can
ciency programs along with landmark buildings and their produce energy sources, while buildings are in operation
characteristics in Malaysia [26]. Their study found that there during the daytime. Thus, this study applied the additive
were inadequate incentives for motivating demand-side approach to improving the performance of ANNs in pre-
management, awareness, ineffective management of quality dicting energy use in building with a solar photovoltaic
services, and inadequate legal and regulatory frameworks. system.
They suggested that energy-efficient building designs should
be considered to reduce energy use over the building life cycle. 3. Methodology
Tian et al. (2019) applied the Bayesian network model to
identify the most energy-efficient primary cooling systems Figure 2 shows the overall structure of AI applications for
[27]. The Bayesian network model was trained using high assessing energy performance in buildings. This flowchart
energy-efficient data in buildings. The trained model was consists of three components, including the buildings and
then used to decide the primary cooling systems in build- IoT network, database management system, and AI-based
ings. Their findings confirmed the applicability of data- building energy analytics. AI-based building energy ana-
driven building design. Zeng et al. (2019) applied the GP lytics may do some tasks such as predictions, classification,
regression [11] for predicting electricity consumption in clustering, alerting, and monitoring. Results from these tasks
buildings. Their conclusions were those complex buildings can recommend or suggest building users for doing further
such as hotels and shopping malls, and the GP regression actions for saving energy and reducing energy costs. This
was not better than those of simple models because of the study focuses on the prediction task in which some AI
inherent complex energy use patterns. techniques were applied. A mathematical theory of AANNs
Chen et al. (2016) developed the electric load prediction for energy prediction models in this study is presented in the
model by integrating the fuzzy time series and global har- following.
mony search algorithm and SVM that can produce reliable
prediction results [28]. Wei et al. (2018) reviewed the data-
driven approaches for assessing building energy [22]. Their 3.1. Artificial Neural Networks. ANNs models have proven
review confirmed that the data-driven approaches have been their effectiveness for engineering problems [30]. The
used widely in the energy domain such as load predictions, multilayer perceptron is a feedforward neural network that
energy pattern profiles, and retrofit solutions. Their investi- reflects inputs onto a set of appropriate outputs. The schema
gation revealed that the ANNs model was the most popular in of ANN models includes a layer for inputs with sensory
applications from energy prediction to retrofit solutions. The input nodes, hidden layers of computation nodes, and an
SVM models were often used for large-scale building energy output layer with a computation node. Equation (1) ex-
analysis due to their simplicity in the training process. presses an activated neuron in a hidden layer.
4 Computational Intelligence and Neuroscience

Building and IoT network Database management

system Artificial intelligence - based
Sensor building energy analytics
IoT network
IoT-based data
collection Transfer

Update
Smart meter Energy data
Weather data
Metadata Prediction, classification,
clustering, alerting,
HTTP/2
and monitoring
Buildings
Recommendation and suggestion

Figure 2: AI techniques for building energy performance assessment.

netj � 􏽘 wji xi and yj � f􏼐netj 􏼑, (1) minimized. Each generation fits a model to the residuals left
by the ANNs on the previous generation. Predictive results
where netj represents an activation of jth neuron, i represents are built by adding the predictions of each model. Reducing
a neuron in the preceding layer, wji represents a weight of the the shrinkage parameter helps prevent overfitting and has a
relationship between neuron j and neuron i, xi represents the smoothing effect. Details of the AANN models are present in
output of neuron i, and yj represents the transferring [31].
function as Figure 3 shows the schema of an ANNs model for
building energy use prediction in this study. The input layer
1
f􏼐netj 􏼑 � , (2) contains the historical energy consumption data, temporal
1 + eλnetj data, insolation data, and weather data (e.g., outdoor tem-
where λ adjusts the function gradient. perature). The hidden layer was used to perform the
Weights wji were updated during the training process of transforming computation between inputs and the output.
ANN models as equation (3). Δji(h) is the difference between The output layer includes prediction results of energy
two iterations as equation (4). consumption in buildings.
In a classical ANNs model, a neuron is presented by the
wji (h) � wji (h − 1) + Δji (h), (3) activating function as
f(xW + b). (5)
Δji (h) � ηδpi χpi + αΔwji (h − 1), (4)
For the AANNs model, a neuron is presented by the
where η represents a learning rate parameter; δpi represents a activating function, where the affine transform is modified
propagated error; χ pi represents a output of neuron i for by using the efficient operator in
record p; α represents a momentum parameter; and Δwji(h-
1) is a change in wji in the previous cycle. f(a•(x◇W) + b), (6)

where • denotes the elementwise multiplication operator.

3.2. Additive Artificial Neural Networks. The AANNs model The neural network (NN) in each neuron represented by
is a meta-model that can improve the effectiveness of a the activating function (equation (6) is called an additive
classical ANNs model [31]. Gradient boosting was used to NN. With regards to the training process of the AANNs
construct the additive regression model by sequentially model, the calculation of the argument derivative
fitting the base learner such as an ANNs model to the current f(a•(x◇W) + b) of the activating function is with pa-
pseudoresiduals by least-squares at each generation. The rameters W, a, b, and inputs x.
pseudoresiduals were gradients of the loss functional being

z(a•(x◇W) + b)
� Diag(x◇W), (7)
za

z(a•(x◇W) + b)
� IM , (8)
zb

z(a•(x◇W) + b) ⎡ a 􏼐sign􏼐Wi,1 􏼑 + 2Wi,1 δ xi 􏼁􏼑

⎣ 1
�⎢ ⎤⎥⎦ ≈ a•sign w 􏼁,
i (9)
zxi aM 􏼐sign􏼐Wi,M 􏼑 + 2Wi,M δ xi 􏼁􏼑
Computational Intelligence and Neuroscience 5

Historical energy Temporal Insolation Weather

Input use data data data
layer

Hidden
layer

Output
layer
Prediction of building energy consumption

Figure 3: Schema of artificial neural networks.

z(a•(x◇W) + b)
� 􏼐aj 􏼐sign xi 􏼁 + 2xi δ􏼐Wi,j 􏼑􏼑􏼑ej ≈ aj xi ej , (10)
zWi,j

where a, b ∈ RM and W ∈ RdxM are the parameters of the unseen data, are then fed into the trained AI model in the
hidden layer, x ∈ Rd is the input of the hidden layer, second phase to evaluate the effectiveness of the investigated
ei ∈ RM represents the ith component of the standard basis of AI models (i.e., the SVR, ANNs, and AANNs models).
RM, wi represents the ith column of W, sign(wi ) � 􏽐M During the evaluation process, the accuracy of the
j�1
models was measured using common statistical indices that
sign(Wi,j )ej ,for i � 1,..., M, and δ represents the function of are mean absolute error (MAE), mean absolute percentage
the Dirac delta. error (MAPE), RMSE, and correlation coefficient (R). These
The mentioned derivative was quickly computed using statistical indices were selected because they have been used
the following equation that was proposed in [32]. to evaluate machine learning models in various studies [34].
d The MAE is a measure of the difference between two
sign(x) � 2δ(x). (11) continuous variables. In this study, the MAE is an average of
dx
the absolute errors between the hourly actual energy con-
In this study, the performance of AANNs was compared sumption values and the hourly predicted energy con-
with baseline models that include the SVR models and sumption values obtained by the AI models. Its formulation
ANNs model in predicting energy use in residential is presented as
buildings. For the SVR model, the radial basis function
(RBF) kernel and the polynomial (PL) kernel were used as 1 n 􏼌􏼌􏼌 􏼌􏼌
MAE � 􏽘 􏼌y − y′ 􏼌􏼌, (12)
kernel functions. The settings of these models were pre- n i�1
sented in the following section.
where y represents the hourly actual net energy consump-
tion data, y’ is the hourly predicted net energy consumption
3.3. Model Settings and Implementation of the Proposed data obtained by the AI models, and n is the number of data
Model. Table 1 summarizes parameter information of the AI points in the sample.
models used in this study. The investigated AI models in- The MAPE presents accuracy as a percentage. This index
clude the SVR model with the PL kernel (SVR-PL), the SVR is commonly used for prediction problems, and in the
model with RBF (SVR-RBF), the ANNs model, and the proposed model, evaluation is due to its intuitive inter-
AANNs model. These AI models were performed in the pretation in terms of relative error. Equation (13) defines the
Weka platform [33] that is an open-source machine learning MAPE calculation. The RMSE is a frequently used index of
platform. The parameter settings of the AI models were the differences between values forecasted by a prediction
implemented in the Weka [33]. model and values measured. Its calculation is presented in
Figure 4 shows the implementation of the AANN model. equation (14).
􏼌 􏼌
To develop the AI models and evaluate their performance, 1 n 􏼌􏼌􏼌y − y′ 􏼌􏼌􏼌􏼌
the original dataset has been divided into two subsets which MAPE � 􏽘􏼌􏼌􏼌 􏼌, (13)
n i�1 􏼌 y 􏼌􏼌
are the training dataset (i.e., the first 90% of the dataset) and
the test dataset (i.e., the last 10% of the dataset). The eval- 􏽶�������������
􏽴
uation process consists of two steps that are the training 1 n 2
phase and the test phase. The AI models were built using the RMSE � 􏽘 y − y′ 􏼁 . (14)
training data which is accounting for 90% of the original data n i�1
in the first phase. The test data, which is considered as the
6 Computational Intelligence and Neuroscience

Table 1: Settings of investigated AI models.

Model Settings
SVR-PL Classifier � SMOreg; c � 1.0; filterType � normalize training data; kernel � PolyKernel; exponent � 1.0
SVR-
Classifier � SMOreg; c � 1.0; filterType � normalize training data; kernel � RBFKernel; gamma � 0.01
RBF
LR AttributeSelectionMethod � M5 method; eliminateColinearAttributes � true, ridge � 10−8
M5Rules BuildRegresionTree � false, unpruned � false; useUnsmoothed � false
ANNs HiddenLayers � a; learningRate � 0.3; momentum � 0.2; normalizeAttributes � true; trainingTime � 500
Classifier � AdditiveRegression_MultilayerPerceptron; numIterations � 10; shrinkage � 1.0; hiddenLayers � a;
AANNs
learningRate � 0.3; momentum � 0.2; normalizeAttributes � true; trainingTime � 500

The training phase

Training
prediction
90% model
Energy consumption data (AANNs)

Training data
Database

Weather data
The test phase
Prediction data
accuracy evaluation

Trained AANNs
10% model
Insolation data Test data

Figure 4: Implementation of the AANNs model.

4. Experiment and Results The SVR-PL model used the simple polynomial kernel as
a kernel function for the prediction. Therefore, its perfor-
4.1. Data Source. Data used in this study were derived from mance did not look good in terms of accuracy indices, which
the Net-Zero Energy Residential Test Facility at the National are 26.18% and 28.60% in terms of the MAPE in the training
Institute of Standards and Technology (NIST) Engineering phase and test phase, respectively. Similarly, the MAE and
Lab [35, 36]. A solar photovoltaic system was used in this RMSE values obtained by the SVR-PL model were relatively
building as a renewable energy source. This data is open and high, up to 236.83 Wh and 430.69 Wh, respectively, for
free. The one-year dataset in hourly resolution was used to predicting residential building energy use profiles. The re-
evaluate the proposed prediction model in this study. Table 2 sults of these statistical indices indicated that the SVR-PL
summarizes the input and output information that was used was not effective in energy use prediction in residential
for training and testing AI models. Attributes of the dataset buildings with renewable energy. Figure 7 provides a vi-
include the net building energy consumption, outdoor dry sualization of the forecasted and recorded values of the net
bulb temperature, insolation, day of the week, hour of the energy consumption obtained by the SVR-PL model. The
day. All data were collected hourly. Figure 5 visualizes the diagonal line in Figure 7 indicates an absolute agreement
outdoor dry bulb temperature, net energy consumption in between the forecasted and recorded values. The scatter plot
the experimental building, and insolation profiles in the in Figure 7 reveals that although most scatter points locate
hourly resolution for a year. For providing readers with a around the black, diagonal line, many points, that were in
clear look, Figure 6 presents the hourly insolation profile and the three dashed red circles, were far from the absolute
hourly energy consumption profile in the building for a agreement. This means the SVR-PL model is still limited to
week. capture an energy use profile.
When the RBF kernel was used as the kernel function in
the SVR model, the performance of the SVR model was
4.2. Results and Discussion. The investigated AI models in slightly enhanced in the prediction as compared to the SVR-
this study include the SVR-PL, SVR-RBF, linear regression PL model. The SVR-RBF model can yield predictive accu-
(LR), M5Rules, ANNs, and AANNs models. Their perfor- racy at 26.38% in the MAPE and 225.06 Wh in the MAE.
mance was assessed using a dataset that was recorded from a Figure 8 shows the scatter plot that compares the recorded
residential building with renewable energy. After the eval- net energy consumption in the experimental building and
uation process, the predictive accuracy of these AI models is the predicted net energy consumption predicted by the SVR-
given in Table 3 via MAPE, MAE, RMSE, and R values RBF model. The scatter plot in Figure 8 presents that the
regarding the training step and test step. SVR-RBF model underestimated energy consumption in the
Computational Intelligence and Neuroscience 7

Table 2: Data attributes for model evaluation.

Symbol Parameter Unit Value
Input
X1 Day of the week — Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday
X2 Hour of the day — 0, 1, 2, . . ., 21, 22, 23
X3 Isolation Wh/m2
X4 Outdoor dry bulb temperature °C
Y_historical Historical building energy consumption Wh
Output
Y Future building energy consumption Wh

1200
Insolation (Wh/m2)

1000
800
600
400
200
0
1
221
441
661
881
1101
1321
1541
1761
1981
2201
2421
2641
2861
3081
3301
3521
3741
3961
4181
4401
4621
4841
5061
5281
5501
5721
5941
6161
6381
6601
6821
7041
7261
7481
7701
7921
8141
Time horizon (01 February, 2015–30 January, 2016)
(a)
Outdoor dry bulb temperature (°C)

40
30
20
10
0
–10
–20
1
221
441
661
881
1101
1321
1541
1761
1981
2201
2421
2641
2861
3081
3301
3521
3741
3961
4181
4401
4621
4841
5061
5281
5501
5721
5941
6161
6381
6601
6821
7041
7261
7481
7701
7921
8141
Time horizon (01 February, 2015–30 January, 2016)

(b)
Net energy consumption (Wh)

5000
4000
3000
2000
1000
0
1
221
441
661
881
1101
1321
1541
1761
1981
2201
2421
2641
2861
3081
3301
3521
3741
3961
4181
4401
4621
4841
5061
5281
5501
5721
5941
6161
6381
6601
6821
7041
7261
7481
7701
7921
8141

Time horizon (01 February, 2015–30 January, 2016)

(c)

Figure 5: The hourly data profile for a year. (a) Hourly insolation profile. (b) Hourly profile of outdoor dry bulb temperature for a year.
(c) Net energy consumption profile.

residential building which is illustrated at scattering points Figure 9 shows a comparing result between the observed
in the dashed red circle located above the black straight line. net energy consumption data and energy consumption
The scattering points in the dashed red circles located below predicted by the ANNs model in the test phase. In terms of
the black straight line depicted over-estimated energy the MAPE, the ANNs models yielded 12.86% in the training
consumption by the SVR-RBF. The R values obtained by the phase and 14.68% in the test phase. The findings in Table 3
SVR-RBF and SVR-PL models were lower than 0.7 in the also revealed that the statistical indices of the ANNs model
training and test steps. Generally, two variants of the SVR were 114.91 Wh in the MAE in the test phase. The R value
models did not perform well in predicting the profiles of achieved by the ANNs model was 0.932 in the test phase
energy consumed in the experimental building. which depicts the good agreement between the actual and
8 Computational Intelligence and Neuroscience

Hourly insolation profile for a week

1000 Sunday Monday Tuesday Wednesday Thursday Friday Saturday

Insolation (Wh/m2)
800
600
400
200
0
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
121
126
131
136
141
146
151
156
161
166
Time horizon (01 February, 2015–07 February, 2015)
(a)
Net energy consumption profile for a week
Net energy consumption (Wh)

4000
Sunday Monday Tuesday Wednesday Thursday Friday Saturday
3500
3000
2500
2000
1500
1000
500
0
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
121
126
131
136
141
146
151
156
161
166
Time horizon (01 February, 2015–07 February, 2015)
(b)

Figure 6: Hourly data profile of (a) insolation and (b) net energy consumption for a week.

Table 3: Predictive accuracy of the machine learning models for building energy consumption.
Accuracy measures in the training step. Training data Accuracy measures in the test step. Test data (N � 835,
ML models (N � 7493, 90%) 10%)
MAPE (%) MAE (Wh) RMSE (Wh) R MAPE (%) MAE (Wh) RMSE (Wh) R
SVR-PL 26.18 240.69 438.45 0.653 28.60 236.83 430.69 0.622
SVR-RBF 24.13 226.58 412.98 0.698 26.38 225.06 412.53 0.659
LR 30.74 265.67 409.80 0.586 36.74 275.68 416.64 0.654
M5Rules 11.96 104.26 172.77 0.953 14.20 112.44 213.43 0.921
ANNs 12.86 106.80 170.28 0.955 14.68 114.91 199.64 0.932
AANNs 12.78 106.88 165.15 0.961 14.04 111.98 188.68 0.940
Predicted net energy consumption in the

Predicted net energy consumption in the

4000 4000
3500 3500
3000 3000
building (Wh)

building (Wh)

2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
0 500 1000 1500 2000 2500 3000 3500 4000 0 500 1000 1500 2000 2500 3000 3500 4000
Actual net energy consumption in the building (Wh) Actual net energy consumption in the building (Wh)
Figure 7: Prediction results by the SVR-PL model. Figure 8: Scatter plot of prediction results by the SVR-RBF model.
Computational Intelligence and Neuroscience 9

ANNs

Predicted net energy consumption in the

4000
3500
3000

building (Wh)
2500
2000
1500
1000
500
0
0 500 1000 1500 2000 2500 3000 3500 4000
Actual net energy consumption in the building (Wh)
Figure 9: Prediction results by the ANNs model.

Table 4: Performance comparison among machine learning models.

Accuracy measures in the test step Improvement percentage (%)
ML models
MAPE (%) MAE (Wh) RMSE (Wh) R MAPE MAE RMSE R
SVR-PL 28.60 236.83 430.69 0.622 103.75 111.50 128.26 33.83
SVR-RBF 26.38 225.06 412.53 0.659 87.96 100.98 118.64 29.89
LR 36.74 275.68 416.64 0.654 161.78 146.19 120.82 30.48
M5Rules 14.20 112.44 213.43 0.921 1.14 0.41 13.12 2.00
ANNs 14.68 114.91 199.64 0.932 4.60 2.61 5.81 0.85
AANNs 14.04 111.98 188.68 0.940

predicted data. These predictive results by the ANN model histogram produced by the SVR-PL model (Figure 12(a)),
were better than those predicted by the SVR-PL and SVR- the SVR-RBF model (Figure 12(b)), the ANNs model
RBF models. As given in Table 3, the M5Rules model (Figure 12(c)), and the proposed AANNs model
achieved good accuracy. Its accuracy indices were 14.2%, (Figure 12(d)). Compared to the predictive performance of
112.44 Wh, 213.43 Wh, and 0.921 in the MAPE, MAE, the SVR-PL model, the AANNs model can significantly
RSME, and R, respectively. The LR model was ineffective in improve the predictive accuracy which was about 103.75% in
predicting energy patterns in buildings with low predictive the MAPE, 111.50% in the MAE, and 128.26% in the RMSE.
statistical indices. The inherent linear characteristic of the LR Similarly, the AANNs outperformed significantly the SVR-
limits its capability in modeling the nonlinear relationship RBF models in the residential building energy consumption
between inputs and the predicted energy consumption. prediction. Because the AANNs model is an enhanced
Table 4 provides the performance comparison among version of the ANNs, its performance was better than those
the investigated ML models. The proposed AANNs model of the ANN model. Compared to the ANNs model, accuracy
achieved a good predictive accuracy, in which its statistical improvement percentages by the AANNs model were 4.6%
measures were 14.04% in the MAPE, 111.98 Wh in the MAE, in the MAPE, 2.61% in the MAE, and 5.81% in the RMSE.
188.68 in the RMSE, and 0.940 in the R for predicting hourly The comparisons in Table 4 confirmed the outperformance
net energy consumption during the test phase. Figure 10 of the AANNs models to other models.
shows the prediction values of energy consumption in the The Wilcoxon signed-rank test is a nonparametric test in
residential building. Most scatter points were close to the which two paired samples were compared to evaluate a
black line which means the AANNs model was effective in significant difference between two population means. This
forecasting the hourly net energy consumption in the statistical test has been used in [28, 37, 38] to confirm the
building with the renewable energy source. Besides, Fig- significance of the accuracy enhancement. Thus, the Wil-
ure 11 plots actual and predicted values of net energy coxon signed-rank test was used in this study. The Wilcoxon
consumption by AANNs in the test phase over the time signed-rank test was used for pair comparison between
horizon. prediction results by the AANNs model with those obtained
The performance comparison among the ML models in by the SVR-PL, SVR-RBF, and ANNs models, respectively.
Table 4 depicts that the AANNs model was the most effective The statistical test results depicted that the computed p value
forecasting model in terms of all performance indices of the was lower than the significance level alpha of 0.05. Thus, the
MAPE, MAE, RMSE, and R. The proposed AANNs model null hypothesis H0 was rejected, and the alternative hy-
obtained the lowest MAPE with 14.04%, followed by the pothesis Ha was accepted. Therefore, the significant differ-
ANNs model with the MAPE of 14.68% and SVR-RBF with ence was confirmed between the performance of the AANNs
the MAPE of 26.38%. Figure 12 presents the prediction error models and other compared models.
10 Computational Intelligence and Neuroscience

AANNs

Predicted net energy consumption in the

4000
3500
3000

building (Wh)
2500
2000
1500
1000
500
0
0 500 1000 1500 2000 2500 3000 3500 4000
Actual net energy consumption in the building (Wh)
Figure 10: Scatter plots of actual and predicted net energy consumption in the building in the test step by the AANNs model.

4000

3500
Net energy consumption (Wh)

3000

2500

2000

1500

1000

500

0
1
21
41
61
81
101
121
141
161
181
201
221
241
261
281
301
321
341
361
381
401
421
441
461
481
501
521
541
561
581
601
621
641
661
681
701
721
741
761
781
801
821
Time horizon for test data

Actual values
Predicted values by the additive ANNs

Figure 11: Actual and predicted net energy consumption by AANNs in test phase.

350 100 300 100

300 250
80 80
Percentage (%)

Percentage (%)
250 200
Frequency

Frequency

200 60 60
150
150 40 40
100 100
20 50 20
50
0 0 0 0
–2702
–2388
–2075
–1761
–1447
–1134
–820
–506
–193
121
435
748
1062
1376
More

–2616
–2323
–2030
–1737
–1443
–1150
–857
–563
–270
23
316
610
903
1196
More

Error bin Bin

Frequency Frequency
Cumulative % Cumulative %

(a) (b)
Figure 12: Continued.
Computational Intelligence and Neuroscience 11

350 120 350 100

300 100 300
80

Percentage (%)

Percentage (%)
250 80 250
Frequency

Frequency
200 200 60
60
150 150 40
100 40 100
20 20
50 50
0 0 0 0
–2087
–1865
–1642
–1419
–1196
–974
–751
–528
–305
–83
140
363
586
808
More

–1876
–1629
–1382
–1135
–889
–642
–395
–148
99
345
592
839
1086
1333
More
Error bin Error bin

Frequency Frequency
Cumulative % Cumulative %

(c) (d)

Figure 12: Prediction error histogram by the prediction models. (a) Error histogram by the SVR-PL model. (b) Error histogram by the SVR-
RBF model. (c) Error histogram by the ANNs model. (d) Error histogram by the AANNs model.

5. Conclusions Data Availability

Energy efficiency is one of the most concerning topics within The data used to support the findings of this study are in-
academic researchers and decision-makers in the energy cluded within the article.
sector. An energy consumption prediction in a building is
the basis for optimizing building performance and reducing Conflicts of Interest
energy costs. This study proposed additive artificial neural
networks (AANNs) that can accurately predict energy The authors declare that they have no conflicts of interest.
consumption in buildings concerning historical data of
energy use and weather conditions. This study also com- Acknowledgments
pared the performance of the AANNs with other ML models
such as the support vector regression with the polynomial This research was funded by Funds for Science and Tech-
kernel function (SVR-PL), the support vector regression nology Development of the University of Danang (B2019-
with the radial basis function kernel function (SVR-RBF), DN02-59).
and the artificial neural networks (ANNs). Their perfor-
mance was assessed using a one-year dataset in the hourly References
resolution that was recorded from a residential building with [1] M. Cellura, F. Guarino, S. Longo, and G. Tumminia, “Climate
renewable energy sources. change and the building sector: modelling and energy im-
The proposed AANNs model achieved a good predictive plications to an office building in southern Europe,” Energy for
accuracy in which its statistical measures were 14.04% in the Sustainable Development, vol. 45, pp. 46–65, 2018.
mean absolute percentage error (MAPE) and 111.98 Watt- [2] T. Zhang, D. Wang, H. Liu, Y. Liu, and H. Wu, “Numerical
hour in the mean absolute error (MAE) for predicting hourly investigation on building envelope optimization for low-energy
net energy consumption. The AANNs model was the most buildings in low latitudes of China,” Building Simulation,
vol. 13, no. 2, pp. 257–269, 2020.
effective forecasting model among the investigated AI
[3] A.-D. Pham, N.-T. Ngo, T. T. Ha Truong, N.-T. Huynh, and
models. Compared to the support vector regression (SVR), N.-S. Truong, “Predicting energy consumption in multiple
the AANNs model can significantly improve the predictive buildings using machine learning for improving energy ef-
accuracy by about 103.75% in the MAPE and 111.50% in the ficiency and sustainability,” Journal of Cleaner Production,
MAE. Similarly, compared to the ANNs model, accuracy vol. 260, p. 121082, 2020.
improvement percentages by the AANNs model were 4.6% [4] J. S. Chou, N. T. Ngo, W. K. Chong, and G. E. Gibson, “16 - big
in the MAPE and 2.61% in the MAE. Thus, the AANNs data analytics and cloud computing for sustainable building
model was recommended as an effective AI-based model for energy efficiency,” in Start-Up Creation, F. Pacheco-Torgal,
predicting net energy consumption in residential buildings E. Rasmussen, C.-G. Granqvist, V. Ivanov, A. Kaklauskas, and
with a solar photovoltaic system. S. Makonin, Eds., pp. 397–412, Woodhead Publishing, Saw-
The contributions of this study include (1) collection of ston, UK, 2016.
[5] A. Raza, T. N. Malik, M. F. N. Khan, and S. Ali, “Energy
building energy use profiles in hourly resolution and their
management in residential buildings using energy hub ap-
associated weather data, (2) investigation of the potential proach,” Building Simulation, vol. 13, no. 2, pp. 363–386,
power of artificial intelligence techniques in forecasting 2020.
future energy consumption in buildings, and (3) develop- [6] Z. Wang and R. S. Srinivasan, “A review of artificial intelli-
ment of the effectiveness and capability of the AANNs in the gence based building energy use prediction: contrasting the
prediction of building energy consumption. capabilities of single and ensemble prediction models,”
12 Computational Intelligence and Neuroscience

Renewable and Sustainable Energy Reviews, vol. 75, pp. 796– [22] W. Tian, Y. Heo, P. de Wilde et al., “A review of uncertainty
808, 2017. analysis in building energy assessment,” Renewable and
[7] H. Song, A. K. Qin, and F. D. Salim, “Evolutionary model Sustainable Energy Reviews, vol. 93, pp. 285–301, 2018.
construction for electricity consumption prediction,” Neural [23] W. Tian, S. Yang, Z. Li, S. Wei, W. Pan, and Y. Liu, “Iden-
Computing and Applications, vol. 32, no. 16, pp. 12155–12172, tifying informative energy data in Bayesian calibration of
2020. building energy models,” Energy and Buildings, vol. 119,
[8] X. Huang, D. Zhang, and X. Zhang, “Energy management of pp. 363–376, 2016.
intelligent building based on deep reinforced learning,” [24] A. Chong, W. Xu, S. Chao, and N.-T. Ngo, “Continuous-time
Alexandria Engineering Journal, vol. 60, no. 1, pp. 1509–1517, Bayesian calibration of energy models using BIM and energy
2021. data,” Energy and Buildings, vol. 194, pp. 177–190, 2019.
[9] E. Jahani, K. Cetin, and I. H. Cho, “City-scale single family [25] D. Hartono, S. H. Hastuti, A. A. Balya, and W. Pramono,
residential building energy consumption prediction using “Modern energy consumption in Indonesia: assessment for
genetic algorithm-based numerical moment matching tech- accessibility and affordability,” Energy for Sustainable De-
nique,” Building and Environment, vol. 172, Article ID velopment, vol. 57, pp. 57–68, 2020.
106667, 2020. [26] P. H. Shaikh, N. B. M. Nor, A. A. Sahito, P. Nallagownden,
[10] X. Juan and G. Weijun, “Analysis on energy consumption of I. Elamvazuthi, and M. S. Shaikh, “Building energy for sus-
rural building based on survey in northern China,” Energy for tainable development in Malaysia: a review,” Renewable and
Sustainable Development, vol. 47, pp. 34–38, 2018. Sustainable Energy Reviews, vol. 75, pp. 1392–1403, 2017.
[11] A. Zeng, H. Ho, and Y. Yu, “Prediction of building electricity [27] Z. Tian, B. Si, X. Shi, and Z. Fang, “An application of Bayesian
usage using Gaussian Process Regression,” Journal of Building Network approach for selecting energy efficient HVAC sys-
Engineering, vol. 28, p. 101054, 2020. tems,” Journal of Building Engineering, vol. 25, p. 100796,
[12] C. Dai, H. Zhang, E. Arens, and Z. Lian, “Machine learning 2019.
approaches to predict thermal demands using skin temper- [28] Y. H. Chen, W.-C. Hong, W. Shen, and N. N. Huang, “Electric
atures: steady-state conditions,” Building and Environment, load forecasting based on a least squares support vector
vol. 114, pp. 1–10, 2017. machine with fuzzy time series and global harmony search
[13] C. Xu, H. Chen, J. Wang, Y. Guo, and Y. Yuan, “Improving algorithm,” Energies, vol. 9, no. 2, p. 70, 2016.
prediction performance for indoor temperature in public [29] K. Amasyali and N. M. El-Gohary, “A review of data-driven
buildings based on a novel deep learning method,” Building building energy consumption prediction studies,” Renewable
and Environment, vol. 148, pp. 128–135, 2019. and Sustainable Energy Reviews, vol. 81, pp. 1192–1205, 2018.
[14] D.-K. Bui, T. N. Nguyen, T. D. Ngo, and H. Nguyen-Xuan, [30] V. Kecman, Learning and Soft Computing: Support Vector
“An artificial neural network (ANN) expert system enhanced Machines, Neural Networks, and Fuzzy Logic Models, MIT
with the electromagnetism-based firefly algorithm (EFA) for Press, Cambridge, MA, USA, 2001.
predicting the energy consumption in buildings,” Energy, [31] J. H. Friedman, “Stochastic gradient boosting,” Computa-
vol. 190, Article ID 116370, 2020. tional Statistics & Data Analysis, vol. 38, no. 4, pp. 367–378,
[15] M. A. Jallal, A. González-Vidal, A. F. Skarmeta, S. Chabaa, and 2002.
A. Zeroual, “A hybrid neuro-fuzzy inference system-based [32] R. N. Bracewell, The Fourier Transform and its Applications,
algorithm for time series forecasting applied to energy con- McGraw-Hill, New York, NY, USA, 2000.
sumption prediction,” Applied Energy, vol. 268, Article ID [33] U. Waikato, https://www.cs.waikato.ac.nz/ml/weka/, 2020.
114977, 2020. [34] M. Zeinolabedini and M. Najafzadeh, “Comparative study of
[16] S. Ganguly, A. Ahmed, and F. Wang, “Optimised building different wavelet-based neural network models to predict
energy and indoor microclimatic predictions using knowl- sewage sludge quantity in wastewater treatment plant,” En-
edge-based system identification in a historical art gallery,” vironmental Monitoring and Assessment, vol. 191, no. 3, p. 163,
Neural Computing and Applications, vol. 32, no. 8, 2019.
pp. 3349–3366, 2020. [35] W. Healy, A. Hunter Fanney, B. Dougherty, and W. Vance
[17] S. Seyedzadeh, F. Pour Rahimian, P. Rastogi, and I. Glesk, Payne, “Tania ullah, lisa Ng, and Farhad Omar, net zero
“Tuning machine learning models for prediction of building energy residential test facility instrumented data,” Year, vol. 2,
energy loads,” Sustainable Cities and Society, vol. 47, Article 2017.
ID 101484, 2019. [36] W. Healy, T. H. Chen, B. Dougherty et al., “Net zero energy
[18] M. Najafzadeh and G. Oliveto, “Riprap incipient motion for residential test facility instrumented data,” Year, vol. 1, 2018.
overtopping flows with machine learning models,” Journal of [37] G.-F. Fan, S. Qing, H. Wang, W.-C. Hong, and H.-J. Li,
Hydroinformatics, vol. 22, no. 4, pp. 749–767, 2020. “Support vector regression model based on empirical mode
[19] S. Chen, K. Mihara, and J. Wen, “Time series prediction of decomposition and auto regression for electric load fore-
CO2, TVOC and HCHO based on machine learning at dif- casting,” Energies, vol. 6, no. 4, pp. 1887–1901, 2013.
ferent sampling points,” Building and Environment, vol. 146, [38] M.-W. Li, Y.-T. Wang, J. Geng, and W.-C. Hong, “Chaos
pp. 238–246, 2018. cloud quantum bat hybrid optimization algorithm,” Non-
[20] S. A. Sharif and A. Hammad, “Developing surrogate ANN for linear Dynamics, vol. 103, no. 1, pp. 1167–1193, 2021.
selecting near-optimal building energy renovation methods
considering energy consumption, LCC and LCA,” Journal of
Building Engineering, vol. 25, Article ID 100790, 2019.
[21] P. Singh and P. Dwivedi, “Integration of new evolutionary
approach with artificial neural network for solving short term
load forecast problem,” Applied Energy, vol. 217, pp. 537–549,
2018.