Energy & Buildings 240 (2021) 110886

Contents lists available at ScienceDirect

Energy & Buildings

journal homepage: www.elsevier.com/locate/enb

A building energy consumption prediction model based on rough set

theory and deep learning algorithms
Lei Lei a, Wei Chen a, Bing Wu a, Chao Chen a, Wei Liu b,⇑
a
School of Architecture and Transportation Engineering, Guilin University of Electronic Technology, 1 Jinji Road, Guilin 541004, China
b
Division of Sustainable Buildings, Department of Civil and Architectural Engineering, KTH Royal Institute of Technology, Brinellvägen 23, Stockholm 100 44, Sweden

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

Article history: The efficient and accurate prediction of building energy consumption can improve the management of
Received 3 October 2020 power systems. In this paper, the rough set theory was used to reduce the redundant influencing factors
Revised 11 January 2021 of building energy consumption and find the critical factors of building energy consumption. These key
Accepted 3 March 2021
factors were then used as the input of a deep neural network with a ‘‘deep” architecture and powerful
Available online 10 March 2021
capabilities in extracting features. Building energy consumption is output of the deep neural network.
This study collected data from 100 civil public buildings for rough set reduction, and then collected data
Keywords:
from a laboratory building of a university in Dalian for nearly a year to train and test deep neural net-
Building energy consumption
Prediction model
works. The test included both the short-term and medium-term predictions of building energy consump-
Rough set tion. The prediction results of the deep neural network were compared with that of the back propagation
Deep learning neural network, Elman neural network and fuzzy neural network. The results show that the integrated
Attribute reduction rough set and deep neural network was the most accurate. The method proposed in this study could pro-
vide a practical and accurate solution for building energy consumption prediction.
Ó 2021 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

1. Introduction ware for building energy simulation includes EnergyPlus, equest,

and octet, etc.[3] In simulating the building energy consumption,
With the rapid development of the global economy, building the physical-model based methods require detailed input informa-
energy consumption keeps increasing and it now accounts for tion, such as physical features of the building, HVAC systems,
more than one-third of total global energy consumption [1]. There- equipment, and occupants’ schedule, etc.[4]. The informative input
fore, reducing building energy consumption to relieve energy pres- would help achieve accurate prediction of building energy con-
sure has become particularly important for realizing sustainable sumption, but as well makes the modeling process quite compli-
goals such as affordable energy and sustainable cities. Accurate cated and slow.
prediction of building energy consumption is one of the effective Data-driven methods are divided into two categories: statistical
means to reduce building energy consumption, since it helps methods and machine learning methods. The statistical methods
achieve better control of power system and improve energy uti- are also called grey box models, which are generally in the form
lization [2]. of regression analysis prediction [5] or time series-based predic-
The methods of performing building energy consumption pre- tion. Regression analysis simulates and predicts energy consump-
diction can be generally divided into physical-model based meth- tion by establishing mathematical expressions between
ods and data-driven methods. Among them, physical-model independent and dependent variables. Guo et al. [6] proposed an
based methods are also called white-box based approaches, which improved multivariable linear regression model to predict the
use the principles of physics to evaluate the building energy con- daily mean cooling load of office buildings in which three main
sumption. For example, to calculate the energy consumption of measures, including the principal component analysis of meteoro-
the heating ventilation, and air conditioning (HVAC) system, the logical factors, cumulative effect of high temperature and dynamic
general methods are variable base degree day methods, equivalent two-step correction. Pombeiro et al. [7] compared and evaluated
full load hours method and bin method. The commonly used soft- the prediction results of several low-complexity models for the
energy consumption of an institutional building. Their research
results indicated that both the developed fuzzy systems and neural
⇑ Corresponding author. networks have better performance and accuracy indicators than
E-mail address: [email protected] (W. Liu). linear regression models. Time series-based analysis is a statistical

https://doi.org/10.1016/j.enbuild.2021.110886
0378-7788/Ó 2021 The Author(s). Published by Elsevier B.V.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Nomenclature

U Domain Acronyms
A Non-empty finite set of attributes HVAC heating, ventilating, and air conditioning
K The range of all attributes SVM support vector machine
f Information function ANN artificial neural network
C Conditional attribute BP back propagation
D Decision attribute RS rough set
V visible layer of restricted Boltzmann machine DBN deep belief neural network
H hidden layer of restricted Boltzmann machine RBM restricted Boltzmann machine
Wmn the weight matrix between the visible layer and the hid- FNN fuzzy neural network
den layer RS-BP rough set combined with BP neural network
b the offset of the visible node RS-SRN rough set combined with Elman neural network
c the offset of the hidden node RS-FNN rough set combined with fuzzy neural network
RS-DBN rough set combined with deep belief neural network

method for dynamic data processing. Jallal et al. [8] proposed a although the gray system theory is simple to calculate, it has
time series forecasting algorithm based on a hybrid neuro-fuzzy a narrow scope of application. It is rarely used to directly pre-
inference system to predict building energy consumption. The dict the energy consumption of the entire building and is gener-
results show that this method is more robust than the traditional ally only suitable for short-term prediction of the energy
adaptive neuro-fuzzy inference system models. However, the pre- consumption of part of the building system.
diction process of this method is more complicated, and the effi-
ANNs are more popular in building energy consumption predic-
ciency of calculating the energy consumption is not high. In tion research because of their strong self-learning and good
general, statistical methods have the advantage of a simple struc- nonlinear function fitting ability, which can maintain a rela-
ture and relatively easy model establishment. However, at the tively high accuracy in energy consumption prediction. Wei
same time, due to the complex interaction between input elements et al. [14] combined blind system identification and feed-
in the gray-box model [9], problems such as a low calculation effi- forward neural network to predict the occupancy level and
ciency and low prediction accuracy are likely to occur. energy consumption of office buildings. Ekici and Aksoy [15]
To overcome the deficiencies of the physical-model based used back propagation (BP) neural network to predict the heat-
methods and the statistical methods, machine learning method ing energy requirements of three different buildings. Rahman
has been developed [10]. Machine learning methods for building et al. [16] proposed a recurrent neural network model that
energy consumption predictions generally include support vector can predict the power consumption of commercial and residen-
machines (SVMs), gray system theory, and artificial neural net- tial buildings for periods longer than a week at one-hour reso-
works (ANNs): lution intervals. Li et al. [17] combined a genetic algorithm
and an adaptive neural fuzzy inference system to predict the

SVMs is a learning method based on structural risk minimiza- energy consumption of campus buildings. These researches on
tion criteria. It can predict building energy consumption with building energy consumption prediction based on shallow neu-
limited samples. Zhong et al. [1] proposed a support vector ral networks show that neural networks have better prediction
regression method that improves the accuracy and generaliza- accuracy than other methods. But they also show that shallow
tion ability of building energy consumption prediction by look- neural networks are prone to losing the original data informa-
ing for the optimal feature space. The model only processes a tion, and the selection of its network topology is generally
small sample of building energy consumption data for one determined by trial and error, which is difficult to find the opti-
month, and it is only suitable for short-term predictions of mal network topology.
building energy consumption. Dong et al. [11] used an SVM
method to predict the monthly building energy consumption In addition to the three main machine learning algorithms men-
of four commercial buildings in the tropics. The error of the pre- tioned above, some studies have applied decision trees [18,19],
diction results is small, which is related to the small data pool extreme learning machines [20,21] and other artificial intelligence
used in the study. The whole data profile in this study consisting algorithms to building energy consumption prediction. In general,
of only 4 years’ monthly utility bills may include less scattered the existing machine learning algorithms are more advanced than
or abnormal data. The SVM method is suitable for nonlinear physical models and statistical methods in all aspects. However,
high-dimensional pattern recognition, but this method is mem- the existing machine learning algorithms have the problems of
ory and computational demanding, which make it difficult to inaccurate prediction results, low calculation efficiency and limited
implement for large-scale training samples. application range. With the continuous development of the artifi-

Gray system theory predicts the development of building cial intelligence, there is now a deep learning algorithm, which is
energy consumption by analyzing the correlation between sys- expected to make up for the shortcomings of existing building
tem factors to generate a data sequence with strong regularity. energy consumption prediction methods and bring accurate pre-
In terms of the application of gray system theory, Robert and dictions to building energy consumption.
Uwe [12] used gray system theory to analyze the thermal pro- Building energy consumption is affected by many factors, such
cess of buildings and predict the thermal energy consumption as weather conditions, building thermal environment, and building
of HVAC system. Guo et al. [13] proposed a method to predict energy usage. The data of building energy consumption have a high
the energy consumption of domestic heat pump water heaters degree of uncertainty and randomness, which makes it difficult for
based on the gray system theory. It should be noted that the prediction model to be of low bias and variance. Deep learning

2
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

is able to make full use of the training data to extract the building where xi is an object that represents a set of data samples. A =
energy consumption features, which further leads to accurate pre- {a1, a2, a3, . . ..an} is a nonempty finite set of attributes and ai(xj)
diction of building energy consumption. Compared with the shal- is the value of  recorded on attribute a. The elements in set A refer
low neural network, the deep neural network contains multiple to the input parameters and building energy consumption. a(x) are
hidden layers and has a ‘‘deep” architecture. These features make the data corresponding to each impact factor. K is the value range
it a very popular artificial intelligence algorithm at present. In of attribute set A. f is the information function set of U and A, and
medicine [22,23], finance [24,25], transportation [26], environ- represents a mapping of U  A ! K, which assigns an information
ment [27] and many other fields, there is related research and value to each attribute of each object, namely, 8a 2 A; x 2 U;
applications of deep learning. Using deep learning as a post infor- f ðx; aÞ 2 K.
mation recognition system for predictive models, the features of If the nonempty finite set of attributes A is composed of a con-
energy consumption data are extracted through the powerful fea- ditional attribute set C and a decision attribute D and satisfies
ture extraction capabilities and ‘‘depth” architecture of deep learn- A ¼ C [ D; C \ D–£, then S is called a decision system. Generally,
ing. In this way, high-precision building energy consumption dealing with decision problems by the rough set method can be
prediction results can be obtained. expressed in the form of decision table. In this study, the decision
table of the information system was expressed as S= (U, C, D, K, f).
Each row in the decision table describes an object (a set of data
2. Research methodology
samples, which is called a decision rule), and each column repre-
sents an attribute of the object (conditional attribute set C or deci-
The prediction of building energy consumption can be divided
sion attribute D). In this study, the conditional attribute set C
into short-term prediction, medium-term prediction and long-
represents various factors that affect the building energy consump-
term prediction according to the length of the prediction time scale
tion, and the decision attribute D represents building energy
[28]. The forecast range of general short-term forecasts is several
consumption.
hours to one week [29,30], the forecast range of medium-term
In engineering applications, the range of the conditional attri-
forecasts is from several weeks to several months [31], and the
butes and decision attributes of the problem mostly involve con-
realization of long-term forecasts is usually based on the establish-
tinuous values. However, when the rough set method is used to
ment of annual forecasts [32]. Due to data limitation, this study
deal with the decision table, the attribute values in the decision
focused on the short-term and mid-term forecasts.
table must be discrete. Therefore, the data need to be discretized
Fig. 1 shows the overall research framework. The parameters Ci
for establishing the decision table. This study discretizes the data
(i = 1, 2, . . ., 20) are the 20 inputs, which represents the hourly
by dividing the range into a finite number of regions by selecting
based weather (air temperature, humidity, wind speed, and solar
breakpoints and each region corresponds to an attribute value.
radiation), building information (levels, area, orientation,
For example, xl (l = 1, 2, . . ., n-1) is selected as a breakpoint in
window-to-wall ratio, thermal conductivity of the external wall
the range of values, and then, the n-1 breakpoint divides the con-
and roof, shading coefficient, and length to width ratio), building
tinuous range into n attribute values, which can reduce the com-
utilization (lighting fixtures, occupant density, and indoor air tem-
plexity of the problem. The discretization of continuous
perature), and air conditioning system (fresh air flowrate, COP, air
attributes can be divided according to expert experience or auto-
supply temperature, fan efficiency, and pump efficiency). Since
matically divided by the system according to certain rules. The
there is larger amount of input data, this investigation firstly
most commonly used discretization methods for continuous attri-
applied rough set theory, which was proposed by Pawlak [33], to
bute are equal distance division, equal frequency division, the L-
eliminate the redundant data and identify the effective hourly
method, the P-method, the Naivescaler algorithm, the SemiNaives-
input data Zi (i = 1, 2, . . ., m; m  20) that mainly determines the
caler algorithm, and Boolean reasoning [35]. This study adopted
building energy consumption. For the short-term, this study also
the equidistant partitioning to discretize the continuous attributes
developed the neural network without rough sets theory as a com-
in the attribute reduction algorithm of rough sets.
parison. For the mid-term prediction, the hourly data Zi are pro-
In the information system of the raw rough set decision table,
cessed into daily data Ti (i = 1, 2, . . ., m) for the neural network.
some of the conditional attributes are critical to the information
In this study, the rough set theory is used as the front-end proces-
system of the decision table, while others may be redundant. The
sor of the prediction model, and deep learning is used as the post-
purpose of attribute reduction is to simplify the information sys-
information recognition system of the prediction model, which
tem by finding the necessary conditional attributes and eliminat-
appropriately combines the advantages of both. In this way, not
ing the redundant and unimportant conditional attributes in the
only simplifies the sample structure and improves the generaliza-
information system. The attribute reduction of a decision table is
tion ability and running speed of the network, but also strengthens
not unique, there are many minimum relative reduction results.
the anti-interference ability and prediction accuracy of the model.
Therefore, to ensure that the classification ability of the informa-
This study collected data from 100 civil public buildings for rough
tion system after attribute reduction remains unchanged, multiple
set reduction, and then collected data from a laboratory building of
minimum relative reduction results was considered.
a university in Dalian for nearly a year to train and test deep neural
networks. The following sections introduces the rough set theory,
deep learning, and the prediction process in sequence.
2.2. Deep learning

2.1. Rough set theory Deep learning is a branch of machine learning. The basic struc-
ture of deep learning is deep neural networks. The difference
The rough set theory is mainly used for learning and induction between deep neural networks and shallow neural networks is
of incomplete data and uncertain knowledge [34]. It is good at that shallow neural networks generally have only two to three lay-
mining the potential relationship between knowledge and finding ers of neural networks. Therefore, shallow neural network has lim-
the key attributes, which make it very suitable for extracting the ited ability to express complex functions. However, deep learning
important influencing factors of building energy consumption. has five or more layers of neural networks and introduces more
This study defined S ¼ ðU; A; K; fÞ as a knowledge expression effective algorithms [36] that can further improve the accuracy
system to classify objects. U = {x1, x2, x3. . .. . .xn} is the domain, of building energy consumption prediction. In this study, the deep
3
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

C1 C2 C20

Rough sets Neural networks

D
Z1 Z2 Zm

Short-term prediction of
Neural networks Data processing building energy consumption
without integration of rough set
and deep neural network

T1 T2 Tm
Short-term prediction of
building energy consumption
based on rough set and deep
Neural networks
neural network

Medium-term prediction of
building energy consumption
based on rough set and deep
neural network

Fig. 1. Research outline.

belief neural network (DBN), which is commonly used in deep shown in Fig. 2. There are no links between the nodes within each
learning, is used to predict building energy consumption. layer. Layer V represents the input data that is visible and layer H is
a hidden layer. In RBM, nodes are random binary variable nodes
(with values of only 0 or 1), and the full probability distribution
2.2.1. Structure of the DBN p (V, H) satisfies the Boltzmann distribution. Different from the tra-
The DBN is made up of multiple restricted Boltzmann machine ditional feed-forward neural network, the link between the visible
(RBM) stacks. The structure of the RBM can be represented by the layer and the hidden layer of the RBM is uncertain and fully linked.
bipartite graph including the visible layer and the hidden layer, as In other words, the values between layers can carry out the two-
way propagation from the hidden layer to the visible layer and vice
versa.
Offset of hidden nodes: c c i1 c i2 c ik The above RBM structure has n visible nodes and m hidden
nodes. Each visible node is only related to m hidden nodes. The
Hidden layer: H Hi
1
Hi
2
Hi
k
nodes in the same layer are independent. This feature makes the
training process of RBM easier. RBM has several parameters, which
1 2
are weight matrix Wmn, the offset of the visible node bi= (bi , bi ,
j
Weight matrix: Wmhn . . ., bi ), and the offset of the hidden node ci= (c1i , c2i , . . ., cki ). These
parameters determine how the RBM network encodes an n-
dimensional input into a m-dimensional output.
RBMs are stacked like bricks to build a network, and this net-
1 2 3 j work is called a DBN. A typical DBN network is shown in Fig. 3.
Visible layer: V Vi Vi Vi Vi
These RBMs in the DBN network have only one visible layer and
one hidden layer, and while the layers are connected to each other.
Offset of visible node: b b i1 b i2 b i3 b ij The W, b, and c of each hidden layer will be trained to represent
high-level features.
Fig. 2. Structure of RBM model.

4
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Output data on building 2.2.2. Training process of DBN

energy consumption The training process of an RBM is to determine the probability
distribution with the highest probability of generating training
samples. Since the decisive factor for the probability distribution
is the weight W, the goal of training an RBM is to find the best
weight. This investigation adopted the greedy layer-wise training
W2 Back propagation
method to train the RBMs one by one.
V2
The training process of the DBN is mainly divided into two
H1 parts: pretraining and fine-tuning. In the first part, unsupervised
RBM W1 Fine-tuning and spontaneous training is performed on each RBM to ensure that
V1 feature information is retained as much as possible when feature
H0 vectors are mapped to different feature spaces. The input layer of
the first RBM is the input layer of the entire neural network. The
RBM W0 Fine-tuning
hidden layer Hi can be regarded as the visible layer of Hi+1, which
is denoted as Vi+1. The pretraining of the greedy layer-wise training
V0 method involves the following steps.

Input data on impact factors of
Step 1: Train the first RBM for all training samples.
building energy consumption
Step 2: Fix the weight and offset of the first RBM, and then train
Fig. 3. Structure and training process of DBN.
the second RBM with V1 = H0.

Raw data

Discretization of continuous data

Establish building energy

consumption decision table

Rough set attribute reduction

Rosetta
operation Optimal sample set of building energy
Test sample
consumption data without redundant attribute

Training samples

Initialize the three parameters

of RBM network: W, b and c

Build Multiple RBM networks to form the DBN

model and enter the pre-training of deep learning

Train separately Each layer of

RBM network with unsupervised

Set the BP classifier in the

Fine-tune the entire DBN
MATLAB last layer of DBN
operation
Training enters the back- Train entity relationship
propagation stage to find classifiers with Supervision
the optimal weight

No
Whether the
weights are optimal

Yes
Determine the training model for DBN

Building energy consumption prediction

The test accuracy is compared with the BP neural

network test model, the fuzzy neural network test
model and the Elman neural network test model.

Discuss forecast results

Fig. 4. Workflow of prediction model for building energy consumption based on rough set and deep learning.

5
 L. Lei, W. Chen, B. Wu et al.
Table 1
Raw data used for attribute reduction in rough sets.

Sample Outdoor Relative Wind Solar Number Building Building Window- Heat Shading Building Heat Lighting Personnel Indoor Fresh COP Supply air Fan Pump Energy
temperature humidity speed radiation of floors area orientation to-wall transfer coefficient length- transfer power density temperature air of the temperature efficiency efficiency consumption
area coefficient to-width coefficient density volume chiller
ratio of external ratio of roof per
walls capita
°C % m/s W/m2 m2 degree % W/(m2K) W/(m2K) W/m2 person / °C m3/h. °C % % kWh
m2 person
1 25.59 86.8 0.83 155.56 4 3520 234 16.3 0.84 0.68 1.51 0.56 7.86 0.25 24.18 28 6.31 20.4 0.66 0.75 39.6
2 25.49 80.1 0.56 310.2 5 5630 263 18.2 0.65 0.59 1.65 0.74 4.69 0.65 24.75 29 6.75 21 0.56 0.56 36.8
3 25.25 79.6 1.20 382.4 4 4856 256 26.8 0.87 0.68 1.85 0.86 11.65 0.99 23.33 36 6.4 21.3 0.59 0.52 88
4 25.05 79.7 0.42 410.84 6 4586 281 17.5 0.96 0.49 1.64 0.94 13 1.11 24.78 25 6.17 21.7 0.85 0.59 67.6
5 24.67 82.1 0.85 390.41 5 7854 256 20.3 0.70 0.95 1.95 0.65 15.92 1.10 23.28 25 5.87 21.9 0.89 0.8 58
6 24.48 80.9 0.80 329.08 7 8561 269 21.2 1.36 0.87 2.64 0.74 6.94 0.48 23.69 28 5.61 22.1 0.90 0.56 68
7 24.98 80.2 0.59 228.42 6 8463 249 18.9 1.29 0.56 2.59 0.60 6.94 0.69 24.21 23 5.61 22.3 0.57 0.59 37.6
8 29.1 80.5 0.34 102.78 5 8564 268 14.6 1.45 0.95 3.15 0.98 8.95 0.59 26.88 37 5.47 22.5 0.67 0.5 70.8
9 29.37 81.1 0.23 266.66 6 8856 235 15 0.71 0.68 1.65 1.00 12.6 0.30 28.98 26 5.47 22.5 0.58 0.68 68.8
6

10 29.58 86.6 0.62 368.47 4 4685 268 17.5 0.85 0.84 2.98 0.94 12.59 0.54 28.62 25 5.54 17.6 0.59 0.8 97.6
11 30.41 90.1 0.38 403.4 6 3698 264 18.8 0.64 0.50 2.36 0.56 12.96 0.98 29.43 23 5.38 22.8 0.81 0.81 65.2
12 31.22 92.3 0.93 394.32 7 8463 284 17.9 0.69 0.59 1.78 0.59 11.59 0.85 29.69 31 5.29 22.8 0.86 0.82 66.4
13 32.06 94.5 0.57 341.07 5 6584 256 20.1 0.84 0.54 3.01 0.83 9.65 0.76 30.10 27 5.23 22.9 0.79 0.76 73.2
14 32.51 95.1 1.04 246.86 6 7569 248 20.3 0.91 0.68 1.99 0.97 9.05 0.29 30.57 25 5.11 19 0.90 0.79 65.2
15 32.93 97.3 0.93 118.82 4 8465 268 23.4 1.25 0.79 2.64 0.76 9.06 1.02 30.22 23 5.18 18.1 0.81 0.72 96
16 33.52 95.9 0.77 309.94 6 4869 269 16.8 1.24 0.76 3.16 1.00 9.00 1.13 31.29 29 5.2 17.2 0.55 0.61 95.6
17 32.83 96 0.38 277.78 7 5896 281 19.5 1.34 0.71 2.46 0.99 5.92 0.95 30.77 28 5.14 17.8 0.59 0.69 86.8
18 31.1 97.4 1.16 361.12 4 5963 264 17.5 0.94 0.81 2.59 0.94 5.62 0.98 28.05 26 5.14 20.4 0.55 0.59 103.6
19 30.01 96.2 0.47 555.56 5 3874 256 18.8 1.35 0.94 2.76 0.96 6.95 1.01 29.15 23 4.95 20.6 0.88 0.54 121.8
... ... ...
100 29.87 91.8 0.92 538.88 6 7658 265 17.6 0.62 0.93 2.54 0.98 8.32 1.09 26.84 25 5.1 19.4 0.81 0.5 98.4

Energy & Buildings 240 (2021) 110886

L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Step 3: Repeat the previous two steps for all RBMs. 3. Data collection

In the second part, the fine-tuning set the classifier model for The field measurements of data had two parts. The first part
the last layer of the DBN. In this study, the classifier model is set was the data collected of 100 civil public buildings for reducing
to the commonly used BP network. By receiving the output feature the inputs of building energy consumption (attributes) by rough
vector of the RBM as its input feature vector, the entity relationship sets. The second part was the data collected of a laboratory build-
classifier is trained with supervision. Each layer of the RBM net- ing for training and testing the deep neural network.
work can only ensure that the weights in its own layer are opti-
mized for the feature vector mapping of the layer but not for
feature vector mapping of the entire DBN. Therefore, the reverse 3.1. Data collection for attribute reduction in rough sets
propagation network also propagates the error information from
top to bottom of each layer of the RBM to fine-tune the DBN net- In this study, civil public buildings were used as experimental
work. The pretraining process of the RBM can be regarded as the objects. Civil public buildings refer to non-residential buildings
initialization the of weight parameters of a deep BP network, for people to carry out various public activities, generally including
which facilitates the DBN to overcome the problem of the BP net- office buildings, commercial buildings, tourist buildings, science,
work’s tendency to fall into a local optimum and the long training education, culture and health buildings, and transportation build-
time due to the random initialization of weight parameters. ings. A total of 20 inputs Ci (i = 1, 2, . . ., 20) that represent building
energy impact factors and the building energy consumption were
collected. For example, the outdoor temperature, relative humidity
and wind speed were measured by hot-wire anemometer TES-
2.3. Prediction process of the model 1341. The solar radiation was measured by solar radiation sensor
HOBO S-LIB-M003. The solar radiation refers to the solar radiation
The main workflow of the building energy consumption predic- outside the building in the natural state. Its value is the total solar
tion is shown in Fig. 4. This study firstly determined the discrete radiation intensity, including the direct solar radiation intensity
breakpoints of each input and discretized the continuous raw data. and the solar scattered radiation intensity. The occupancy was esti-
Through attribute reduction by rough set, the important inputs for mated by measuring the concentration of CO2 in the building with
building energy consumption was obtained. Then the data sample a carbon dioxide detector Telaire-7001. The supply air temperature
set was divided into training data and test data. The training data of HVAC units was measured by air temperature and humidity sen-
was used for DBN model training, and the test data was used for sor HOBO UX100-003. The building energy consumption was
verifying the model accuracy. The training of the DBN model obtained by real-time monitoring. Building energy consumption
started by initializing the three parameters W, b, and c of the was determined by electricity consumption, which considered
RBM. This investigation then performed unsupervised pre- the energy used by the HVAC system, lighting fixtures, and electri-
training and unsupervised fine-tuning of the DBN. In this way, cal appliances.
the optimal weights for mapping the entire DBN were obtained. The data collection were conducted on 100 civilian public build-
In this study, Rosetta [37], a rough set software, is used to ings in Dalian during the period from May 2017 to September
reduce the attributes in the rough set decision table formed by 2017, and one set of test data on building energy consumption
the sample data. Rosetta is a tabular logical data tool based on and the influencing factors of building energy consumption were
the rough set theory framework. It is often used for the algorithmic collected for each building. The test period for each set of data
acquisition of the rules and attribute reduction in rough set theory. was one hour. Some of the data are shown in Table 1. Among them,
The attribute reduction method based on the genetic algorithm the indoor temperature refers to the average of measured air tem-
provided in the software is selected to reduce the 20 inputs of perature in the room over one hour. The heat transfer coefficient of
building energy consumption. The remaining conditional attri- the exterior wall is the average heat transfer coefficient of all exte-
butes after reduction are the corresponding important inputs of rior walls. The shading coefficient refers to the external shading
building energy consumption. The establishment, training, and test coefficient, which is the ratio of the solar radiation heat gained
of the neural networks were performed in MATLAB [38]. through an external window with an external shading structure

Fig. 5. Hourly energy consumption of the laboratory building.

7
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

(a) (b)

(c) (d)
Fig. 6. Hourly meteorological data of Dalian throughout the year (a) air temperature; (b) relative humidity; (c) solar radiation; (d) wind speed.

Fig. 7. Daily energy consumption of the laboratory building.

to that through the same external window without an external took the shading structure size ratio  as a parameter to calculate
shading structure. This study adopted a simplified formula and the external shading coefficient by SD = ax2 + bx + 1, where a and b

8
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

(a) (b)

(c) (d)
Fig. 8. Daily meteorological data of Dalian throughout the year (a) Daily outdoor average temperature; (b) Daily average humidity; (c) Daily solar radiation; (d) Daily average
wind speed.

are regression coefficients [39]. The values of a and b are depen- from the 1000th hour to the 2000th hour was spring, and the lab-
dent on the climate zone where the building is located, the basic oratory building had low energy consumption. The season around
type of external shading, and the orientation of the windows. the 4000th to 5000th hour was summer, which is the period that
the laboratory building consumes the most energy throughout
3.2. Data collection for the deep neural network the year. Please note that the energy consumption refers to the
electrical energy consumption, the energy consumption by the
After reducing the number of inputs by rough sets, the effective central heating system in winter was no considered. Fig. 6 shows
inputs Zi (i = 1, 2, . . ., m; m  20) that determines the building the corresponding hourly meteorological data of Dalian, which is
energy consumption are obtained. This investigation collected data part of the input data. Among them, the period around 4000th to
including the effective inputs Zi and building energy consumption 5000th hour was summer, so the air temperature, relative humid-
in a laboratory building of a university in Dalian. The laboratory ity and solar radiation were relatively high.
building is a frame-shear structure, with five floors above ground The measured hourly data throughout the year is for short-term
and one floor underground, with a total construction area of forecast of building energy consumption. The hourly data were fur-
9533 m2 and an air-conditioning area of 8579.7 m2. ther processed into daily data for the medium-term forecast of
Fig. 5 shows 8176 h of hourly building energy consumption of building energy consumption. For example, the summation of the
the laboratory building from January 1, 2018 to December 7, hourly energy consumption in a day is the daily energy consump-
2018. The dots lie in the two shaded areas represent the two peri- tion and average of the hourly outdoor air temperature in a day is
ods (168 h each) for testing the short-term predictions. The period the daily outdoor air temperature. Fig. 7 shows the daily energy
9
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Table 2
Discrete breakpoints of each attribute.

Inputs (influence factors of building energy consumption) Label Minor impact (1) Slight impact (2) Moderate impact (3) Significant impact (4)
Outdoor temperature C1 <=26 (26,29) (29,32) >32
Relative humidity C2 <=60 (60,75] (75,90] >90
Wind speed C3 >1.6 (1.6,1.3] (1.3,1.0] <=1
Outdoor solar radiation illuminance C4 <=150 (150,300] (300,450] >450
Number of floors C5 <=3 (3,5] (5,7] >7
Building area C6 <=3000 (3000,6000] (6000,9000] >9000
Building orientation (general orientation: south) C7 <=240 (240,270] (270,300] >300
Window-to-wall area ratio C8 <=0.15 (0.15,0.2] (0.2,0.25] >0.25
Heat transfer coefficient of external walls C9 <=0.8 (0.8,1.1] (1.1,1.4] >1.4
Shading coefficient C10 <=0.5 (0.5,0.65] (0.65,0.8] >0.8
Building length-to-width ratio C11 <=1.5 (1.5,2.5] (2.5,3.5] >3
Heat transfer coefficient of roof C12 <=0.6 (0.6,0.75] (0.75,0.9] >0.9
Lighting power density C13 <=6 (6,9] (9,12] >12
Personnel density C14 <=0.4 (0.4,0.7] (0.7,1.0] >1.0
Indoor temperature C15 <=25 (25,29] (29,32] >32
Fresh air volume per capita C16 <=30 (30,40] (40,50] >50
COP of the chiller C17 >6 (5,6] (4,5] (0.7,1.0]
Supply air temperature C18 >24 (21,24] (18,21] (29,32]
Fan efficiency C19 >0.8 (0.7,0.8] (0.6,0.7] <=0.6
Pump efficiency C20 >0.8 (0.6,0.8] (0.4,0.6] <=0.4

Table 3 as the inputs of the deep neural network and the building energy
The levels of energy consumption. consumption is the output of the deep neural network. The data
Rank sequence Characterization state Energy consumption collected in section 3.2 was used for training and testing the deep
description neural network.
1 Low energy consumption <=50
2 Moderate energy consumption (50,80]
3 High energy consumption (80,110] 4.1. Identification of effective inputs for building energy consumption
4 Ultra-high energy consumption >110

Since rough sets cannot process continuous data, the original

consumption of 341 days. The shaded part represents the 30-day data used for attribute reduction in the rough sets must be dis-
test period for mid-term prediction of building energy consump- cretized first. The equidistance division is used as the ideological
tion. Fig. 8. Shows the daily meteorological data of Dalian through- basis of the discretization process. Combined with the basic char-
out the year. acteristics of the data and the degree of centralization, the discrete
points of the individual influencing factors of building energy con-
sumption are slightly adjusted. The influencing factors of building
4. Results energy consumption are used as energy consumption indicators.
The value range of each attribute is discretized into four intervals,
In order to identify the effective inputs for the building energy coded as 1, 2, 3, and 4, which indicate that the input has minor,
consumption, this study first applied the rough set with the data slight, moderate, and significant impact on the energy consump-
collected in section 3.1. The determined effective inputs are used tion, respectively. C1, C2, C3, . . ., C20 represent the 20 influencing

Table 4
Decision table of building energy consumption in the laboratory building.

Sample Conditional attributes (inputs) C Decision attribute (output) D

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15 C16 C17 C18 C19 C20 Decision value
1 1 3 4 2 2 2 1 2 2 3 2 1 2 1 1 1 1 3 3 2 1
2 1 3 4 3 2 2 2 2 1 2 2 2 1 2 1 1 1 3 4 3 1
3 1 3 3 3 2 2 2 4 2 3 2 3 3 3 1 2 1 2 4 3 3
4 1 3 4 3 3 2 3 2 2 2 2 4 4 4 1 1 1 2 1 3 2
5 1 3 4 3 2 3 2 3 1 4 2 2 4 4 1 1 2 2 1 2 2
6 1 3 4 3 3 3 2 3 3 4 3 2 2 2 1 1 2 2 1 3 2
7 1 3 4 2 3 3 2 2 3 2 3 1 2 2 1 1 2 2 4 3 1
8 3 3 4 1 2 3 2 1 3 4 3 4 2 2 2 2 2 2 3 3 2
9 3 3 4 2 3 3 1 1 1 3 2 4 4 1 2 1 2 2 4 2 2
10 3 3 4 3 2 2 2 2 2 4 3 4 4 2 2 1 2 4 4 2 3
11 3 4 4 3 3 2 2 2 1 1 2 1 4 3 3 1 2 2 1 1 2
12 3 4 4 3 3 3 3 2 1 2 2 1 3 3 3 2 2 2 1 1 2
13 3 4 4 3 2 3 2 3 2 2 3 3 3 3 3 1 2 2 2 2 2
14 3 4 3 2 3 3 2 3 2 3 2 4 3 1 3 1 2 3 1 2 2
15 3 4 4 1 2 3 2 3 3 3 3 3 3 4 3 1 2 3 1 2 3
16 3 4 4 3 3 2 2 2 3 3 3 4 2 4 3 1 2 4 4 2 3
17 3 4 4 2 3 2 3 2 3 3 2 4 1 3 3 1 2 4 4 2 3
18 3 4 3 3 2 2 2 2 2 4 3 4 1 3 2 1 2 3 4 3 3
19 3 4 4 4 2 2 2 2 3 4 3 4 2 4 3 1 3 3 1 3 4
... ...
100 3 4 4 4 3 3 2 2 1 4 3 4 2 4 2 1 2 3 1 3 3

10
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Table 5 The 20 inputs for building energy consumption C1, C2, C3, . . ., C20
Reduction of conditional attributes. are used as the 20 conditional attributes, and the conditional attri-
Serial Minimal relative reduction set (ie Strength Number bute values are the four coding level values in Table 2. The amount
number remaining after reduction) of energy consumption is taken as a decision attribute, and the
1 {C4, C13, C15} 100 3 decision value is the sequence value of the four energy consump-
2 {C9, C12, C19} 100 3 tion grades determined according to the energy consumption in
3 {C8, C9, C10, C14} 100 4 Table 3. This study considered a data sample as an object and con-
4 {C8, C9, C15, C19} 100 4
5 {C9, C10, C17, C20} 100 4
dition variables (the inputs) jointly determined the decision vari-
6 {C10, C12, C15, C18} 100 4 able (building energy consumption or output). Based on the data
7 {C9, C15, C17, C19} 100 4 collected in Table 1, the decision table for building energy con-
8 {C4, C14, C17, C18} 100 4 sumption is shown in Table 4.
9 {C2, C9, C19, C20} 100 4
After the attribute reduction operation, a total of 28 reductions
10 {C8, C12, C14, C17} 100 4
11 {C10, C12, C15, C19} 100 4 with a strength of 100 are obtained, as shown in Table 5. From the
12 {C1, C9, C19, C20} 100 4 reduction results, C1, C2, C4, C7, C8, C9, C10, C12, C13, C14, C15, C17, C18,
13 {C14, C15, C19, C20} 100 4 C19, and C20 repeatedly appear in the multiple minimum relative
14 {C1, C9, C10, C14} 100 4 reduction results, which shows that these 15 conditional attributes
15 {C9, C12, C15, C18} 100 4
(reassigned as Z1, Z2, . . ., Z15) are the core of the decision table.
16 {C10, C14, C15, C19} 100 4
17 {C2, C9, C10, C18} 100 4 Therefore, through the attribute reduction of the rough set, five
18 {C4, C7, C17, C18} 100 4 miscellaneous and unimportant inputs of building energy con-
19 {C2, C9, C12, C20} 100 4 sumption, including wind speed, number of floors, building area,
20 {C1, C14, C17, C19} 100 4
building length-to-width ratio and fresh air volume per capita,
21 {C2, C7, C9, C10} 100 4
22 {C1, C8, C9, C18} 100 4 are eliminated.
23 {C4, C7, C12, C17} 100 4
24 {C2, C4, C10, C17, C19} 100 5
25 {C2, C8, C9, C10, C13} 100 5
4.2. Prediction of building energy consumption
26 {C1, C4, C9, C14, C18} 100 5
27 {C7, C9, C10, C13, C14} 100 5 In the short-term prediction of building energy consumption,
28 {C2, C4, C8, C9, C17} 100 5 there are 8176 sets of original data, which are divided into two
parts: test data and training data. Among them, 336 sets are used
as test data for two test periods in the low-energy season and
factors of building energy consumption. The specific division of the high-energy season, and the rest are used as training data. In the
discrete points of each factor is shown in Table 2. medium-term prediction of building energy consumption, a total
The original domain of the 100 power consumption datasets of 341 sets of data are obtained, 30 sets are used for testing, and
used for attribute reduction of the rough set is discretized into four 311 sets are used for training. For comparison, this study also
energy consumption levels, as shown in Table 3. Less than 50 kwh trained and tested the BP neural network, Elman neural network
is low energy consumption, which is represented by level 1. From and fuzzy neural network (FNN). This study used trial and error
50 kwh to 80 kwh is medium energy consumption, which is repre- method to find the optimal structure and parameters of these three
sented by level 2. From 80 kwh to 110 kwh is higher energy con- neural networks. For example, the optimal structure of BP neural
sumption, which is represented by level 3. >110 kwh is high network was a double hidden layer neural network. The number
energy consumption, which is represented by level 4. of neuron nodes in the first hidden layer was 10, and the number
of neuron nodes in the second hidden layer was 14.

Fig. 9. The prediction of building energy consumption by DBN and RS-DBN for (a) low energy consumption season and (b) high energy consumption season.

11
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

(a) (b)
Fig. 10. The prediction error of DBN and RS-DBN for (a) low energy consumption season; (b) high energy consumption season.

Table 6
Quantitative evaluation of DBN model and RS-DBN model.

Prediction model Low energy consumption season High energy consumption season
MAPE RMSPE MAPE RMSPE
DBN 0.0552 0.0515 0.0484 0.0498
RS-DBN 0.0388 0.0347 0.0291 0.0265

Fig. 11. Comparison of short-term prediction of building energy consumption with the experimental data for the laboratory building in summer.

4.2.1. Short-term prediction of building energy consumption without consumption for the laboratory building during the low-energy
the rough set season and the high-energy season. Fig. 9 shows the prediction
This section investigated the necessity of integrating rough set of the DBN model without rough set and the RS-DBN model.
with neural network. Without using rough sets as the front-end Fig. 10 shows the error of the DBN without rough set and RS-
processor of the prediction model, the 20 inputs of building energy DBN in the two time periods. The results show that the prediction
consumption were directly used as the inputs of the four neural performance of the RS-DBN model is better than that of the DBN
networks. The test used short-term prediction of building energy model without the rough set.

12
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Fig. 12. Comparison of short-term prediction of building energy consumption with the experimental data for the laboratory building in summer from the 60th hour to the
72nd hour.
Absolute error (kWh)

Hours (h)
Fig. 13. The absolute error of the four models in short-term prediction of building energy consumption in the summer.

13
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Fig. 14. Prediction results of four models for short-term prediction of building energy consumption in low energy consumption season of the laboratory building.

By calculating average absolute relative error (MAPE) and root clearer and more intuitive comparison between the predictions
mean square relative error (RMSPE), and experimental data, Fig. 12 presents the results from the 60th
  hour to the 72nd hour. The comparison shows that the predicted
1 Xn yq  yp  energy consumption of DBN has the highest agreement with the
MAPE ¼ 
i¼1 
 ð1Þ
n yq  actual energy consumption.
Fig. 13 shows the overall performance of the four models by cal-
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
1 1 Xn culating the absolute error in the short-term prediction of building
RMSPE ¼ i¼1 q
ðy  yp Þ2 ð2Þ energy consumption in the summer. The Elman neural network
yq n
and fuzzy neural network, the deep neural network is more accu-
the prediction accuracy of DBN model and RS-DBN model was rate than the BP neural network in predicting the energy consump-
quantitatively evaluated. Among them, yq and yp represent actual tion of the laboratory building in the summer season.
value and predicted value, respectively, and yq represents the aver- Spring is representative of the low energy consumption season
age of actual value. The calculation results are shown in Table 6. of the laboratory building. From March 20 to March 26, short-term
For both seasons, the calculated MAPE and RMSPE of the RS- prediction of building energy consumption was carried out by the
DBN model were smaller than the DBN model. It indicates that four neural networks as shown in Fig. 14. The results show that the
without rough set participation, the prediction of building energy prediction of FNN deviated a lot with the experimental data, while
consumption by a neural network alone has a lower prediction the predictions of other three neural networks agreed well with
accuracy than that by the integration of the rough set and a neural the experimental data. The major difference between the spring
network. Therefore, the following tests used the integrated rough and summer in terms of energy consumption is the usage of air-
set and neural networks. conditioning system. This comparison confirmed that DBN model
can accurately accomplish short-term prediction of building
4.2.2. Short-term prediction of building energy consumption with the energy consumption with/without air-conditioning system.
rough set The experimental results from the 60th hour to the 72nd hour is
This study conducted the prediction of building energy con- enlarged in Fig. 14 to further compare the prediction effects of the
sumption in the summer, which is the season with high energy RS-BP, RS-SRN and RS-DBN in the low energy consumption season.
consumption in the laboratory building, and spring, which is the The comparison result is shown in Fig. 15. The predicted energy
season with low energy consumption. After rough sets was used consumption curve of the RS-DBN has the highest fitting degree
to reduce the influencing factors of building energy consumption, with the actual energy consumption curve, and its prediction effect
15 important influencing factors of building energy consumption is the best. This result indicates that the RS-DBN prediction model
remained. These remaining factors were used as the inputs of the is still more advantageous than other models in short-term predic-
deep belief neural network, and hourly building energy consump- tion of building energy consumption in the low energy consump-
tion was output. tion season of the laboratory building.
From June 23 to June 29 (a total of 168 h, summer), the hourly Fig. 16 is a comparison of the overall error performance of the
building energy consumption for the laboratory building is pre- short-term prediction of building energy consumption by the four
dicted. Fig. 11 compares the predicted hourly energy consumption prediction models during the low energy consumption season of
with the measured one of the laboratory building. The trend of the the laboratory building. The area surrounded by the absolute value
predicted energy consumption of the four neural networks is gen- curve of the error of the RS-DBN model and the X axis is mainly
erally consistent with that of the measured data, which shows that covered by the area representing the errors of the other three mod-
the combination of the rough set and neural network to predict els. The overall error performance of the RS-DBN model is the best,
building energy consumption is effective and feasible. To make and its prediction error for most test samples is at a minimum. In

14
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Fig. 15. Partial results of short-term prediction of building energy consumption in low energy consumption season of the laboratory building: (a) RS-DBN; (b) RS-BP; (c) RS-
SRN; (d) RS-FNN.

general, the accuracy performance of the four models in the low To facilitate a clear comparison of the prediction accuracy of the
energy consumption season is roughly similar to that in the high four models on the medium-term prediction of building energy
energy consumption season. The RS-DBN model can accurately consumption, a needle chart is used to analyze the absolute error
and effectively predict short-term building energy consumption of each model. The absolute errors of the four models for each test
under the influence of the air conditioning system factors to vary- data are shown in Fig. 18. Among them, at most test sample points,
ing degrees. the length of the needles representing the RS-BP or RS-FNN is the
longest and the fluctuation range is large. The needle length of
the RS-DBN model is basically the shortest and its range of fluctu-
ation is small. The RS-DBN model has the smallest absolute error in
4.2.3. Medium-term prediction of building energy consumption with the mid-term prediction of building energy consumption and its
the rough set error has the least variations. Table 7 further shows the relative
The suitability of the RS-DBN prediction model is also investi- error of the RS-DBN model for the mid-term prediction of building
gated for mid-term prediction of building energy consumption. energy consumption. The maximum relative error of the RS-DBN
The hourly energy consumption data of the laboratory building model does not exceed 7%. Therefore, the RS-DBN model is not only
are processed into daily energy consumption data, and a 30-day suitable for short-term prediction of building energy consumption,
mid-term prediction of building energy consumption is conducted but also accurate and effective for mid-term prediction of building
in November (during winter). The prediction results of the four energy consumption.
models of the RS-DBN, RS-BP, RS-SRN and RS-FNN for daily build-
ing energy consumption in November are shown in Fig. 17. The
predictions by the RS-DBN and RS-SRN have a high degree of fitting 5. Discussions
with the actual energy consumption. Therefore, the RS-DBN and
RS-SRN are more suitable than the RS-BP and RS-FNN for mid- This study used considerably high percentage of data for train-
term prediction of building energy consumption. ing because it is dependent on the nature of the deep neural net-
15
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Absolute value of error (kWh)

Hours (h)
Fig. 16. The overall error performance of the four models in the prediction experiment of low energy consumption season.

Fig. 17. The prediction results of the four models in the medium-term prediction experiment of building energy consumption.

work. The main training method of deep belief neural network is civil public buildings. In terms of schedule, the open time of those
unsupervised learning, which requires a large amount of training buildings are generally from 7:00 am to 10:00 pm. In terms of
data for the algorithm itself to extract the characteristics of the building energy consumption, the laboratory building is similar
data. In contrast, the shallow neural network generally uses super- with other civil public buildings as most of its energy is consumed
vised learning for training method, which requires less amount of by the HVAC system, computers, experimental equipment, and
training data. For example, the two studies [25] and [21] used con- lighting. Another reason is that the required data for rough set the-
siderably high percentage of data for training. Li et al. [21] used ory is its diversity but not quantity. Therefore, this study collected
>90% of the data for training. the data of 100 civil public buildings for only one hour each. How-
One can notice that this study deducted the relevant input vari- ever, the data used for training and testing the deep neural net-
ables by using the dataset of public buildings and applied the work requires long-term monitoring, the data collection is
results for a laboratory building in training and testing the deep difficult in those civil public buildings. For a laboratory building
neural network. One reason is that the civil buildings refer to located in a university, the long-term monitoring of data in a con-
non-residential buildings where people carry out various public trollable range could be realized.
activities, such as office buildings, commercial buildings, tourist The ‘‘feature” in this study refers to the implicit connections
buildings, science, education, culture and health buildings, and between the input and the output data of the deep neural network.
transportation buildings, etc. The laboratory building belongs to These connections will directly affect the output information after

16
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

Error (kWh)

Fig. 18. The absolute errors of the four models in the medium-term prediction experiment of building energy consumption.

Table 7
Relative error analysis of RS-DBN model in the medium-term prediction experiment of building energy consumption.

Days Actual energy consumption value (kwh) RS-DBN predicted energy Relative error of RS-DBN model in
consumption value (kwh) the medium-term prediction (%)
Day 1 1407.2 1419.8 0.90%
Day 2 1394 1391.4 0.19%
Day 3 1087.6 1102.1 1.33%
Day 4 1097.2 1117.9 1.89%
Day 5 1367.6 1382.2 1.07%
Day 6 1489.2 1485.6 0.24%
Day 7 1738 1787.8 2.87%
Day 8 1581.6 1559.2 1.42%
Day 9 1455.6 1459.2 0.25%
Day 10 1197.6 1208.9 0.94%
Day 11 1192.4 1215.3 1.92%
Day 12 1652 1740 5.33%
Day 13 1621.2 1621.1 0.01%
Day 14 1650.4 1737.4 5.27%
Day 15 1778.4 1811.1 1.84%
Day 16 1665.2 1751.1 5.16%
Day 17 1537.6 1528.6 0.59%
Day 18 1535.2 1524.9 0.67%
Day 19 1730 1782.9 3.06%
Day 20 1489.6 1493 0.23%
Day 21 1633.6 1734.2 6.16%
Day 22 1638.4 1736.8 6.01%
Day 23 1506.8 1491.9 0.99%
Day 24 1285.2 1275.8 0.73%
Day 25 1337.6 1340.4 0.21%
Day 26 1394.8 1396.6 0.13%
Day 27 1595.2 1589.5 0.36%
Day 28 1594 1585.4 0.54%
Day 29 1650.4 1736.7 5.23%
Day 30 1712.4 1777.9 3.83%

passing through layer-by-layer and bottom-up unsupervised learn- 6. Conclusions

ing. This study did not discuss the actual feature extraction
because deep learning used an algorithm for learning features This paper integrated the rough sets and deep learning to estab-
automatically. The training method of the deep neural network lish a prediction model for building energy consumption. The
was unsupervised. It used uncalibrated data to train the parame- model uses rough set theory as the front-end processor for deep
ters of each layer spontaneously and does not require artificial learning. The attribute reduction method based on the genetic
intervention in the process of extracting features. algorithm in rough set theory is used to reduce the miscellaneous

17
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

influencing factors of building energy consumption. The model [6] G. Qiang, T. Zhe, D. Yan, Z. Neng, An improved office building cooling load
prediction model based on multivariable linear regression, Energy Build. 107
uses a deep neural network as the post information recognition
(2015) 445–455.
system. The remaining influencing factors of building energy con- [7] H. Pombeiro, R. Santos, P. Carreira, et al., Comparative assessment of low-
sumption after rough set reduction were used as the input of the complexity models to predict electricity consumption in an institutional
DBN, and the building energy consumption are output from the building: Linear regression vs. fuzzy modeling vs. neural networks, Energy
Build. 146 (2017) 141–151.
DBN. This investigation tested the developed model by comparing [8] M.A. Jallal, A.G. Vidal, A.F. Skarmeta, A hybrid neuro-fuzzy inference system-
its performance with BP, Elman and fuzzy neural networks in both based algorithm for time series forecasting applied to energy consumption
short-term and medium-term predictions of building energy con- prediction, Appl. Energy 268 (2020) 114977, Energy Build. 111 (2016) 380-392.
[9] Z.W. Li, Y.M. Han, P. Xu, Methods for benchmarking building energy
sumption. The results led to the following conclusions: consumption againstits past or intended performance: an overview, Appl.
Energy 124 (7) (2014) 325–334.

The implementation of rough set theory was able to eliminate [10] J. Braun, N. Chaturvedi, An inverse gray-box model for transient building load
prediction, HVAC&R Res. 8 (1) (2002) 73–99.
redundant influencing factors of building energy consumption. [11] B. Dong, C. Cao, S.E. Lee, Applying support vector machines to predict building
The DBN with reduced number of inputs had improved accu- energy consumption in tropical region, Energy Build. 37 (5) (2005) 545–553.
racy in building energy simulation. [12] L. Robert, R. Uwe, Comparison between support vector regression and grey
system theory based models for adaptive load forecasting of building thermal

The DBN had more accurate prediction of either short-term or energy consumption, EG-ICE 2014, European Group for Intelligent Computing
long-term building energy consumption than the shallow neu- in Engineering-21st International Workshop: Intelligent Computing in
ral networks such as BP, Elman and fuzzy neural networks. Engineering 2014. 2014.
[13] J.J. Guo, J.Y. Wu, R.Z. Wang, A new approach to energy consumption prediction

This study presented a procedure to make use of machine learn-
of domestic heat pump water heater based on grey system theory, Energy
ing for building energy simulation. The developed prediction Build. 43 (6) (2011) 1273–1279.
model has great potential to predict the energy consumption [14] Y. Wei, L. Xia, S. Pan, J. Wu, X. Zhang, M. Han, W. Zhang, J. Xie, Q. Li, Prediction
of various types of buildings on different time scales. It can of occupancy level and energy consumption in office building using blind
system identification and neural networks, Appl. Energy. 240 (2019) 276–294.
bring more accurate prediction results, which can enable energy [15] B.B. Ekici, U.T. Aksoy, Prediction of building energy consumption by using
demand managers to better schedule and control the fluctuat- artificial neural networks, Adv. Eng. Softw. 40 (5) (2009) 356–362.
ing power supply for demand management continuously in real [16] A. Rahman, V. Srikumar, A.D. Smith, Predicting electricity consumption for
commercial and residential buildings using deep recurrent neural networks,
time. Appl. Energy 212 (2018) 372–385.
[17] K. Li, H. Su, J. Chu, Forecasting building energy consumption using neural
CRediT authorship contribution statement networks and hybrid neuro-fuzzy system: a comparative study, Energy Build.
43 (10) (2011) 2893–2899.
[18] P. Domingos, A few useful things to know about machine learning, Commun.
Lei Lei: Funding acquisition, Project administration, Supervi- ACM 55 (10) (2012) 78–87.
sion, Conceptualization, Writing - review & editing, Resources, [19] A. Liaw, M. Wiener, Classification and regression by randomForest, R News. 2
(3) (2002) 18–22.
Methodology. Wei Chen: Writing - original draft, Formal analysis, [20] S. Naji, A. Keivani, S. Shamshirband, U.J. Alengaram, M.Z. Jumaat, Z. Mansor, M.
Investigation, Data curation, Software, Validation. Bing Wu: Writ- Lee, Estimating building energy consumption using extreme learning machine
ing - review & editing. Chao Chen: Data curation. Wei Liu: Funding method, Energy. 97 (2016) 506–516.
[21] C.D. Li, Z.X. Ding, D.B. Zhao, J. Yi, G. Zhang, Building energy consumption
acquisition, Supervision, Conceptualization, Writing - review &
prediction: an extreme deep learning approach, Energies. 10 (10) (2017) 1525.
editing, Formal analysis. [22] K. Shankar, A.R.W. Sait, D. Gupta, S.K. Lakshmanaprabu, A. Khanna, H.M.
Pandey, Automated detection and classification of fundus diabetic retinopathy
images using synergic deep learning model, Pattern Recogn. Lett. 133 (2020)
Declaration of Competing Interest
210–216.
[23] I.R.I. Haque, J. Neubert, Deep learning approaches to biomedical image
The authors declare that they have no known competing finan- segmentation, Inf. Med. Unlocked 18 (2020) 100297.
[24] A. Kim, Y. Yang, S. Lessmann, T. Ma, M.-C. Sung, J.E.V. Johnson, Can deep
cial interests or personal relationships that could have appeared
learning predict risky retail investors? A case study in financial risk behavior
to influence the work reported in this paper. forecasting, Eur. J. Operation. Res. 283 (1) (2020) 217–234.
[25] M.A. Amanullah, R.A.A. Habeeb, F.H. Nasaruddin, A. Gani, E. Ahmed, A.S.M.
Acknowledgments Nainar, N.M. Akim, M. Imran, Deep learning and big data technologies for IoT
security, Comput. Commun. 151 (2020) 495–517.
[26] H. Fujiyoshi, T. Hirakawa, T. Yamashita, Deep learning-based image
This research was supported by the National Natural Science recognition for autonomous driving, IATSS Res. 43 (4) (2019) 244–252.
Foundation of China (Grant No.: 51708146), Guangxi Natural [27] P. Hähnel, J. Mareček, J. Monteil, F. O’Donncha, Using deep learning to extend
the range of air pollution monitoring and forecasting, J. Comput. Phys. 408
Science Foundation (Grant No.: 2018GXNSFAA281283), Guangxi (2020) 109278.
Science and Technology Project (Grant No.: Guike AD18281046). [28] F. Zhang, C. Deb, S.E. Lee, J. Yang, K.W. Shah, Time series forecasting for building
Wei Liu acknowledges the financial support from Energimyn- energy consumption using weighted Support Vector Regression with
differential evolution optimization technique, Energy Build. 126 (2016) 94–103.
digheten (Swedish Energy Agency, grant No. 50057-1). [29] F. Rodrigues, C. Cardeira, J.M.F. Calado, The daily and hourly energy
consumption and load forecasting using artificial neural network method: a
References case study using a set of 93 households in Portugal, Energy Proc. 62 (2014)
220–229.
[30] P. Lusis, K.R. Khalilpour, L. Andrew, et al., Short-term residential load
[1] H. Zhong, J. Wang, H. Jia, Y. Mu, S. Lv, Vector field-based support vector
forecasting: Impact of calendar effects and forecast granularity, Appl. Energy
regression for building energy consumption prediction, Appl. Energy. 242
205 (2017) 654–669.
(2019) 403–414.
[31] M. Ghiassi, D.K. Zimbra, H. Saidane, Medium term system load forecasting
[2] C.D. Li, Z.X. Ding, D.B. Zhao, et al., Building energy consumption prediction: an
with a dynamic artificial neural network model, Electr. Power Syst. Res. 76 (5)
extreme deep learning approach, Energies 10 (10) (2017) 1–20.
(2006) 302–316.
[3] K. Amasyali, N.M. El-Gohary, A review of data-driven building energy
[32] M. Cunkas, A.A. Altun, et al., Long term electricity demand forecasting in
consumption prediction studies, Renew. Sustain. Energy Rev. 81 (2018)
Turkey using artificial neural networks, Energy Source Part B: Econ. Plann.
1192–1205.
Policy. 5 (2010) 279–289.
[4] C. Fan, F.u. Xiao, Y. Zhao, A short-term building cooling load prediction method
[33] Z. Pawlak, Rough set, Int. J. Comput. Inform. Sci. 11 (2) (1982) 341–356.
using deep learning algorithms, Appl. Energy 195 (2017) 222–233.
[34] Q.L. Zhou, H.M. Mo, Y. Deng, A new divergence measure of pythagorean fuzzy
[5] K. Li, C. Hu, G. Liu, W. Xue, Building’s electricity consumption prediction using
sets based on belief function and its application in medical diagnosis,
optimized artificial neural networks and principal component analysis, Energy
Mathematics 8 (1) (2016) 389–411.
Build. 108 (2015) 106–113.

18
L. Lei, W. Chen, B. Wu et al. Energy & Buildings 240 (2021) 110886

[35] A. Skowron, H.S. Nguyen, Boolean reasoning scheme with some applications in [37] T. L£ken, Vibration Analysis using Rosetta-A Practical Application of Rough
data mining, in: European Conference on Principles of Data Mining & Sets, The Norwegian University of Science and Technology, 1998.
Knowledge Discovery, Springer-Verlag, 1999, pp. 107–115. [38] J. Sunarso, S.S. Hashim, J.Y.J. Yeo, J.J. Chew, MATLAB-based project assessment
[36] X. Zhu, Y. Qian, B. Sun, et al., Image Inpainting Forensics Algorithm in process modelling unit: A case study from Swinburne University of
Based on Deep Neural Network, Acta Optica Sinica. 38 (11) (2018) Technology Sarawak Campus, Educ. Chem. Eng. 33 (2020) 17–26.
105–113. [39] J.P. Liu, Building Physics, China Architecture & Building Press, Beijing, 2009.

19