TSINGHUA SCIENCE AND TECHNOLOGY

ISSNll1007-0214 12/22 pp143–157

DOI: 1 0 . 2 6 5 9 9 / T S T . 2 0 2 3 . 9 0 1 0 0 0 8
Volume 29, Number 1, February 2024

Heating-Cooling Monitoring and Power Consumption Forecasting

Using LSTM for Energy-Efficient Smart Management of Buildings:
A Computational Intelligence Solution for Smart Homes

Omid Akbarzadeh , Sahand Hamzehei, Hani Attar, Ayman Amer, Nazanin Fasihihour,
Mohammad R. Khosravi, and Ahmed A. Solyman

Abstract: Energy management in smart homes is one of the most critical problems for the Quality of Life (QoL)
and preserving energy resources. One of the relevant issues in this subject is environmental contamination, which
threatens the world’s future. Green computing-enabled Artificial Intelligence (AI) algorithms can provide impactful
solutions to this topic. This research proposes using one of the Recurrent Neural Network (RNN) algorithms known
as Long Short-Term Memory (LSTM) to comprehend how it is feasible to perform the cloud/fog/edge-enabled
prediction of the building’s energy. Four parameters of power electricity, power heating, power cooling, and total
power in an office/home in cold-climate cities are considered as our features in the study. Based on the collected
data, we evaluate the LSTM approach for forecasting parameters for the next year to predict energy consumption
and online monitoring of the model’s performance under various conditions. Towards implementing the AI predictive
algorithm, several existing tools are studied. The results have been generated through simulations, and we find them
promising for future applications.

Key words: design-builder; Besos; smart cities; smart building; neural network; Long Short-Term Memory (LSTM)

1 Introduction related problems. For example, Ref. [2] stated that the
The increase in the number of industries due to the usage of remote-controlled robots would be expanded to
expansion of the industry has made environmental include building and system maintenance in hazardous
issues more pressing today. According to Ref. [1], if situations.
air pollution is not decreased by 2030, an oxygen kit Nowadays, researchers have considered different
will be necessary to breathe comfortably since the air techniques to increase the number of Electric Vehicles
would be toxic. Artificial Intelligence (AI) is predicted (EVs) in the city and make them optimize to reduce the
to be used more in the future to tackle environmental- negative impact of using vehicles in cities[3, 4] . In this

 Omid Akbarzadeh, Sahand Hamzehei, and Nazanin Fasihihour are with the Department of Electronics and Telecommunications,
Politecnico di Torino, Turin 10129, Italy. E-mail: [email protected]; [email protected];
S258228@studenti. polito.it.
 Hani Attar and Ayman Amer are with the Department of Energy Engineering, Zarqa University, Zarqa 13110, Jordan. E-mail:
[email protected]; [email protected].
 Mohammad R. Khosravi is with Shandong Provincial University Laboratory for Protected Horticulture, Weifang University of Science
and Technology, Weifang 262799, China. E-mail: [email protected].
 Ahmed A. Solyman is with the Department of Electrical and Electronics Engineering, Faculty of Engineering and Architecture,
Nişantaşı University, Istanbul 25370, The Republic of Türkiye. E-mail: [email protected].
 To whom correspondence should be addressed.
Manuscript received: 2022-11-27; revised: 2023-02-02; accepted: 2023-02-13
C The author(s) 2024. The articles published in this open access journal are distributed under the terms of the
Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/).
 144 Tsinghua Science and Technology, February 2024, 29(1): 143–157

case, one could theoretically achieve better results by control, improving energy efficiency, safety, usability,
optimizing the techniques considered by Akbarzadeh et and accessibility[12, 13] . Machine learning is another
al.[5] branch of research that may be used for smart building.
The different buildings are also considered during The objective is to forecast what will occur in the next
these years, such as making hospitals smarter by year. In this situation, it is feasible to determine the
considering the Internet of Things (IoT) and machine amount of energy required for each building’s cooling,
learning techniques, such as the system developed by heating, electricity, and total power. By making smarter
Refs. [6–8]. In addition, cloud computing can be used decisions regarding heating and lighting use based on
to manage the data when data collection saves a huge the various needs of a building’s occupants, machine
amount of data, like the research done by Khan et learning has the potential to improve energy management
al.[9] Moreover, to solve environmental problems, many and reduce a building’s carbon footprint[14–16] .
solutions can contribute to reducing air pollution in This study seeks to simulate an office in a city with
major cities, for example, using an EV, which provides a cold climate in order to comprehend four distinct
many advantages. As stated by Ref. [10], EV is a characteristics, namely power heating, power electricity,
technology that eliminates the local Nitrogen Oxide power cooling, and total power, and to assess a time
(NOx ) and delicate Particulate Matter (PM) emissions. series method in the case of the forecast for the following
However, expanding the public transport system and year. Figure 1 depicts an overview of the stages, from
increasing the number of EVs on the market play designing an office to collecting data and making
a crucial role in reducing air pollution. One of the predictions based on that information.
most important factors is the optimization of buildings,
which is crucial because Ref. [11] stated that smart 2 Data Collection
buildings can be a solution that not only reduces costs Understanding the various stages of the project
and carbon footprint, but also positively affects the is the first step in evaluating and predicting the
environment. To make a building more intelligent, IoT- under investigation parameters. This research includes
based technologies might be useful for online monitoring the following steps shown in Fig. 1. Utilizing the
and controlling building energy consumption. By adding DesignBuilder program, create an office. After that,
sensors and a gateway, it is possible to monitor the generate two office configurations with Controlled
cooling and heating of the interior temperature and adjust Natural Ventilation (CNV) enabled or disabled and their
the cooling and heating power in an energy-efficient corresponding *. idf files using DesignBuilder. Then,
manner. For instance, if a household forgets to turn run simulations using Besos for each configuration’s
off an additional light before leaving home, he/she energy requirements, altering three parameters: Window-
may either utilize the application, or the light can be to-Wall Ratio (WWR) from 10% to 90%; two shading
turned off automatically. IoT helps intelligent buildings modes (on/off); and three insulation thicknesses (0.001,
with security, automated administration, and online 0.05, 0.2). All cases will be input into Besos to determine

Fig. 1 Summary of three main steps of method.

 Omid Akbarzadeh et al.: Heating-Cooling Monitoring and Power Consumption Forecasting Using LSTM for : : : 145

the ideal values that minimize energy use. Moreover, window consideration is necessary. WWR quantifies the
utilizing EnergyPlus to conduct a dynamic energy proportion of window area to the external wall area of a
simulation to get hourly energy consumption figures, and structure[24] . It is calculated
P as follows: 2
in the last step, use Long Short–Term Memory (LSTM) Glazing area .m /
WWR.%/ D P (1)
to forecast the energy consumption estimates for the Gross exterior wall area .m2 /
following year. The total area of the walls separating the building’s
2.1 DesignBuilder interior and exterior is the building’s gross exterior wall
area. The glass portion of a building’s facade or internal
DesignBuilder is an EnergyPlus-based software surfaces is referred to as glazing[25] . By adding 10% in
application for measuring and controlling energy, each stage, this paper varies the WWR between 10%
carbon, lighting, and comfort[17–19] . The designed and 90%; one parameter helps run the Besos. Another
office under study is located in Winnipeg, Manitoba parameter is Shading, defined as blocking the sun’s rays
province, Canada, with a total net area of 4036.47 in the building. The exterior shading of the facade
square feet. This office is divided into two zones with is a crucial method to improve a building’s energy
twelve windows, two floors, two roofs, and seven walls. efficiency in a subtropical area[26] . Understanding that
The year-round cold is one factor that led to Winnipeg’s
the installation of shading can affect the amount of
selection. Winnipeg is one of the coldest cities on the
artificial lighting in the building is critical to realizing
globe, according to Ref. [20]. This is important since
why Shading plays a crucial role in smart buildings. The
people’s energy consumption habits vary greatly in
energy required for artificial lighting is shown in the
regions with unpredictable weather. In an environment
following:
with moderate temperatures, energy is predicted to be
WŒkW
h=year D WL C WP (2)
used for power cooling in the spring and summer and
heating in the autumn and winter. However, this is not where WP is the amount of energy required for
the case for places such as Winnipeg. According to Ref. emergency lighting and lighting control systems, and
[21], the maximum April temperature in Winnipeg is WL is the amount required to ensure that the intended
10 ıC, indicating that power heating is more likely than illumination requirements are met. WL and WP are
power cooling. This phase results in various *. idf files, calculated in the following:
essential for initiating the next stages, namely, Besos PN FC FO .tD FD C tN /
WL D (3)
and EnergyPlus. 1000
where,
2.2 Besos and EnergyPlus  PN is the total installed lighting power (W);
Besos is a Python library with the latest version  FC is the constant illuminates factor;
of 2.1.3. It is defined by different modules to  FO is the occupancy dependence factor;
model and optimize buildings to assist academics and  tD is the daylight time usage (h);
engineers in constructing more sustainable and district-  FD is the daylight dependence factor;
integrated structures[22] . Besos is used to conduct several  tN is the non-daylight time usage (h).
simulations with changing parameters, as stated in PP C .tY t0 / C Pe m  te m
WP D (4)
Table 1, to determine the building’s energy usage. 1000
where,
Describing the parameters is essential to  PP C is the power needed for all control systems,
understanding why they should be considered. such as emergency and standby (W);
Windows have a significant impact on the performance  tY is the standard year time (h);
of a structure[23] . It implies that changing the number or  t0 is the annual operating time (h);
size of windows may alter a home’s energy consumption.  Pe m represents the total charging power installed
In the same building, when the window width is N for the emergency lighting (W);
centimetres, the indoor temperature may be close  te m is the emergency lighting charge time (h).
to the outside temperature, and vice versa when the Controlling shading smartly may increase the amount
window width is N M centimetres. This is why of sunshine to maintain warmer internal temperatures
Table 1 Parameters configuration. and spread the light. The climate of the area where the
WWR Shading Insulation thickness (m) building is located is one factor that should be considered.
0%–90% On and off 0.0001–0.3500 Reference [27] claimed that shading control can be
 146 Tsinghua Science and Technology, February 2024, 29(1): 143–157

classified into three groups. Optimizing the shading with Besos. Of course, the energy needed also depends
is crucial for warmer climates. In a mixed environment, on the outside temperature, which is supplied to the
it is possible to shade the building in the summer and let simulation as input with an *. epw file containing various
the sun in during the winter. Ultimately, it is preferable information about the outside temperature for a chosen
to reduce shading in a cold area. Given that Winnipeg is city. Other factors include,
a cold climate and the building under study is an office, it  the schedule (here are working hours);
is assumed that by employing the shade off, the amount  the surfaces’ transmittance; and
of sunlight may rise and make the interior warmer than  the building’s occupancy (including the number of
the exterior. If people do not shade throughout the day, people and machines like computers that naturally emit
the temperature may drop from the absence of sunshine, heat).
and artificial lighting use may grow. Engineers, architects, and researchers can be
Insulation thickness is the last characteristic examined estimated parameters such as energy consumption for
in this study. Understanding the U -value, which heating, cooling, etc. by considering EnergyPlus[30] . One
evaluates how efficient a material is as an insulator[28] , of the project’s goals is to simulate a building’s complete
is preferable before deciding on the insulation thickness. data monitoring system, so the ideal total energy usage
The lower the U -value, the better for the best insulating simulation was run using the Besos settings. After the
materials close to zero[29] . The method for calculating simulation, it is feasible to anticipate how much energy
U -value is shown in the following: will be used in the upcoming year. Figures 2–5 represent
1 the data collected from Besos and EnergyPlus, with the
U D (5)
RSI C RS C C RA C Ri y-axis demonstrating energy usage in the kWh unit and
where, the x-axis representing the data based on the hourly data
 RSI is the thermal resistance of the internal surface for one year.
(m2
K/W);
 RSC is the thermal resistance of the outside surface 3 Forecast Algorithm
(m2
K/W);
 RA is the thermal resistance of unvented air The performance gap is defined as the discrepancy
cavities; between the current level of performance and the one
 Ri is the thermal resistance of building components i that is desired[31] . According to Ref. [32], the amount
(m2
K/W). of energy used in buildings accounts for forty percent
The energy consumption required to keep the internal of the total energy used in the United States. Modelling
temperature within comfortable ranges can be seen and projecting building energy consumption is vital for

Fig. 2 Collected data of electricity power.

 Omid Akbarzadeh et al.: Heating-Cooling Monitoring and Power Consumption Forecasting Using LSTM for : : : 147

Fig. 3 Collected data of heating power.

Fig. 4 Collected data of cooling power.

addressing building energy efficiency issues and meeting and performance.
the present difficulties of human comfort. Urbanization 3.1 Summary of forecasting
growth is one of the major reasons for the increase The first step is understanding the difference between
in energy consumption[33] . To better understand, it prediction and forecasting, and why forecasting is
is necessary to comprehend the methodologies used considered in this paper. The available dataset is time
to anticipate the building’s energy consumption; as series, defined as recording the data over time[35] . For
Ref. [34] stated, AI and machine learning have been example, the data are recorded hour by hour in a year. A
proposed for forecasting buildings’ energy consumption forecast is a computation or assessment that combines
 148 Tsinghua Science and Technology, February 2024, 29(1): 143–157

Fig. 5 Collected data of total power.

information from past occurrences and current trends generate high-dimensional features. Ideally, a decision
to predict the result of a future event. At the same time, a on a given test vector x should not depend on all vectors
prediction states that something will occur in the future in the training set, but only on a select few. This approach
regardless of the availability of past knowledge[36] . reduces the computational complexity and localizes
This study focuses on forecasting power consumption the decision-making, meaning that the classifier uses
based on hourly data gathered over the period of a year; training set vectors similar to x to determine x’s class.
hence, the time series issue is the emphasis of this Considering this, let’s introduce the function L.y; /, as
research. Before comparing alternative methodologies, shown in the following:
summarized in Table 2, it is essential to recall the L.y; / D max.0; 1 y/ (7)
purpose of the study. Each machine learning and neural
where L.y; / represents the loss function, which
network is tailored to a certain study area.
calculates the difference between the actual output y
3.2 SVM and the predicted output . The goal of the SVM is
The SVM method aims to discover a hyperplane in N - to minimize this loss function, thereby enhancing the
dimensional space, where N is the number of features to model’s prediction accuracy.
classify the data points separately[37] . It is, nevertheless, It turns out that minimizing Eq. (8) is proportioned to
mainly employed to solve classification problems[38] . minimizing this loss function which depends on the loss
SVM employs the kernel technique to accomplish the function plus a regularization term, essentially penalizes
goal of applying a linear decision boundary to high- weighted vectors with a very large norm,
N
dimensional data. SVM has based on the premise that 1 2
X
min jjwjj C c max.0; 1 yi i / (8)
the classifier is simple. The method for calculating the w;w0 2
i D1
estimated class as a function is denoted in the following: After that, end up with a solution with the form
yO D w T x C w0 (6) expressed in the following:
N
where yO is the predicted class label, w represents the X
weight vector and w0 is the bias term, and x is the input wO D ˛i xi ; with ˛ spare (9)
i D1
feature vector. Through the kernel trick it is possible to
where vector w, O which is needed to set up the classifier,
Table 2 Under study algorithms. is a linear combination of the training vectors xi with
Support Vector Machine (SVM)
coefficients ˛i (a sparse vector is a vector that has
Gaussian Process (GP)
most coefficients ˛i equal to zero) and this leads to a
Artificial Neural Network (ANN)
classification stated in the following:
 Omid Akbarzadeh et al.: Heating-Cooling Monitoring and Power Consumption Forecasting Using LSTM for : : : 149

N
X insulation. While accounting for uncertainty, Ref. [43]
O
y.x/ D sgn .wO 0 C wO T x/ D sgn .wO0 C ˛i xiT x/ proposed a scalable probabilistic framework enabling
i D1 large-scale expenditures in energy retrofitting buildings.
(10)
Additionally, Ref. [9] provided a framework for GP
Results in Ref. [39] demonstrate that the SVM
modelling to calculate energy savings and uncertainty
approach might achieve more accuracy and
rates in measuring methods, they claim that GP models
generalization than the Back-Propagation (BP) neural
networks model, which is excellent for constructing may capture complicated nonlinear and multi-variable
cooling load prediction models. Moreover, Ref. [40] interactions and multi-resolution energy behaviour
compared four algorithms for predicting cooling in the patterns in contrast to linear regression. Since GP
office building in Guangzhou, China, models are created within a Bayesian framework,
 Back Propagation Neural Network (BPNN); they may account for many sources of uncertainty
 Radial Basis Function Neural Network (RBFNN); more systematically. Under the specific condition, the
 General Regression Neural Network (GRNN); performance of the GP is acceptable, but here is another
 SVM. question: is the GP the best option if there is a limit of
However, all of them are appropriate for building uncertainty in the data set?
cooling load prediction; SVM and GRNN methods can 3.4 ANN
reach better accuracy. Reference [41] applied a parallel
Neural networks’ novel techniques can be seen in
implementation of the SVM approach to estimate energy
different fields of studies[44] . The ANN phrase is taken
consumption in several buildings utilizing massive
from biological neural networks, which are responsible
time-series information for the first time. Numerous
for the development of the structure of the human
research studies have forecasted the power consumption
brain[45] . The reason why a typically more sophisticated
in buildings, as observed in mentioned articles, but other
model is considered, and also a model whose parameters
approaches must be included to compare them ideally
are calculated automatically, will be defined in the
since the considered dataset consists of data for both
following. Suppose a simple problem in machine
power heating and cooling, and the data simulated just
learning, “digit recognition” (the situation that is given
for one year and for one building.
some handwritten digits and set up a classifier to classify
3.3 GP the ten possible classes); it is pretty hard to relate these
When the Gaussian probability distribution is small pictures to a statistical distribution. It is also hard
generalized, GPs are considered, which can be a to describe this to a computer[46] .
foundation for advanced non-parametric machine Figure 6 shows the neural network learns the model
learning classification and regression techniques[42] . If (not related to a statistical distribution, but it is a
the joint probability density function of the random numerical model that computes some function, which is
used as a classifier). The neural network automatically
variables X.t1 /; X.t2 /; : : : ; X.tN / is Gaussian for any
computes the f .x/ parameters. Consequently, a
value of N , a random process is Gaussian, as stated in
classifier is a sophisticated, which is a multidimensional
the following:
1 function with several inputs (vectorized data samples)
fX .x1 ; x2 ; : : : ; xN I t1 ; t2 ; : : : ; tN / D p ex and outputs (class probabilities). Reference [47]
N
.2 / det.R/
(11) considered neural network to anticipate annual
where, consumption, which exemplifies an ANN strategy based
1 on a supervised Multi-Layer Perceptron (MLP) network
 x is equal to Œx mT R 1 Œx mI for electrical energy consumption forecasting, and
2
 x D ŒX.t1 /; X.t2 /; : : : ; X.tN /T is the column
vector of random variables;
 x is the column vector Œx1 ; x2 ; : : : ; xN T ;
 m is the mean of the column vector;
 R is the covariance matrix.
In building energy modelling, ideal values for several
parameters are often uncertain, such as envelope
Fig. 6 Neural network example.
 150 Tsinghua Science and Technology, February 2024, 29(1): 143–157

demonstrates that it can estimate annual consumption different parameters. For example, it is necessary to
with reduced error. In addition, Ref. [48] evaluated understand the relationship between cooling power and
several building energy consumption forecast algorithms inside or outside temperature. Data correlation is defined
for residential structures. Their findings indicated that as finding the linear relation between two variables.
deep neural networks and ANN have the highest Assume the situation in which a random vector, in this
performance accuracy, with ANN requiring less training example, has only two entries that X D Œx1 ; x2 . It is
time than deep neural networks. easier to talk about the lack of correlation, meaning if x1
and x2 are statistically independent, it is clear that there
4 Data Processing
should be no relationship between those values; or in
Data was procured as described in Section 2 other terms, the value of random variable x1 as a piece
after processing through DesignBuilder, Besos, and of knowledge should not say anything about the value
EnergyPlus. Table 3 presents a segment of the dataset, that x2 takes. For example, looking at the pixels of an
featuring six entries from June 15, 2022 at different image, they are typically similar between adjacent pixels
times. “Time” relates to the hour-by-hour simulation and neighbouring pixels. So it is possible to say that
parameter. tin and tout , represent the inside and outside there is some statistical correlation between the values
temperatures, respectively. PH is the heating power of those two pixels modelled by two random variables x1
(in kW
h), PC is cooling power (in kW
h), PE is the
and x2 , and a tool should be considered. The correlation
electricity power (in kW
h), and PT is the total power
coefficient measures the amount of correlation between
(in kW
h). In the end, the dataset consists of 8 columns
two random variables. The correlation coefficient is in
and 8760 rows, representing a year’s worth of simulated
the interval Œ 1; C1. If the correlation coefficient equals
data.
zero, there is no correlation between random variables. If
4.1 Data correlation it is equivalent to C1, it means they have a considerable
One of the steps to understanding a dataset more correlation. If it equals 1, there is a correlation between
effectively is to consider the correlation between x1 and x2 . Figure 7 represents the correlation between
Table 3 Dataset example with six entries.
Time tin (ıC) tout (ıC) PH (kW
h) PC (kW
h) PE (kW
h) PT (kW
h)
10:00:00 22.8 14.19 0.51 0.0 10.73 0.51
11:00:00 22.8 14.55 0.51 0.0 10.71 0.51
12:00:00 23.01 14.89 0.38 0.0 10.7 0.38
13:00:00 23.05 14.81 0.38 0.0 10.7 0.38
14:00:00 22.97 14.57 0.51 0.0 10.7 0.51

Fig. 7 Correlation between data available in the dataset.

 Omid Akbarzadeh et al.: Heating-Cooling Monitoring and Power Consumption Forecasting Using LSTM for : : : 151

data available in the dataset. the hyperbolic tangent is an alternative activation

To forecast each parameter of the power cooling, function that has a link to the sigmoid function, as shown
power heating, total power, and power electricity, it in the following:
needs to understand the most relevant features based 1 C tanh. z2 /
D (13)
on Fig. 7. This process, named feature selection, can 2
be helpful in reduction of the computational time and The key difference lies in the output range; unlike
improving model performance by reducing the number the Sigmoid function that varies between 0 and 1, the
of input variables. hyperbolic tangent function’s output ranges from 1 to
4.2 LSTM C1. The Tanh function has a derivative of up to 1.0,
which makes parameter changes far more relevant; this
The technique used to predict the parameters is LSTM. makes the Tanh function virtually always superior to the
It is the RNN architecture used in deep learning. LSTM Sigmoid function as an activation function for hidden
is an RNN architecture that shows temporal sequences layers.
and long-range dependencies more precisely than
4.4 Number of epochs
conventional RNNs. After understanding the appropriate
algorithm to forecast the dataset and determine the Another parameter that needs setting is the number
correlation between data, it is time to understand the of epochs, which represents complete passes through
LSTM configuration. This method forecasts power the training dataset[35] . Figure 8 illustrates the
cooling, heating, electricity, and total power in an office required number of epochs for the total power. After
building in Winnipeg, Canada. approximately ten epochs, the behavior stabilizes, as
evidenced by the intersection between the training loss
4.3 Activation functions
and validation loss curves. However, we set the epoch
Two activation functions are considered in this paper, value to 50 for the total power as the decrease in loss is
Sigmoid and Tanh. Sigmoid formula is in the following: still ongoing. While increasing the number of epochs
1 can lead to longer computation time, the effect on the
S.x/ D (12)
1Ce x model’s performance may be minimal, rendering the
where S.x/ is a sigmoid function. trade-off acceptable. Figure 9 indicates the electricity
The input to the function is simply a linear power epochs behaviour, where the situation is more or
combination of the input plus the bias. When the input less similar to the total power. The difference is that
is largely negative or positive, the function goes to zero after ten epochs, there are no changes in the case of
or one; however, the transition between zero and one is training and validation loss. The number of epochs for
smooth. power electricity is equal to 25. As Fig. 10 represents,
The chosen activation function is both continuous and the number of epochs for the heating power has a
differentiable. This characteristic means that even slight different story, where after ten epochs, there is a slight
changes to the weights within the neural network can decrement in the performance. For heating power, 50 is
result in substantial variations in the output. In contrast, considered the number of epochs. Cooling power epochs

Fig. 8 Training loss vs. validation loss for total power.

 152 Tsinghua Science and Technology, February 2024, 29(1): 143–157

Fig. 9 Training loss vs. validation loss for electricity power.

Fig. 10 Training loss vs. validation loss for heating power.

are represented by Fig. 11, where the performance is model’s predictions. A smaller MSE value indicates a
unacceptable since the lack of available data resulted in better estimation. In this equation, n is the number of
that. As the data is related to the cold climate, most days, data points, AEi represents the actual or observed values,
there is no cooling power in the office. and PEi denotes the predicted or estimated values from
the model. The difference between these two values,
4.5 Model evaluation
AEi PEi , indicates the error of each prediction. The
This paper considered R2 and MSE as metrics to squaring operation penalizes larger discrepancies more
evaluate model performance. The MSE, presented in severely than smaller ones, highlighting the model’s
Eq. (14), is commonly used to assess the quality of the ability to avoid large errors.

Fig. 11 Training loss vs. validation loss for cooling power.

 Omid Akbarzadeh et al.: Heating-Cooling Monitoring and Power Consumption Forecasting Using LSTM for : : : 153
n
1 X and specific days of such months are unavailable. MSE
MSE D .AEi PEi /2 (14)
n
i D1
is estimated to be 3.76 using LSTM, whereas R2 is
where AEi is the actual value at the i -th instance and 0.72. As shown in Fig. 14, cooling power has the
PEi is the predicted value at the i -th instance by the lowest performance of all factors studied in this study,
model. On the other hand, R2 is also considered, which confirming once again that increasing the number of data
is defined in Eq. (15). also enhanced the model’s performance. The total power
On the other hand, R2 is also considered, which is result is illustrated in Fig. 15, in which the performance
defined in the following: is more similar to the heating power, with the MSE
N equals to 2.76 and R2 equals to 0.78.
y.n/2
P
O
Œy.n/
2 e2 nD1 6 Conclusion
R D1 D1 (15)
y2 N
Œy.n/ y2 Technology use might result in a mix of benefits and
P
nD1 drawbacks. Controlling the energy that people use daily
R2 should be as close as possible to 1. is one thing that must be done to stop the destruction
of the environment. In this particular scenario, some
5 Results and discussion
research areas, such as machine learning and IoT, play
This paper’s collected data are excessively significant an essential role. Smart building is one topic that
since they include all energy consumption-related contributes to solving the problems related to using
characteristics, including cooling power, heating power, energy in buildings. The model evaluation of the LSTM
electricity power, and total power. In addition, the represents that the RNN algorithm has an acceptable
city under study is one of the coldest places in the result since its accuracy is in a good range for different
world, which may affect people’s energy consumption cases. In this paper, RNN is also considered a method to
for factors such as cooling power. In this situation, forecast data related to energy consumption.
the expectation is whether the provided data and the The paper’s primary issues pertain to two distinct
performance of the LSTM are adequate. Comparing topics. Initially, we chose a city with mostly low
original and forecasted data demonstrates, as shown temperatures throughout the year, which may result
in Fig. 12, that the performance of electricity power in data unavailability for cooling energy consumption.
is acceptable. Table 4 presents the model assessment Therefore, we analyzed the LSTM algorithm when
based on the MSE and R2 , which are 4.46 and imperfect data was supplied. In addition, the workplace’s
0.82 for electricity power, respectively. The result can heating, cooling, and energy usage are evaluated so that
be attributed to the consistent availability of hourly the most accurate projections may be made for the next
electricity power data throughout the year. However, years. In this instance, the performance of the LSTM
the model’s performance may be influenced by the was thoroughly evaluated in various conditions. For
absence of heating power usage data during warmer future expansions, it is strongly suggested to test LSTM
months in the workplace. This lack of data during high- in a city with hot temperatures to establish if LSTM’s
temperature periods could potentially affect the model’s performance is acceptable.
predictive accuracy. However, heating power’s MSE and
References
R2 are 2.76 and 0.78, respectively. This demonstrates the
LSTM’s adequate performance. Figure 13 contrasts the [1] K. R. Systems, What will happen if the level of air pollution
original and prediction of the model for heating power. continues to increase? https://www.kent.co.in/blog/what-
Cooling power is the most difficult aspect of this paper will-happen-if-thelevel-of-air-pollution-continues-to-increase,
since, in a cold region, cooling power reduces daytime 2020.
[2] O. Akbarzadeh, Evaluating latency in a 5G infrastructure
consumption. Consequently, data for the fall and winter for ultralow latency applications, http://webthesis.biblio.
polito.it/id/eprint/22652, 2021.
Table 4 Model evaluation. [3] Enel X, Do electric cars cause pollution? How do they
Energy characteristic MSE R2 reduce it? — Enel X, https://corporate.enelx.com/
Electricity power 4.46 0.82 en/question-and-answers/how-much-does-an-electric-car-
Heating power 2.91 0.77 pollute, 2022.
Cooling power 3.76 0.72 [4] Business Watch, What are the benefits of smart buildings?
Total power 2.76 0.78 https://www.businesswatchgroup.co.uk/what-are-the-
 154 Tsinghua Science and Technology, February 2024, 29(1): 143–157

benefits-of-smart-buildings/, 2021. evaluating critical challenges and practices for software

[5] O. Akbarzadeh, M. Baradaran, and M. R. Khosravi, IoT- vendor organizations to secure big data on cloud computing:
based smart management of healthcare services in hospital An AHP-based systematic approach, IEEE Access, vol. 9,
buildings during COVID-19 and future pandemics, Wireless pp. 107309–107332, 2021.
Commun. Mobile Comput., vol. 2021, p. 5533161, 2021. [10] BESOS, BESOS: Building and energy systems optimization
[6] Is machine learning the next big thing in and surrogate-modelling, https://pypi.org/ project/besos/,
smart buildings?—Intelligent building Europe, 2022.
www.intelligentbuildingeurope.com, 2022. [11] L. Troup, R. Phillips, M. J. Eckelman, and D. Fannon,
[7] Altensis, DesignBuilder Software, https://www.altensis. Effect of window-to-wall ratio on measured energy
com/en/services/designbuildersoftware, 2015. consumption in US office buildings, Energy Build., vol.
[8] AccuWeather, 5 of the coldest cities in the world, (in 203, p. 109434, 2019.
Chinese), https://www.accuweather.com/en/weather-news/ [12] Calculator Academy, Window to wall ratio calculator-
5-of-thecoldest-cities-in-the-world-2/434260, 2022. calculator academy, calculator. academy, https://calculator.
[9] A. W. Khan, M. U. Khan, J. A. Khan, A. Ahmad, K. academy/window-to-wall-ratiocalculator/, 2022.
Khan, M. Zamir, W. Kim, and M. F. Ijaz, Analyzing and [13] CIOB, Glazing-designing buildings, https://www.

Fig. 12 Original vs. prediction for power electricity.

Fig. 13 Original vs. prediction for heating power.

 Omid Akbarzadeh et al.: Heating-Cooling Monitoring and Power Consumption Forecasting Using LSTM for : : : 155

Fig. 14 Original vs. prediction for cooling power.

Fig. 15 Original vs. prediction for total power.

designingbuildings.co.uk/wiki/Glazing, 2022. [19] C. Symonds, How to detect a performance gap (+5 of

[14] Y. Q. Xiao, B. H. Qi, and W. Dong, Research on the the most common causes)? https://factorialhr.com/blog/
layout of external shading components and the design performance-gap/, 2022.
of microstructure in combined-façade of subtropical [20] C. Robinson, B. Dilkina, J. Hubbs, W. W. Zhang, S.
architecture, in Proc. 2011 Int. Conf. on Remote Sensing, Guhathakurta, M. A. Brown, and R. M. Pendyala, Machine
Environment and Transportation Engineering, Nanjing, learning approaches for estimating commercial building
China, 2011, pp. 1519–1524. energy consumption, Appl. Energy, vol. 208, pp. 889–904,
[15] YourHome, Shading, https://www.yourhome.gov.au/ 2017.
passive-design/shading, 2022. [21] M. Bourdeau, X. Q. Zhai, E. Nefzaoui, X. F. Guo,
[16] Energlaze, U-values explained, https://www.energlaze.ie/u- and P. Chatellier, Modeling and forecasting building
values-explained/, 2022. energy consumption: A review of data-driven techniques,
[17] TheGreenAge, Getting to grips with U-values, https://www. Sustainable Cities Soc., vol. 48, p. 101533, 2019.
tegreenage.co.uk/getting-to-grips-with-u-values/, 2022. [22] S. Seyedzadeh, F. P. Rahimian, I. Glesk, and M. Roper,
[18] EnergyPlus, https://energyplus.net/, 2022. Machine learning for estimation of building energy
 156 Tsinghua Science and Technology, February 2024, 29(1): 143–157

consumption and performance: A review, Visualization [36] J. Brownlee, Difference between a batch and an
Eng., vol. 6, no. 1, p. 5, 2018. epoch in a neural network, https://machinelearning
[23] Investopedia, Understanding time series, https:// mastery.com/difference-between-a-batch-and-an-epoch/,
www.investopedia.com/terms/t/timeseries.asp, 2022. 2022.
[37] S. Hamzehei, Gateways and wearable tools for
[24] Key Differences, Difference between forecasting and
monitoring patient movements in a hospital environment,
prediction, https://keydifferences.com/difference-between-
https://webthesis.biblio.polito.it/22711/, 2022.
forecasting-and-prediction.html, 2022.
[38] K. Rezaee, G. Jeon, M. R. Khosravi, H. H. Attar, and
[25] R. Gandhi, Support vector machine—Introduction to A. Sabzevari, Deep learning-based microarray cancer
machine learning algorithms, https://towardsdatascience. classification and ensemble gene selection approach, IET
com/supportvector-machine-introduction-to-machine- Syst. Biol., vol. 16, nos. 3&4, pp. 120–131, 2022.
learning algorithms-934a444fca47, 2022. [39] L. Z. Kong, G. S. Li, W. Rafique, S. G. Shen, Q. He,
[26] Analytics Vidhya, SVM—Support vector machine M. R. Khosravi, R. L. Wang, and L. Y. Qi, Time-aware
algorithm in machine learning, https://www. missing healthcare data prediction based on ARIMA
analyticsvidhya.com/blog/2017/09/understaingsupport- model, IEEE/ACM Trans. Comput. Biol. Bioinf., doi:
vector-machine-example-code/, 2022. 10.1109/TCBB.2022.3205064.
[27] Q. Li, Q. L. Meng, J. J. Cai, H. Yoshino, and A. Mochida, [40] F. Wang, L. N. Wang, G. S. Li, Y. L. Wang, C. Lv, and L. Y.
Applying support vector machine to predict hourly cooling Qi, Edge-cloud-enabled matrix factorization for diversified
load in the building, Appl. Energy, vol. 86, no. 10, pp. 2249– APIs recommendation in mashup creation, World Wide Web,
2256, 2009. vol. 25, no. 5, pp. 1809–1829, 2022.
[41] L. Z. Kong, L. N. Wang, W. W. Gong, C. Yan, Y. C. Duan,
[28] Q. Li, Q. L. Meng, J. J. Cai, H. Yoshino, and A.
and L. Y. Qi, LSH-aware multitype health data prediction
Mochida, Predicting hourly cooling load in the building:
with privacy preservation in edge environment, World Wide
A comparison of support vector machine and different
Web, vol. 25, no. 5, pp. 1793–1808, 2022.
artificial neural networks, Energy Convers. Manage., vol.
[42] F. Wang, H. B. Zhu, G. Srivastava, S. C. Li, M. R. Khosravi,
50, no. 1, pp. 90–96, 2009.
and L. Y. Qi, Robust collaborative filtering recommendation
[29] H. X. Zhao and F. Magoulès, Parallel support vector with user-item-trust records, IEEE Trans. Comput. Soc.
machines applied to the prediction of multiple buildings Syst., vol. 9, no. 4, pp. 986–996, 2022.
energy consumption, J. Algorithms Comput. Technol., vol. [43] L. Y. Qi, W. M. Lin, X. Y. Zhang, W. C. Dou, X. L. Xu,
4, no. 2, pp. 231–249, 2010. and J. J. Chen, A correlation graph based approach for
[30] J. Brownlee, Gaussian processes for classification personalized and compatible web APIs recommendation in
with python, https://machinelearningmastery.com/gaussian- mobile APP development, IEEE Trans. Knowl. Data Eng.,
processes-for-classification-with-python/, 2020. vol. 35, no. 6, pp. 5444–5457, 2023.
[31] Y. Heo, R. Choudhary, and G. A. Augenbroe, Calibration [44] C. M. L. Etoundi, J. De Dieu Nkapkop, N. Tsafack, J. M.
of building energy models for retrofit analysis under Ngono, P. Ele, M. Wozniak, J. Shafi, and M. F. Ijaz, A novel
uncertainty, Energy Build., vol. 47, pp. 550–560, 2012. compound-coupled hyperchaotic map for image encryption,
[32] Y. Heo and V. M. Zavala, Gaussian process modeling for Symmetry, vol. 14, no. 3, p. 493, 2022.
[45] C. Chen, C. Y. Wang, B. Liu, C. He, L. Cong, and S. Wan,
measurement and verification of building energy savings,
Edge intelligence empowered vehicle detection and image
Energy and Build., vol. 53, pp. 7–18, 2012.
segmentation for autonomous vehicles, IEEE Trans. Intell.
[33] javaTpoint, Artificial neural network tutorial, https://
Transp. Syst., doi: 10.1109/TITS.2022.3232153.
www.javatpoint.com/artificial-neural-network, 2022. [46] Y. R. Wu, H. F. Guo, C. Chakraborty, M. Khosravi, S.
[34] M. A. Azadeh and S. Sohrabkhani, Annual electricity Berretti, and S. H. Wan, Edge computing driven low-light
consumption forecasting with neural network in high energy image dynamic enhancement for object detection, IEEE
consuming industrial sectors of Iran, in Proc. 2006 IEEE Trans. Netw. Sci. Eng., doi: 10.1109/TNSE.2022.3151502.
Int. Conf. on Industrial Technology, Mumbai, India, 2006, [47] C. Chen, H. F. Li, H. Li, R. F. Fu, Y. Y. Liu, and S. H. Wan,
pp. 2166–2171. Efficiency and fairness oriented dynamic task offloading in
[35] R. Olu-Ajayi, H. Alaka, I. Sulaimon, F. Sunmola, and internet of vehicles, IEEE Trans. Green Commun. Netw.,
S. Ajayi, Building energy consumption prediction for vol. 6, no. 3, pp. 1481–1493, 2022.
residential buildings using deep learning and other machine [48] Y. Xia, S. R. Qu, and S. H. Wan, Scene guided colorization
learning techniques, J. Build. Eng., vol. 45, p. 103406, using neural networks, Neural Comput. Appl., vol. 34, no.
2022. 13, pp. 11083–11096, 2022.
 Omid Akbarzadeh et al.: Heating-Cooling Monitoring and Power Consumption Forecasting Using LSTM for : : : 157

Omid Akbarzadeh received the BEng Ayman Amer received the PhD degree
degree in electrical and electronic from Tennessee University, USA in 1991.
engineering from Shiraz University, He is currently the dean of the Department
Iran. He is currently a master student in of Energy Engineering, Zarqa University,
Communications and Computer Networks Jordan. His research interest topics are
Engineering (CCNE) at Politecnico di renewable energy, energy control, and
Torino, Italy. His research interests include energy storage. Moreover, smart design
ultralow latency networks, machine for energy-efficient designs is an interesting
learning, and cyber-physical systems. topic for his research.

Sahand Hamzehei is a master student Ahmed A. Solyman received the BEng and
in Information and Communications MEng degrees in electrical and electronics
Technology (ICT) for smart societies engineering from Technical Conllege of
from Politecnico di Torino, Italy. His Egypt in 1999 and 2006, respectively,
research interests include machine learning, and the PhD degree in electrical and
neural networks and deep learning, digital electronics engineering from the University
image processing, Machine Learning for of Strathclyde, UK in 2013. He is now
Healthcare (MLHC), and smart cities. an assistant professor at the Department of
Electrical and Electronics Engineering, Faculty of Engineering
and Architecture, Nişantaşı University, The Republic of
Hani Attar received the PhD degree from
Türkiye. His research interests include wireless communication
University of Strathclyde, UK in 2011.
networks, IoT, smart grids, artificial intelligence, and renewable
Since 2011, he has worked as an electrical
energy systems.
engineering and energy systems researcher.
He is now a university lecturer at Zarqa
University, Jordan. His research interests Mohammad R. Khosravi is an
include renewable energy systems, efficient international research fellow at Shandong
computing, cyber-physical systems, and Provincial University Laboratory for
wireless communications. Protected Horticulture, Weifang University
of Science and Technology, China. He
received the BEng and MEng degrees
Nazanin Fasihihour is a master student in
in electrical engineering from Shiraz
ICT for smart societies from Politecnico di
University, Iran in 2013 and Persian Gulf
Torino, Italy. Her research interests include
University, Iran in 2015, respectively, and the PhD degree
machine learning, neural networks, and
in communications engineering from Shiraz University of
deep learning.
Technology, Iran in 2020. His main topics of interest are
multimedia systems, and high-performance computing and
communications.