energies

Article
Enabling Optimal Energy Management with Minimal IoT
Requirements: A Legacy A/C Case Study
Panagiotis Michailidis 1,2, *, Paschalis Pelitaris 1 , Christos Korkas 1,2 , Iakovos Michailidis 1,2 , Simone Baldi 3,4
and Elias Kosmatopoulos 1,2

1 Electrical and Computer Engineering Department, Polytechnic School of Xanthi,

Democritus University of Thrace, 67100 Xanthi, Greece; [email protected] (P.P.); [email protected] (C.K.);
[email protected] (I.M.); [email protected] (E.K.)
2 Centre for Research & Technology—Hellas (CE.R.T.H.), Information Technologies Institute (I.T.I.),
57001 Thessaloniki, Greece
3 School of Mathematics, Jiulonghu Campus, Southeast University, Nanjijng 211189, China; [email protected]
4 Center for Systems and Control, Delft University of Technology, 2628 CD Delft, The Netherlands
* Correspondence: [email protected]

Abstract: The existing literature on energy saving focuses on large-scale buildings, wherein the
energy-saving potential is substantially larger than smaller-scale buildings. However, the research
intensity is significantly less for small-scale deployments and their capacities to regulate energy use
individually, directly and without depreciating users’ comfort and needs. The current research effort
focused on energy saving and user satisfaction, concerning a low-cost—yet technically sophisticated—
methodology for controlling conventional residential HVAC units through cheap yet reliable ac-



tuation and sensing and auxiliary IoT equipment. The basic ingredients of the proposed experi-


mental methodology involve a conventional A/C unit, an Arduino microcontroller, typical wireless
Citation: Michailidis, P.; Pelitaris, P.;
IoT sensors and actuators, a configured graphical environment and a sophisticated, model-free,
Korkas, C.; Michailidis, I.; Baldi, S.;
optimization-and-control algorithm (PCAO) that portrays the ground basis for achieving improved
Kosmatopoulos, E. Enabling Optimal
performance results in comparison with conventional methods. The main goal of this study was
Energy Management with Minimal
IoT Requirements: A Legacy A/C
to produce a system that would adequately and expeditiously achieve energy savings by utilizing
Case Study. Energies 2021, 14, 7910. minimal hardware/equipment (affordability). The system was designed to be easily expandable in
https://doi.org/10.3390/en14237910 terms of new units or thermal equipment (expandability) and also to be autonomous, requiring zero
user interventions at the experimental site (automation). The real-life measurements were collected
Academic Editors: Daniele D. Giusto over two different seasonal periods of the year (winter, summer) and concerned a conventional
and Sandro Nizetic apartment in the city of Xanthi, Northern Greece, where summers and winters exhibit quite diverse
climate characteristics. The final results revealed the increased efficiency of PCAO’s optimization in
Received: 27 October 2021 comparison with a conventional rule-based control strategy (RBC), as concerns energy savings and
Accepted: 19 November 2021
user satisfaction.
Published: 25 November 2021

Keywords: building energy-management systems; domestic automation; centralized building

Publisher’s Note: MDPI stays neutral
optimization and control; energy-sustainable buildings; HVAC control; IoT
with regard to jurisdictional claims in
published maps and institutional affil-
iations.

1. Introduction
Energy saving in buildings represents a substantial scientific area for curbing energy
consumption and decreasing energy wastage. Recent scientific studies have exposed
Copyright: © 2021 by the authors.
numerous methodologies of and applications for mitigating the environmental impact of
Licensee MDPI, Basel, Switzerland.
This article is an open access article
buildings’ operation through intelligent energy management [1,2].
distributed under the terms and
During the last few decades, HVAC devices and equipment have been extensively
conditions of the Creative Commons utilized to improve indoor comfort in residential or commercial buildings [3,4]. In fact,
Attribution (CC BY) license (https:// this has undoubtedly improved the living quality of occupants; but, it has also dramatically
creativecommons.org/licenses/by/ increased the amount of energy consumed in such buildings. According to the literature,
4.0/). there is a large amount of energy consumed in buildings by the extensive use of HVAC

Energies 2021, 14, 7910. https://doi.org/10.3390/en14237910 https://www.mdpi.com/journal/energies

Energies 2021, 14, 7910 2 of 25

equipment, contributing significantly to the overall energy demand for climating purposes,
which accounts for 32% of overall energy demand [5]. To this end, there is an urgent need for
curbing energy wastage in buildings by utilizing different methods and practices in existing
legacy infrastructure. Recent IoT-based innovations, exploiting cheap microelectronic
devices, have been proven reliable in enabling smartness and optimal load management,
paving the way toward a gradual domestic digital transition that could significantly
contribute to a carbon-free economy [6].
Considering indoor climating (i.e., heating, cooling and ventilation) as the most
energy-intensive type of energy end-use in buildings, several methodologies and research
attempts have focused on decreasing the related energy account by revamping the residual
operational capacity of the appliance itself, increasing compression efficiency factors, higher
thermal capacity coolants, etc. However, recent advances in computer science and AI have
highlighted energy-saving potential through the efficient exploitation and control of HVAC
capabilities: out-balancing with any available promising grid elements (such as renewable
energy sources (RES)), batteries, ventilation and temperature control. Gathering indoor
and outdoor sensorial data, those control approaches may be elaborated even more by
utilizing efficient artificial intelligence approaches [7–9].
However, real-life applications are usually limited to industrial-scale climating devices
and large-scale deployments, where control policies, operational automation and digital
connectivity are implemented by manufacturers.
When supported by the appropriate sensors, actuators and other IoT instruments,
a building’s micro-grid forms a digital ecosystem at the network edge, enabling seam-
less integration with third-party smart services. The main goal of the current study is
to utilize cheap IoT infrastructure, combined with an intelligent, model-free, adaptive,
energy-management and -control tool, to achieve reduced energy use and sufficient thermal
comfort at the same time. The idea is to enable the implementation of a smart control-
optimization approach, integrating, as its main ingredient, a novel, self-adaptive, recon-
figurable, plug-n-play A.I. algorithm, denoted as PCAO [10], in conventional domestic
environments where the budget for networking, IoT-sensing and computing capacities are
usually quite limited [11].
State-of-the-art novel techniques that are widely accepted for delivering the intelligent
management and control of energy savings in HVAC and other building appliances include
fuzzy logic, adaptive fuzzy logic and rule-based approaches. Yet, while their adaptation
performances exhibit significant energy-saving results, their operation in real life still
exhibits numerous drawbacks, leaving a huge potential for curbing energy consumption
“out of the box”. Conventional state-of-the-art methodologies have proved incapable of
effectively coping with dynamic alterations in such non-linear, time-changing systems [8].
For instance, rule-based strategies (RBC) are dependent on simplified generic rules that
do not correspond efficiently to every (or even similar) type(s) of buildings. Their lack
of efficiency becomes more obvious when different occupancy behaviors occur, resulting,
sometimes, in the opposite result, where energy savings are concerned [9]. Additionally,
RBC approaches, in general, demand time-consuming and manual activities in order to be
established correctly, as well as adequate knowledge of their potential tasks [12].
Moreover, more elaborate methodologies, such as model-predictive control and
data-intensive neural networks, are considered as highly demanding of computational
and memory resources; a fact that renders their applicability in lightweight domestic
cases poor [3,8,9].
For example, model-predictive control (MPC) methodologies are usually based on
simplified white-box modelling schemes that are suitable mostly for non-stochastic, time-
invariant systems and, thus, their mathematical models can adequately describe the thermal
behavior of the building plant [12]. To this end, conventional domestic application cases
do not usually preserve the capacity for adopting elaborate methodologies, while lighter
adaptive control approaches are usually difficult to design, in terms of mitigating the
emerging thermal behaviors of the underlying building plant [9].
Energies 2021, 14, 7910 3 of 25

Contrarily, the examined PCAO methodology employs a built-in, lightweight per-

ception mechanism that is able to adjust its adaptive control policy on-the-fly and, thus,
effectively mitigate potential unnecessary excessive energy demands or thermal discom-
fort peaks. The PCAO methodology has already been evaluated several times in the
past, in large-scale tertiary building cases exhibiting significant energy-saving results [3].
In this scientific effort, a PCAO algorithm was evaluated for the first time in a mini-scale
apartment, adopting conventional cheap IoT equipment with low computational capac-
ity, thereby exhibiting its overall efficiency in real-life, everyday domestic applications.
Taking advantage of the utilized cost-efficient IoT sensors and interconnected actuation
equipment, the testbed implementation methodology concerned a single-room student
apartment, thermally managed by a single-inverter HVAC unit.
Our PCAO was eventually able to reconfigure itself and adapt to the new perceived
situation accordingly, offering optimized energy use by producing efficient control de-
cisions with respect to thermal conditions and forecasted data about weather and other
exogenous disturbances, such as occupancy schedules [3,8].
Moreover, it should be mentioned that the examined methodology exhibits seamless
scalability and implementation flexibility, enabling the integration of additional HVAC
devices—or other equipment—with just a few modifications and auxiliary equipment
additions. The current study is structured as follows:
• In Section 1, related work and its novelty are stated, while the overall test case’s
experimental approach is described with respect to the real-life testbed apartment
along with the system architecture and the HVAC features (A/C unit).
• In Section 2, the two control strategies—the benchmark rule-based control (RBC)
strategy and also the PCAO control strategy—are thoroughly described, along with
their attributes.
• In Section 3, the PCAO control application objectives are evaluated in detailed: the
system architecture, the PCAO control design and the PCAO algorithmic steps are
presented, along with the experimental control setup.
• In Section 4, the overall cost function is defined as a summary of energy-cost and
comfort-cost aspects. Energy cost- and comfort cost-evaluation approaches are being
thoroughly described.
• Section 5 includes the experimental PCAO and RBC results that are compared in detail
in their overall performance levels of energy saving and comfort. Outdoor conditions
that concern the summer and winter periods are illustrated. Next, experiments that
concern summer cooling and winter heating are presented and evaluated thoroughly
by extracting meaningful conclusions about the behavior of the PCAO control—in
comparison with the benchmark RBC approach.
• In Section 6, the final conclusions, as well as the future potential of the current research
effort, are thoroughly assessed.

1.1. Related Work and Novelty of the Paper

It should be noted that the conventional strategies for micro-grid load optimization
and control, at the building level, can be divided to three major categories: rule-based
control strategies (RBC), model-based predictive control strategies (MBPC), and model-
free adaptive control strategies (MFAC), each having their individual advantages and
disadvantages:
Rule-Based Control Strategies: RBC are practically driven by experience and observa-
tions, where a set of simplified logical commands (if/then) are defined in order to form a
multi-level policy that minimizes unnecessary excessive energy consumption. For instance,
such a rule-based conventional strategy may turn on/off the heater when the IoT sensorial
measurements (e.g., temperature) reach a critical point. Unfortunately, such approaches
tend to fail in many real-life scenarios where case-specific characteristics in different in-
dividual buildings impose highly diversified behavioral features: HVAC infrastructure
thermal capacity, building structure, dynamic occupancy schedules, building micro-climate
Energies 2021, 14, 7910 4 of 25

and location, user needs, infrastructural aging effects. Thus, rule-based approaches usually
present poor performance, due to the fact that their control policies are typically static and
inflexible to emergent dynamics [9].
It should be also noted that such systems usually demand periodic manual recalibra-
tions of their implemented control scheme in order to adequately guarantee some level of
energy efficiency. This strategy considers a manual and tedious procedure that requires
the involvement of a human expert with additional costs but cannot always guarantee
close-to-optimal behavior, since it’s impossible even for experienced individuals to de-
velop a highly elaborate set of rules able to mitigate the complex nonlinear dynamics of
the system [13].
Model-Based Predictive Control Strategies: MBPC model-driven (or model-dependent)
control strategies are mostly focusing on constructing an elaborate performance model
in order to efficiently assess and choose the appropriate decisions that are estimated to
present the greatest energy-efficiency potential [14–20].
Due to the unavailability of such models, the potential model-construction proce-
dure portrays a cumbersome activity that significantly limits the applicability of such
methodologies. Sophisticated white-box simulation approaches can provide templates
from pre-established use-cases, where manual analytic modeling or grey-box method-
ologies are considered. Complications and obscurity in such approaches are ordinary
and thus—in order to be suppressed—model oversimplifications are widely adopted [21].
To this end, using Model-driven approaches for determining a building’s micro-grid be-
havior is not a trivial task to be addressed, usually imposing a lot of abstractions and
generalizations that propagate major inefficiencies to the optimization and control schemes
that may even lead to even poorer performance than had previously existed.
Model-Free Adaptive Control Strategies: Model-free adaptive control (MFAC), is an
adaptive control methodology that does not demand process models. Model-free control is
mainly considered as alternative approach to adequately controlling complicated cyber–
physical systems—such as buildings—by utilizing a simplified representation of the system
(ultra-local model) and following, a mathematical estimation for constructing a simple,
but effective, controller [22]. Especially in cases that face real-life engineering problems in a
dynamic environment, without a robust and accurate model with which to begin, the most
appropriate and efficient approach is probably to utilize a model-free approach.
The PCAO Approach: The PCAO approach represents a next-generation model-free
adaptive control-optimization technique that is able to overcome multiple barriers and
obstacles facing state-of-the-art optimization approaches. Contrary to the aforementioned
widely utilized practices, the examined centralized approach—namely parameterized
cognitive adaptive optimization (PCAO) [3], which has been proved capable of efficiently
coping with the absence of a model, as well as with limited computational resources in a
domestic application case. PCAO was able to construct an internal lightweight performance
model to assess the effects of randomly perturbed control policies under uncertain dynam-
ics. PCAO is self-contained with respect to performance dynamics, requiring just a single
measurement point of the overall energy and comfort efficiency to readjust periodically,
while the overall objective function is analytically unknown.
It should be also underlined that the efforts of this study are aimed at extended
verification tests on a real-life, small-scale apartment—in contrast with previous studies,
which have extensively focused on large-scale/high-thermal-inertia structures (in real life
and simulations) [3,7–9].
The introduced real-life IoT measurements and conclusions of this paper highlight
the high efficiency of PCAO, where the same optimization problem is being determined
in a parallel manner by a conventional Rule-based Strategy for the same apartment room.
To this end, the PCAO approach was thoroughly contradicted by the utilized RBC strategy,
in diverse climate environments (summer, winter) over the year.
Energies 2021, 14, 7910 5 of 25

1.2. The Test-Case Building

The apartment in which the experiments took place represents a typical Greek stu-
dent chamber of multi-purpose use—as a bedroom and an office—with an area of 15 m2 .
The apartment leads to the rest of the residence via a typical wooden-dashboard door that
remained closed during the experiments in order to eliminate possible energy wastages,
to stabilize measurements and also to decrease the complexity level of the whole procedure.
It should be noted that at the east side there is a double aluminum reinforced door that
leads to the balcony of the apartment. The insulation of external walls that surrounding
the whole room apartment consists of dow insulation material 50-mm thick. The following
images illustrate the floor plan of the room as well as a 3D representation that illustrates
the interior furniture layout and also the location of the HVAC—an air-conditioning (A/C)
unit that was the only equipment responsible for regulating the indoor room climate.
Figures 1 and 2 illustrate the top-down perspective and the 3D perspective of the apart-
ment, respectively:

Figure 1. Room apartment floor plan.

Figure 2. 3D representation of the room apartment.

Energies 2021, 14, 7910 6 of 25

The room apartment that the experiments took place is positioned in northern Greece
in the city of Xanthi. Winters and summers are of diverse nature and thus our optimization
control tool proved its efficiency both in cold and hot climate conditions.

1.3. HVAC System Features

For manipulating the interior climate conditions of the apartment, an already-installed
HVAC wall unit was utilized. The air conditioner—A/C–HVAC (split-type) that integrated
a cooling power of 9000 Btu, a heating power of 12,000 Btu and maximum power consump-
tion of 853 kW. It should be underlined that the A/C unit integrates inverter technology,
which provides additional energy efficiency in energy consumption and also minimizes the
possible large fluctuations in internal temperature for maximizing user comfort. The eval-
uated approach delivered a next-generation energy-efficiency and user-comfort scheme,
introducing artificial intelligence and minimal low-cost IoT sensorial equipment to the
already-installed conventional A/C. The following Table 1 describes explicitly the technical
characteristics of the unit.

Table 1. Technical characteristics of the A/C unit.

Cooling Power 8531 BTU

Heating Power 11,943 BTU
Energy Efficiency Ratio (EER) 4.1
Coefficient of Performance (COP) 4
Cooling capacity 2.5 kW
Heating capacity 3.5 kW
Energy Start A
Airflow rate 8 m3 /min
Dehumidification performance 1 lit/h

Measurements: In this particular scientific experiment we have measured four (4)

main parameters using low-cost IoT sensors: (1) the interior temperature of the apartment;
(2) the interior humidity of the apartment; (3) external temperature; (4) external humidity.

2. Control-Optimization Methodologies
2.1. Control Strategies
The measurements that conducted in this particular experiment concerned two sepa-
rate and diverse seasonal intervals: the first concerns the summer season of 2017 and the
second concerns the winter of 2017. Those two intervals have been selected in order to
examine the performance of the PCAO and RBC in diverse climate conditions (temperature,
humidity). It should also be mentioned that those intervals were chosen in order to examine
a proper number of days wherein A/C was necessary to achieve adequate indoor comfort.
Last—but not least—the selected periods were chosen in order to generate effects that
exhibit the overall system’s potential in terms of energy savings and indoor comfort.
Most specifically the two intervals concern August (15–30 August 2017) and January
(25 December–10 February 2017). The final experimental selection was made, after many
attempts, so as to achieve the appropriate parameters, since we needed our experiments to
be as representative as possible. The most significant parameters of those metrics concern:
experimental duration (time intervals of the experiment and overall duration), the day
hours that concern the experiment, the RBC initial temperature setting point for the air
conditioner and Eigenvalues (e1 , e2 ).
Energies 2021, 14, 7910 7 of 25

2.1.1. Rule-Based Control Strategy

The selection of the appropriate experiment scenario was made after many attempts
to find the appropriate parameters. We wanted the experiments to be as representative as
possible. The most basic of these parameters are:
• Time intervals of the experiment and its duration: the experiments’ time intervals
were chosen to last 6 h per day. This choice was made both for practical reasons and
also after an empirical study of the hours that concerned the A/C unit under normal
conditions.
• Hours of the experiment: the experiments were carried out in the morning hours
of each day for the summer period and in the afternoon hours of the winter season,
because these hours required intensive utilization of A/C and users were indoors.
• Rule-based controller (RBC): this parameter concerns the initial temperature setting
for the air conditioner for each day. The RBC value was 26 ◦ C for the summer period
and 22 ◦ C for the winter period
• Eigenvalues (e1 , e2 ): experiments were performed with different eigenvalues for both
periods. The relationship between them was chosen as expressed in Equation (1),
as follows:

e2 = 4 × e1 (1)
Some values (e.g., e1 = 10−6 ) produced major alterations in the controller outputs,
while others (e.g., e1 = 10−4 ) changed them slightly. Thus, these parameters were de-
fined for generating the most viable best results. The results of the RBC algorithm mea-
surements with the eigenvalues that produced the optimal control were e1 = 5 × 10−5 ,
e2 = 4 * e1 = 2 × 10−4 , respectively.
In addition to the measurements with the algorithm in use, there were also mea-
surements conducted for which the A/C needed to operate at a constant temperature.
The purpose of these measurements was to obtain a reference base in order to compare
the outcomes of the proposed PCAO algorithm with the outcomes of a conventional RBC
operation. The temperatures at which the A/C operated during these measurements were
25 ◦ C and 22 ◦ C for the summer and winter months, respectively. These temperatures
were chosen since they produce the highest user satisfaction, as generated by the Fanger
equation; here, in the Fanger equation as it concerns clothing insulation (clo), clo was 0.4
and 0.75 for the two experimental periods, respectively.

2.1.2. PCAO Control Strategy

There are 11 variables that were imported to the controller. Those variables concern:
• The conditions table: four values that describe the conditions inside and outside
the room (indoor temperature, indoor humidity, outdoor temperature and outdoor
humidity)
• Predictions Table: contains the forecast of external conditions, such as temperature
or humidity for the next n hours (six values). The value of n was three throughout the
whole experimental procedure.
• The operating temperature of the AC unit
Thus, the controller-status vector results in Equation (2) as follows:

11 + 1
11 × × L = 66 × L (2)
2
where L represents the number of linear controllers that are being alternated/surrogated
between each other. For all the conducted experiments the number L was equal to 1.
The PCAO optimization control algorithm that has been utilized is responsible for moni-
toring the profits of each controller, which are used to determine the actions of each room
element at each step. The two aforementioned control strategies are thoroughly compared
Energies 2021, 14, 7910 8 of 25

in Section 5.3 PCAO and RBC Comparison, between days that preserve quite similar
external/internal conditions in order to fairly exhibit their performance differences

3. Control Application Objectives

3.1. System Architecture
The architecture of the system, able to deliver significant interior comfort by utilizing
the optimal amount of energy, is based on the communication between the interconnected
components of a conventional PC, an IoT micro-controller and the pre-installed A/C unit.
• The PC: the personal computer of the user acts as the heart of the system, where the
PCAO control-optimization algorithm is installed and run. Gathering all the available
IoT sensorial information, the PC processes them, producing the finalized control
commands through the PCAO tool. Next the control commands are transmitted
to the A/C unit through the microcontroller. It should also be noted that, the PC
acts as a communication channel between the user and the overall system, thereby
offering a user-interface environment that transforms the potential user into a high-
level operator able to manually intervene, adding their individual high-level demands
and parameters.
• The IoT Arduino microcontroller: The microcontroller performed as an intermedia-
tor between the IoT sensors, gathering information on the existing conditions of the
interior and the exterior, and also transmitting them to PC (PCAO algorithm). More-
over, the microcontroller also used its IoT actuators as an intermediator between the
PC (PCAO) and the A/C unit, transferring the PCAO-generated control commands in
order of their execution
• The A/C unit: The unit applied the control commands that had been received by
the microcontroller, modifying its operation and, thus, altering the interior climate
conditions accordingly.
In order to fully explain the experimental procedure a simple diagram of the whole
experimental architecture is presented in 1–5 steps.
The 1–5 steps that the aforementioned diagram (Figure 3) represents are described
as follows:
1. The IoT sensors measuring the conditions and transmit the values to the microcon-
troller
2. The microcontroller is transferring measurements to the PC where the necessary
control commands are being generated through the PCAO algorithmic calculations
3. The control commands are directed back to microcontroller
4. Microcontroller IoT actuators deliver them to the A/C unit
5. The A/C applies the control commands by modifying the interior climate through
its operation

Figure 3. System architecture implementation framework.

Energies 2021, 14, 7910 9 of 25

3.2. PCAO Control-System Design

The following Figure 4 represents the setup of the optimization technique as well as
the optimization algorithm.

Figure 4. Controller and algorithmic setup.

The system simulation model is connected to a parametrized controller and the

proposed optimization algorithm (PCAO). At each step, it evaluates these parameters
utilizing the condition results of the apartment. For each PCAO effect in the controller,
the presented closed-loop system is simulated and, thus, the software provides appropriate
results in selecting the next controller. More specifically, the process is depicted within the
steps of the following section.

3.3. PCAO Algorithmic Steps

Before analyzing the measurement equipment for energy cost as well as for comfort
cost, its useful to exhibit the basic information for the novel PCAO algorithm. It should
be noted that the methodology has already been extensively evaluated and tested in
diverse real-life implementations, as well as in simulated environments. The algorithm
has exhibited its advantages in numerous applications without the need for cumbersome
pre-application procedures and calibration [23].
The mathematical implementation of PCAO is briefly presented in Table 2 below:

Table 2. Implementation of PCAO algorithm.

Step PCAO Step Description
Initially, define the central control matrix P(0) in order to be semi-definitely positive about the e1 and
e2 eigenvalues (where e2 > e1 > 0), and initialize a scalar time-decaying continuous function a(t) that
1
will be utilized as the continually decaying exploration radius for the central candidates that would
be produced in Step 4.
Determine the refreshed Th time interval and deploy the ongoing control matrix for the entire update
time duration. At the end of the time duration, compute the overall performance index, Eg,PCAO ; this
2
factor represents the approximation error of the optimal cost-to-go function and derives from the
Jacobi–Bellman equation [24].
Next, a central, parametrically linear estimator needs to be trained by utilizing the Eg,PCAO values,
3
as previously computed, and the respective central control matrices as the central regressor vector.
Produce a randomly version of the deployed central control matrix that is perturbed by a(t), utilizing
4 it as the perturbation center and using the price of the time-decaying function, as determined in Step
1, as the exploration radius.
Evaluate the performance the produced central competitors accordingly with Step 4, utilizing the
5 respective central, parametrically linear estimator (Step 3) and nominate the one exhibiting the
greatest overall performance Pb (i).
Establish the elected matrix as the current one and return to Step 2, repeating until achieving the
6
desired performance.
Energies 2021, 14, 7910 10 of 25

For supplementary information concerning the mathematical procedures and formu-

lations, the interested reader may address [24].
Figure 5 portrays an algorithmic procedure that is conducted iteratively until the
system performance converges to an optimized point.

Figure 5. PCAO algorithmic diagram.

3.4. Experimental Control Setup

Having illustrated the system architecture, PCAO control design and PCAO steps,
we need to mention that the experimental setup considers the following metrics:
Metric Summer Period Winter Period
Experiments Interval 15–30 August 25 November–10 December
Experiment Time Zone morning afternoon
HVAC Control Window 6h 6h
HVAC Control Set-Point every 15 min every 15 min
Controller Update every 24 h every 24 h
A/C Operating Temperatures 25–22 ◦ C 25–22 ◦ C

4. Defining the Overall Cost Function

4.1. The Overall Cost-Function Methodology
As has already been mentioned, the utilized controller in this experimental effort
had, as primary targets, (a) the optimization of energy consumption, and also (b) of user
satisfaction, by minimizing the control function of the system. Also, in order to avoid
the complexities during evaluation, usually introduced when considering Fanger and
PPD (predicted percentage of dissatisfied) metrics, the optimization criterion considered
a simpler formulation that blends non-renewable energy and user discomfort as factors,
as the following Equation (3) describes [25]:

Total Cost = t ∗ Energy Cost + (1 − t) ∗ Com f ort Cost (3)

where Energy Cost is defined as the consumption of energy using the A/C unit, and Comfort
Cost as the satisfaction degree of the occupant for each interval of the experiment (the
method for calculating Comfort Cost is defined explicitly in the following Section). The factor
t is a positive decimal number (0 < t < 1) and it is used in order to potentially manipulate
the weight of each factor (Energy Cost and Comfort Cost) in the Total Score; t represents a
mathematical way to manipulate the importance of each factor relative to each other.
It should be underlined that the Energy Cost and Comfort Cost factors that comprise
the Total Cost are acting competitively with one and other. Achieving and maintaining an
adequate comfort level for occupants requires additional energy consumption, while a
strict adherence control strategy in the energy economy often fails to satisfy tenants. In this
case, energy use may be considered wastage, since it does not meet the satisfaction criteria
that concern the consumed energy. Considering an equitably dependent optimization
standard between Energy Cost and Comfort Cost, t was set to 0.5. It needs to be underlined
Energies 2021, 14, 7910 11 of 25

that the sophisticated research concerning the selection outcomes of this diverse factor, t,
have already been noted in the literature [26].

4.1.1. Measuring Energy Cost

In order to accurately measure the energy consumption (Energy Cost) the EGM-PWML
energy meter was utilized. Among its alternatives, EGM-PWML has the ability to measure
energy, voltage, intensity, resistance and power factor, but what really assisted the execution
of this particular experiment was its ability for directly connect to the PC via a USB port
and the fact that it also integrates 128 KB of built-in internal flash memory for storing
measurements when disconnected from the PC. Those capabilities limiting cumbersome
manual control procedures, without the demand for physical presence.
Additionally, EGM-PWML is accompanied by software that is able to display its
measurements on the computer’s screen and also provides the ability to create graphical
representations for a descriptive display of measurements and their variations for the
monitoring device (A/C). Finally, it has the ability to export measurements to a Microsoft
Excel file for storing and processing.
It should be underlined that, for the efficient operation of the whole experiment,
the demand for real-time control optimization required the processing of measurements
during its execution and not at the end of its time interval—as the software provided by
the manufacturer otherwise allowed. To this end, a suitable function was developed in
Simulink MATLAB environment for adequate real-time processing and the use of such
measurements by our PCAO optimization algorithm.

4.1.2. Measuring Comfort Cost

According to the American Scientific Society for Heating, Cooling and Air Condition-
ing (ASHRAE, American Society of Heating, Refrigerating and Air Conditioning), thermal
comfort is defined as the psychosomatic state in which a person desires no thermal change
in the indoor environment and is satisfied with the prevailing thermal conditions [27].
Thermal comfort relates human sensation and perception to a number of environmen-
tal and physical parameters. Extensive studies have led to a number of thermal comfort
equations. These equations are mainly based on statistical studies or based on lookup
tables created from statistical data.
The most widely used model is the one developed in 1970 by O.P. Fanger, which
integrates heat-balance equations and empirical studies of skin temperature to determine
comfort [28]. The three indicators derived from Fanger’s comfort equation concern: (1)
predicted mean vote (PMV); (2) predicted percentage dissatisfied (PPD); and (3) lowest
possible percentage dissatisfied (LPPD). The first parameter estimates the average thermal
comfort selection and ranges from cold (−3) to very hot (3), with desired limits in [−0.5 +
0.5] field. The second parameter predicts/estimates the number of dissatisfied users under
given conditions. It is expressed as a percentage (%) and cannot be reset to zero, as shown
in Figure 6.
The environmental parameters that concern the thermal environment and affect the
heat dissipation from the human body—and thus the thermal comfort concern: temperature
(air, radiant, surface), humidity, wind velocity and multiple parameters such as wearables,
clothes and human movement or activity, metabolic rate etc., as is shown in the figure
below. These parameters play a significant role in the thermal comfort level of a particular
space [25,28].
Energies 2021, 14, 7910 12 of 25

Figure 6. PPD, PMV diagram.

The thermal balance between the human body and the environment is calculated from

M − W = H + Ec + Cres + Eres (4)

According to Fanger, calculations of the predicted mean vote (PMV) are described by
the following Equation (5):
 
PMV = 0.303e−0.036M + 0.028 × [( M − W ) − H − Ec − Cres − Eres ] (5)

Since some of the above variables are difficult to measure, the previous equation can
be expressed as in Equation (6):

n
PMV = 0.303e−0.036M + 0.0028 × ( M − W ) − 3.96 × 10−8 f cl [(tcl + 273)4 − tr + 273)4 − f cl hc (tcl − t a )


−3.05[5.73 − 0.0007( M − W ) − p a ] − 0.42[( M − W ) − 58.15] − 0.0173M(5.87 − p a ) (6)

−0.0014M(34 − p a )}

With the clothing-area factor being described by Equation (7) or Equation (8) below:

Fcl = 1 + 2 × Icl (7)

Or:
Fcl = 1.05 + 2 × Icl (8)
The convective heat transfer coefficient is described by Equation (9):
1
hc = 12.1 × V 2 (9)

The thermal resistance of clothing is described by Equation (10):

Rcl = 0.155 × Icl (10)

The clothing surface temperature is described by Equation (11):

tcl = 35.7 − 0.0275 × ( M − W ) − Rcl × (( M − W ) − 3.05 × (5.73 − 0.007 × ( M − W ) − p a )−0.42 × (( M

(11)
−W ) − 58.15) − 0.0173 × M × (5.87 − p a ) − 0.0014 × M × (34 − t a ));
Energies 2021, 14, 7910 13 of 25

where M represents the rate of metabolic heat production in W/m2 , W represents rate
of mechanical work accomplished in W/m2 , fcl represents clothing area factor (dimen-
sionless), tcl represents the clothing surface temperature in ◦ C, tr represents the Mean
radiant temperature in K, hc represents the convective heat transfer coefficient in W/m2 K,
ta represent the ambient temperature in ◦ C, pa represents the water vapor pressure in
ambient air in kPa, Rcl represents the thermal resistance of clothing in m2 K/W, V represents
wind velocity in m/s, Icl represents thermal resistance, clothing (clo unit), L represents the
thermal load on the human body in W/m2 . H represents energy losses in the form of heat
from the surface of the body through treatment, convection and radiation in W/m2 , EC the
amount of heat dissipated due to evaporation in a state of thermal equilibrium in W/m2 ,
Cres represents the amount of heat that is transferred by treatment during respiration in
W/m2 , Eres represents the amount of heat dissipated due to evaporation during respiration
in W/m2 and e portrays Euler’s number (2.718).
Having calculated the above PPD is given by the following Equation (12):
4 +0.2197× PMV 2
PPD = 100 − 95 × e−0.03353× PMV (12)

For the needs of this study, the dissatisfaction index (PPD) was utilized, representing
a factor that needs to be reduced in order to meet adequate comfort criteria.

5. Results and Discussions

5.1. Outdoor Conditions
5.1.1. Summer Period Outdoor conditions
The two control strategies were evaluated in the same apartment and in the same
seasons (summer and winter) but on different days. Note that we needed first to have
evaluated RBC, and then, secondly, the PCAO approach. This needed to take place contin-
uously since our PCAO algorithm integrates self-learning capabilities and its performance,
therefore, was increasing from day to day. Thus, in order to compare them fairly, we chose
days that possessed similar climatic characteristics. During the summer season and espe-
cially during August 2017, the observed temperatures varied between 23–34 ◦ C, while the
relative humidity ranged between 31–41%. Figures 7 and 8 depict the overall evolution of
the external conditions, which concern outdoor temperature and humidity for the summer
days on which experimentatio took place.

Figure 7. Outdoor temperature (summer).

Energies 2021, 14, 7910 14 of 25

Figure 8. Outdoor humidity (summer).

5.1.2. Winter Period Outdoor Conditions

During the experimental winter period the observed temperatures varied between
−1 ◦ C–14 ◦ C and humidity measurements ranged between 39–50%. Figures 9 and 10 depict
the procession of the external conditions over the winter period, which concern the winter
outdoor temperature and humidity, respectively.

Figure 9. Outdoor temperature (winter).

Figure 10. Outdoor humidity (winter).

Energies 2021, 14, 7910 15 of 25

5.2. Experiments
5.2.1. Summer Cooling
During the summer experimental period, the conducted measurements were obtained
utilizing both the RBC approach and the PCAO algorithm for controlling the air conditioner.
The main goal during this batch of experiments was to define a temperature-setpoint policy
for energetically and thermally optimized cooling purposes. The time interval for control-
ling the A/C was 16 days, overall. The primary aim was to adapt the PCAO algorithm
towards optimal operation—its learning/adaptation process—and thus, to align the PCAO
with the apartment’s parameters and the overall real-life user activity. Our concern was to
support the PCAO’s learning process with an adequate amount of time i.e., 14 days out
of the 16 experimental days during summer, in order to efficiently deliver the expected
optimization results.
The experiments were performed with an initial RBC temperature cooling setpoint
point of 25 ◦ C. As already mentioned, the RBC measurements were conducted while the
A/C was operating at a constant temperature. The selected temperature was 25 ◦ C—a
temperature point that Fangers’ equation suggests as the ideal set point for achieving
maximum occupant satisfaction.
The summarized results of these experiments are shown in Table 3, illustrating the
Initial Indoor temperature of the apartment at the start of the experiment, the Mean
Outdoor temperature, expressed as the average of the temperatures observed during
the daily available six-hour window, the total energy consumption, in Wh, as Energy
Cost, the PPD index, defined as the costs concerning user satisfaction (Comfort Cost) and
the Total Cost, which concerns energy and comfort costs and is given by Equation (3),
using t = 0.5.

Table 3. Experiment results (Summer).

Mean Outdoor Initial Indoor Energy PPD/Comfort

Day Strategy Total Cost
Temp Temp Cost Cost
1 PCAO 26.7 27 1372.1 704.3 1083.203199
2 PCAO 29.3 28 2109.9 1128.6 1619.266995
3 PCAO 30.6 30 1655.7 362.2 1008.957756
4 PCAO 29.9 28 624.4 1027.4 825.9022707
5 RBC 31.5 29 2466.6 687.2 1576.89854
6 PCAO 28.1 28 975.8 684.0 829.8855836
7 PCAO 30.5 28 1425.2 1068.5 1246.844298
8 PCAO 27.3 28 274.8 1221.6 748.241688
9 PCAO 27.6 27 292.7 1180.9 736.8000257
10 PCAO 27.5 27 216.3 1253.0 734.6413202
11 RBC 27.5 27 504.2 991.5 747.8687881
12 PCAO 27.8 27 561.0 739.9 650.465108
13 PCAO 29.5 27 1293.8 817.5 1055.650061
14 PCAO 29.6 27 960.4 718.6 839.478698
15 PCAO 29.5 28 933.8 899.6 916.6987783
16 PCAO 29.2 27 965.6 863.6 914.6379437

A/C temperature control points (summer): The variations in A/C operating temper-
ature over the days when the PCAO algorithm was utilized are illustrated in Figure 11.
Figure 12 portrays the average outdoor temperature and the set points/costs (energy cost,
PPD/comfort cost and total cost) on the days of the experimental procedure.
It is noticeable that, on the second day (which was warmer than the first) the PCAO
approach lowered the room temperature to almost 20 ◦ C in an attempt to mitigate the
increased PPD index values. Although, the third day was even warmer than the previous
two; the control setpoints were lowered to almost 25 ◦ C setpoint from the previous internal
temperature, since the PCAO perception mechanism was able to balance out the previous
day’s energy peak while achieving low thermal dissatisfaction levels (PPD). It is also
obvious that a similar phenomenon took place between the sixth and seventh days.
Energies 2021, 14, 7910 16 of 25

Figure 11. Set points / outdoor temperature (summer).

Figure 12. Controlling points/costs (summer).

Between 8th and 12th days, when the external conditions were similar to each other,
small alterations emerged in the set points, achieving similar total performance. This im-
provement became more evident on the 12th day, when the daily mean ambient tem-
perature was slightly higher from the previous ones, the total cost index decreased by
85 units compared with the 10th day. Note that from Table 3, on day 11, where the RBC was
considered, the respective total cost from the 8th–12th PCAO experimental days is slightly
better, presenting balanced energy–comfort index values, a fact related to the parameter t
in the cost function being set to 0.5.
On the 13th and 14th days the outdoor temperatures and the initial indoor temper-
ature rose approximately 2 ◦ C. Finally, the last four consecutive days of the experiments
presented similar ambient temperature levels; the PCAO was able to present a more effi-
cient performance (reducing the total cost) in comparison to RBC, even though the built-in
perception mechanism had been trained with sparse similar daily data. This means that
the PCAO algorithm managed to balance energy consumption and user satisfaction, while
showing acceptable Total Cost values as compared with day 11, on which the RBC strategy
was considered for similar conditions’ instantiation (see Table 3).

5.2.2. Winter Heating

The same types of measurements were also conducted during the winter. The main
goal during this batch of experiments was to define a temperature-setpoint policy for
Energies 2021, 14, 7910 17 of 25

energetically and thermally optimized heating. The PCAO algorithm had defined the room
conditions in real time for more than 20 days, utilizing IoT sensorial data. During this
time, the system was able to estimate the parameters of the control problem and offer an
adequate optimization effect. The initial heating setting point was 22 ◦ C.
Similarly, to the summer experiments, the RBC strategy was considered for more
than eight daily experiments wherein the A/C unit was constantly operating at the set
point of 22 ◦ C, ensuring adequate comfort (optimal cost satisfaction, as denoted by the
Fanger equation).
The summarized results of the performed experiments during the winter period
are presented in Table 4, using the same layout as developed in the previous section
concerning summer.

Table 4. Experiment results (winter).

Mean Outdoor Initial Indoor Energy PPD/Comfort

Day Strategy Total Cost
Temp Temp Cost Cost
1 PCAO 12.6 19 1846.5 918.7 1382.6
2 PCAO 8.7 18 2415.6 547.7 1481.65
3 RBC 8.3 19 2382.33 455 1418.665
4 PCAO 6.76 20 2452.5 548.7 1500.6
5 PCAO 4.53 17 2604 400.7 1502.35
6 RBC 5.2 18 2225 395.97 1310.485
7 PCAO 4.8 17 2344 390.44 1367.22
8 PCAO 0.5 16 3400 848.65 2124.325
9 RBC 0.07 16 3450 1193.03 2321.515
10 PCAO −0.5 14 3910 1100 2505
11 PCAO 0 15 4025.86 1180.47 2603.165
12 PCAO 0 14 4012.69 914.2 2463.445
13 PCAO 3.3 18 2450.26 865.09 1658.08
14 PCAO 4.6 16 2418.49 819 1618.745
15 PCAO 4 17 2310 800.58 1555.29
16 RBC 5.7 16 2303 791.5 1547.25
17 PCAO 7 16 2062 758.65 1410.325
18 PCAO 6 16 2610.59 819.035 1714.8125
19 PCAO 5 16 2247.86 789.74 1518.8
20 PCAO 6.07 15 2610.74 885.45 1748.095
21 PCAO 3.38 15 2837.91 866.26 1852.085
22 RBC 3.3 16 2840.2 1289.5 2064.85
23 PCAO 5.69 16 2236 913.18 1574.59
24 PCAO 6.07 17 2482.05 698.17 1590.11
25 RBC 6.2 18 2957.04 646.12 1801.58
26 RBC 11.69 18 1649.45 443.93 1046.69
27 PCAO 9.69 19 1710.46 257.81 984.135
28 RBC 7.85 17 2003.47 919.81 1461.64

A/C temperature-control points (winter): The output’s evolution of the A/C oper-
ating temperatures for the days when the control algorithm was utilized are presented
in Figure 13, along with the average outside temperature. The costs for those days are
presented in Figure 14.
It is noticeable that although the experiments commenced on a relatively warm day
(with an average outdoor temperature close to 8.7 ◦ C), a gradual decrease of temperature
took place from days 8–12, during which temperatures were below 0 ◦ C. By observing
the costs in Figure 14 and then dividing the first eight days into two groups according to
the external conditions, we obtain: group A, consisting of days 1, 2, 4, 5, 7, and group B,
of days 8, 10, 11, 12. We indicate that, although the total cost in the first days of each group
increased following the decreasing temperature, in the last days of each group the total
cost shows a decrease, defying the low outdoor temperature.
During the second day it is also noticeable that the PCAO controller, knowing that
it would be colder, raised the indoor temperature during the last hours of the available
Energies 2021, 14, 7910 18 of 25

time-control windows, aiming at maintaining an acceptable PPD during these late-day

hours where the hourly ambient temperature dropped significantly.

Figure 13. Set points and outdoor temperatures (winter).

Figure 14. Controlling points and costs (winter).

This happened again on the fourth day. Although the Total Cost increased during
the second and third days, the satisfaction index was maintained at high levels. Thus,
the increase in Total Cost was caused mainly by the increase in energy consumption
(Energy Cost)—a fact that was anticipated due the dropping ambient temperature.
A similar behavior pattern was observed during days 8–11, on which the daily mean
ambient temperature dropped close to 0 ◦ C. The PCAO control scheme attempted to main-
tain indoor thermal conditions by increasing the setpoints to 24 ◦ C, during the late control
hours, to acceptable dissatisfaction levels at the expense of increased energy consumption.
For the next two days (12th and 13th), the PCAO controller made decisions and presented
performance profiles similar to days 6 and 7, on which similar outdoor conditions occurred,
while the initial indoor temperature was significantly reduced, as expected (much colder
overnight conditions).
Energies 2021, 14, 7910 19 of 25

After the 13th day, a gradual increase of the average outdoor temperature was ob-
served. For the next 10 days, the average outdoor temperature stabilized at about 6 ◦ C.
Especially on 26th and 27th days, an overall total cost improvement was observed,
since the ambient temperature followed an increasing trend, resulting in much warmer
days that allowed the PCAO controller to keep the A/C heating setpoints to low levels and
save energy without jeopardizing comfort.

5.3. PCAO and RBC Comparison

Despite the obvious improvement that seemed to have been achieved in the cost func-
tion, the achievements can be also denoted through comparison between the PCAO control
algorithm’s performance and the RBC conventional control-approach’s performance.
We compared the individual PCAO operating days’ performances with those of the
RBC, which served as a benchmark. In order to efficiently compare the two aforementioned
control approaches, these two strategies were thoroughly compared between days that
preserved quite similar external conditions in order to fairly exhibit their performance
differences in terms of initial outdoor temperature and average outdoor temperature.
Tables 5–14 show the aforementioned data and the comparisons thereof. It should
be noted that Total Cost is acquired as a result of the sum of Energy Cost and Comfort Cost
(satisfaction cost) considering a balanced t factor value of 0.5.

Table 5. Comparison between the PCAO (third) and RBC (fifth) summer days.

Day Strategy Initial Temp. Mean Temp. Out Power Cons PPD Cost
3 PCAO 30 30.6 1655.7 362.2 2017.9
5 RBC 29 31.5 2466.6 687.2 3153.8
−1 0.9 33% 47% 36%

Table 6. Comparison between the PCAO (9th) and the RBC (11th) summer days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
9 PCAO 27 27.6 292.7 1180.9 1473.6
11 RBC 27 27.5 504.2 991.5 1495.7
0 −0.1 42% −19% 1%

Table 7. Comparison between the PCAO (10th) and RBC (11th) summer days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
10 PCAO 27 27.5 216.3 1253.0 1469.3
11 RBC 27 27.5 504.2 991.5 1495.7
0 0 57% −26% 2%

Table 8. Comparison between the PCAO (12th) and RBC (11th) summer days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
12 PCAO 27 27.8 561.0 739.9 1300.9
11 RBC 27 27.5 504.2 991.5 1495.7
0 −0.3 −11% 25% 13%

Table 9. Comparison between the PCAO (15th) and RBC (5th) summer days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
15 PCAO 28 29.5 933.8 899.6 1833.4
5 RBC 29 31.5 2466.6 687.2 3153.8
1 2 62% −31% 42%
Energies 2021, 14, 7910 20 of 25

Table 10. Overall aggregated comparison results (summer).

Days
Initial Temp Mean Temp Out Power Cons PPD Cost
[PCAO vs. RBC]
3–5 −1 0.09 33% 47% 36%
9–11 0 −0.1 42% −19% 1%
10–11 0 0 57% −26% 2%
12–11 0 −0.3 −11% 25% 13%
15–5 1 2 62% −31% 42%
Average 0 0.338 37% −1% 19%

Table 11. Comparison between the PCAO (19th) and RBC (16th) winter days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
19 PCAO 16 5 2247.86 789.74 3037.6
16 RBC 16 5.7 2303 791.5 3094.5
0 0.7 2% 0% 2%

Table 12. Comparison between the PCAO (21st) and RBC (22nd) winter days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
21 PCAO 15 3.38 2837.91 866.26 3704.17
22 RBC 16 3.3 2840.2 1289.5 4129.7
1 −0.08 0% 33% 10%

Table 13. Comparison between the PCAO (24th) and RBC (25th) winter days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
24 PCAO 17 6.07 2482.05 698.17 3180.22
25 RBC 18 6.2 2957.04 646.12 3603.16
1 0.13 16% −8% 12%

Table 14. Comparison between the PCAO (27th) and RBC (3rd) winter days.

Day Strategy Initial Temp Mean Temp Out Power Cons PPD Cost
27 PCAO 19 9.69 1710.46 257.81 1968.27
3 RBC 19 8.3 2382.33 455 2837.33
0 −1.39 28.20% 43.34% 30.63%

5.3.1. Summer Cooling

Comparing the fifth (RBC) with the third day (PCAO)—two days that had quite
similar environmental conditions—we identified a significant improvement in Total Cost
(about 36%). It should be noted that this improvement occurred after only three days of
operation with the PCAO controller. Of course, despite the fact that the fifth day (RBC) was
cooler than the third (PCAO) and that higher energy consumption and higher PPD were
expected, the PCAO strategy outperformed the RBC strategy even during more difficult
instantiation conditions. In order to extract explicit comparative results, we would need to
investigate additional factors and building dynamics that were not relevant to the context
of this specific study. Table 5 illustrates the comparison between the PCAO’s on the third
and the RBC on the fifth summer days, as follows:
By comparing the 11th day, using the RBC, with the 9th and 10th, using the PCAO,
it seems that the potential improvement is much lower than previously—with power
consumption (Power Cons) being significantly in favor of the PCAO. Here, the PCAO
seemed to apply a strategy that targeted reducing energy consumption in a well-balanced
manner, as compared with PPD. The increase in Total Cost, here, clearly was a consequence
of the increased dissatisfaction index. Table 6 illustrates the comparison between the PCAO
on the 9th and the RBC on the 11th summer days.
This was also confirmed by comparing of the 10th day (PCAO) with the conditions of
the 11th (RBC), observing that they were similar. Energy consumption was further reduced,
Energies 2021, 14, 7910 21 of 25

but so was comfort/user satisfaction. Thus, the improvement in the PCAO’s Total Cost over
these days was not significantly better than from the conventional RBC. Table 7 illustrates
the comparison between the PCAO on the 10th and the RBC on the 11th summer says.
Examining the 12th day, we identify that the overall Total Cost improvement was
significant: at this point, comfort/user satisfaction exhibited a major improvement, while
energy consumption increased insignificantly. This combination produced an overall
improvement in Total Cost of 13%, which was the largest of these three days (9th, 10th and
12th). Table 8 illustrates the comparison between the PCAO on the 12th and RBC on the
11th summer days, as follows:
As concerns the last three experimental days, the PCAO’s performance still remained
significantly better when compared with the 5th day under RBC control. Thus, it seems
the PCAO algorithm produces progressively better results while trying to find the optimal
ratio between energy savings and comfort/user satisfaction.
Comparing the results from the 5th day and the 15th (i.e., a day towards the end of
the summer experiments), we see that the controller managed to reduce the total cost index
by almost 40%. Although the conditions were more favorable on the days of the PCAO’s
use, such a big improvement and difference in the energy and comfort levels cannot be
justified by this difference alone. Table 9 illustrates the comparison between the PCAO on
the 15th and the RBC on the 5th summer days.
The aggregated results of the above comparisons are illustrated in Table 10, where
their average values are extracted.
From Table 10 it can be concluded that, for the days that had similar mean outdoor
temperatures, an improvement of 37% in consumed energy and an insignificant drop
in satisfaction (−1%) were observed. By summarizing the results from the aforemen-
tioned comparative analysis for the summer period, the overall PCAO improvement in
comparison with the conventional RBC control methodology is determined at 19%.

5.3.2. Winter Season Heating

Similar to summer season, the comparison resulted to significant improvements as
well. More specifically, when comparing the 16th RBC Day with the 19th PCAO day, a very
small improvement in Total Cost measures of about 2% can be observed. Table 11 illustrates
the comparison between the PCAO 19th day and RBC 16th winter day as follows:
However, when comparing the 20th and 21st days (PCAO) versus the 25th and
22nd days (RBC), respectively, a Total Cost improvement of 3% and 10% were observed,
respectively. Noticeably, on the 20th (PCAO) winter day the initial room temperature
was 3 ◦ C lower in comparison with the 25th (RBC) day and the external conditions were
colder, justifying the decreased (negative) satisfaction index. Nevertheless, the PCAO
controller improved energy consumption by 12%, achieving an overall improvement of 3%
in Total Cost.
The same applies to the 21st day, on which, although the indoor temperature was
1 ◦ C, the two cost parameters were improved, resulting in an overall improvement of 10%.
Table 12 illustrates the comparison between the 21st (PCAO) and 22nd (RBC) winter days.
Accordingly, the comparison of 24th (PCAO) winter day and 25th (RBC) winter day
exposes that the PCAO controller behaved robustly, despite the unfavorable conditions
for its operation. Although the comparison table reveals a deterioration in satisfaction
measures, a close observation between the individual measurements of the two exposes
that the negative difference was generated during the first hours of operation, where the
difference of 1 ◦ C in the initial indoor temperature needed to be compensated for. Table 13
illustrates the comparison between the 24th (PCAO) and 25th (RBC) winter days.
Finally, comparing the results of 27th (PCAO) with those of the 3rd (RBC) winter days,
we observed a major improvement in Total Cost of about 30%, while both cost parameters
were improved significantly. Although there was a difference of about 1 ◦ C in the mean
external temperature, favoring the optimization algorithm, this difference alone hardly
Energies 2021, 14, 7910 22 of 25

explains the large improvement in overall performance. Table 14 illustrates the comparison
between the 27th (PCAO) and 3rd (RBC) winter days.
Table 15 show the overall aggregated comparison results in winter, during the days
with colder indoor temperatures (0.4 ◦ C, on average) and during warmer days, presenting
higher outdoor temperatures (0.126 ◦ C, on average). The PCAO controller generated a
9.8% improvement in terms of energy consumption, 10.6% in terms of satisfaction and an
overall improvement of 10.52%, in terms of Total Cost.

Table 15. Overall aggregated comparison results (winter).

Day
Mean Temp
[PCAO– Initial Temp Power Cons PPD Cost
Out
RBC]
19–16 0 0.7 2% 0% 2%
21–22 1 −0.08 0% 33% 10%
23–16 0 0.01 3% −15% −2%
24–25 1 0.13 16% −8% 12%
27–3 0 −1.39 28% 43% 30.6%
Average 0.4 −0.126 9.8% 10.6% 10.52%

Another important group of winter days, in terms of the PCAO controller performance,
were days 8, 10 and 12, on which the lowest temperatures were recorded. On eighth winter
day, the PCAO algorithm demonstrated a rapid improvement in Total Cost of 8%, with an
enormous increase in satisfaction, of about 30%. Table 16 illustrates the comparison between
the 8th (PCAO) and 9th (RBC) winter days.

Table 16. Comparison between the PCAO (8th) and RBC (9th) winter days.

Initial Mean Power

Day Strategy PPD Cost
Temp Temp Out Cons
8 PCAO 16 0.5 3400 848.65 4248.65
9 RBC 16 0.07 3450 1193.03 4643.03
0 −0.43 1% 29% 8%

Table 17 summarizes the results of the above three comparisons. It is observed that for
the days presenting 1 ◦ C average lower initial indoor temperature and a nearly consistent
average outdoor temperature, the PCAO algorithm generated an overall optimization
of 10.1% in terms of Total Cost, improving energy consumption by 8.33% and increasing
the satisfaction index by 13.66%. It is also evident that, gradually, the PCAO controller
improved its performance in every index (energy consumption, satisfaction index and
Total Cost). Moreover, in a larger-scale experiment, the PCAO approach would probably
generate even greater results in terms of energy-consumption cost, satisfaction index and,
consequently, Total Cost.

Table 17. Overall aggregated comparison of results (winter).

Day
Mean Temp
[PCAO vs. Initial Temp Power Cons PPD Cost
Out
RBC]
8–9 0 −0.43 1% 29% 8%
19–16 0 0.7 2% 0% 2%
21–22 1 −0.08 0% 33% 10%
23–16 0 0.01 3% −15% −2%
24–25 1 0.13 16% −8% 12%
27–3 0 −1.39 28% 43% 30.6%
Average 0.33 −0.176 8.33% 13.66% 10.1%
Energies 2021, 14, 7910 23 of 25

6. Conclusions
The current research has illustrated an easy and affordable way to implement a novel
PCAO control-optimization strategy in a small-scale apartment by fine-tuning the opera-
tion of a single A/C unit. As has been illustrated, the creation of such a system integrated
a simplified installation of IoT sensors and low-cost equipment using conventional com-
puting resources—an Arduino, a computer and inexpensive IoT sensors. It should be
also noted that these computing resources could be further simplified by utilizing a single
microprocessor board, equipment that is widely commercially.
Thus, this approach may be of concern a vast number of potential residents that
represent similar conditions (e.g., spaces of similar type and the same basic characteristics,
such as the presence of an A/C unit) and can be easily adopted in conventional residencies
for delivering improved energy efficiency, as well as improved indoor comfort. That,
in comparison with cumbersome conventional manual approaches or even the automated
common conventional approaches (e.g., the commonly used benchmark of rule-based
control approaches (RBC)). Considering the potential applicability of such a simplified
concept, a vast amount of energy may be optimally saved, resulting in major economic,
social and environmental benefits in residential buildings.
As the results have indicated, the examined simplistic IoT-based system proved quite
capable of meeting demanding criteria for significant energy savings without any kind
of subsequent degradation in user comfort or satisfaction, contrary to the majority of
state-of-the-art techniques. It seems apparent that the optimization effect of the PCAO was
sufficiently balanced between energy savings and indoor comfort, generating an improved
energy efficiency result that significantly surpassed the performance of state-of-the-art RBC
optimization. In other words, the system proved adequate to provide intelligent automated
control to at least one air-conditioning unit, resulting in a significantly optimized energy-
efficiency outcome, as well. It should be noted that the current research concerned an
independent small apartment, wherein the identification of major energy-saving results is
cumbersome in comparison with large-scale systems (e.g., large commercial buildings)—
where, of course, the energy-saving potential is larger.
Another important aspect arising from the experimental analysis was that the behavior
of PCAO algorithm in increasing efficiency over a relatively short period of time, while also
generating a significant optimization improvement over the long term when compared with
a conventional RBC approach. It was noticeable that, after some time period, the algorithm
had aligned with the various dynamics of the apartment (occupancy, weather behavior,
climate), producing significant improvements in the overall performance measured as
“Total Cost”, in comparison with the RBC methodology.
All the aforementioned evaluated observations revealed that the current PCAO
methodology, in small residential buildings, may represent an initial step for future research
pursuits, e.g., the introduction of other sophisticated parameters that affect optimization
challenges, such as the presence or absence of users in the space, decision making about
external planning (user activities, etc.), the addition of other apartments (or even in an
entire residential building) may offer fruitful research opportunities.
Moreover, another potential research pathway may arise from benchmarking other
control strategies (e.g., fuzzy logic) or even combing the methods for further optimization
and configuration of such systems. Additionally, another interesting research avenue may
concern the connection of the current system to a smart grid or its supply of renewable
energy sources, for utilizing electricity when it is most available.
We could also focus even more on models that utilize optimization over longer dura-
tions. The operation of the system on a 24-h basis would provide the opportunity to test
different strategies during the day and night. To verify and improve upon the robustness of
the system, we envisage performing future experiments in which the IoT sensors, actuators
or air-conditioning units may malfunction. The aim will be the rapid perception and
treatment of such problems by the automatic calibration of equipment and informative
messages sent to the user and to other sites.
Energies 2021, 14, 7910 24 of 25

Finally, research would be useful in developing a better method for comparing the
results between different experimental conditions. This method could be used, both, as a
better way of evaluating obtained improvements and also as a complementary way of
evaluating control.

Author Contributions: Conceptualization, E.K.; Data curation, P.M., I.M., C.K., S.B. and P.P.; Formal
analysis, P.M. and I.M.; Funding acquisition, E.K.; Investigation, P.M.; Project administration E.K.;
Software P.P.; Writing—original draft, P.M.; Writing—review and editing, P.M. and I.M. All authors
have read and agreed to the published version of the manuscript.
Funding: The research leading to these results was partially funded by the European Commission
H2020-EU.2.1.5.2.—Technologies enabling energy-efficient systems and energy-efficient buildings
with a low environmental impact (Grant agreement ID: 768735) PLUG-N-HARVEST—https://www.
plug-n-harvest.eu/, accessed on 22 November 2021.
Institutional Review Board Statement: Not applicable.
Conflicts of Interest: The authors declare no conflict of interest.

Nomenclature

A/C air-chiller/-conditioning unit

IoT internet of things
LPPD lowest possible percentage dissatisfied
HJB Hamilton–Jacobi–Bellman
HVAC heating ventilating and air conditioning
MPC model-predictive control
MBPC model-driven predictive control
PCAO parameterized cognitive adaptive optimization
PMV predicted mean vote
PPD predicted percentage dissatisfied
RES renewable energy sources
RBC rule-based controller

References
1. Aguilar, J.; Garces-Jimenez, A.; R-Moreno, M.D.; García, R. A systematic literature review on the use of artificial intelligence in
energy self-management in smart buildings. Renew. Sustain. Energy Rev. 2021, 151, 111530. [CrossRef]
2. Antonopoulos, I.; Robu, V.; Couraud, B.; Kirli, D.; Norbu, S.; Kiprakis, A.; Flynn, D.; Elizondo-Gonzalez, S.; Wattam, S. Artificial
intelligence and machine learning approaches to energy demand-side response: A systematic review. Renew. Sustain. Energy Rev.
2020, 130, 109899. [CrossRef]
3. Michailidis, I.; Sangi, R.; Michailidis, P.; Schil, T.; Fuetterer, J.; Mueller, D.; Kosmatopoulos, E. Balancing Energy Efficiency with
Indoor Comfort Using Smart Control Agents: A Simulative Case Study. Energies 2020, 13, 6228. [CrossRef]
4. Porta-Gándara, M.A.; Rubio, E.; Fernández, J.L.; Rubio-Royo, E. Economic feasibility of passive ambient comfort in Baja California
dwellings. Build. Environ. 2002, 37, 993–1001. [CrossRef]
5. Bhagwat, A.; Teli, N.; Gunaki, P.; Majali, V. Review Paper on Energy Efficiency Technologies for Heating, Ventilation and Air
Conditioning (HVAC). Int. J. Sci. Eng. Res. 2015, 6, 106–116.
6. Terzopoulos, M.; Korkas, C.; Michailidis, I.T.; Kosmatopoulos, E. Overview of Legacy AC Automation for Energy-Efficient
Thermal Comfort Preservation. In International Conference on Computer Vision Systems, Proceedings of the 12th International Conference
ICVS 2019, Thessaloniki, Greece, 23–25 September 2019; Tzovaras, D., Giakoumis, D., Vincze, M., Argyros, A., Eds.; Lecture Notes in
Computer Science; Springer: Cham, Switzerland, 2019; Volume 11754. [CrossRef]
7. Michailidis, I.; Baldi, S.; Kosmatopoulos, E.; Boutalis, Y. Optimization-based Active Techniques for Energy Efficient Building
Control Part I: Optimization Algorithms. In Proceedings of the International Conference on Buildings Energy Efficiency and
Renewable Energy Sources, Kozani, Greece, 1–3 June 2014; pp. 35–38.
8. Michailidis, I.; Baldi, S.; Kosmatopoulos, E.; Boutalis, Y. Optimization-based Active Techniques for Energy Efficient Building
Control Part II: Real-life Experimental Results. In Proceedings of the International Conference on Buildings Energy Efficiency and
Renewable Energy Sources, Kozani, Greece, 1–3 June 2014; pp. 39–42.
9. Michailidis, I.; Schild, T.; Sangi, R.; Michailidis, P.; Korkas, C.; Fütterer, J.; Müller, D.; Kosmatopoulos, E. Energy-efficient HVAC
management using cooperative, self-trained, control agents: A real-life German building case study. Appl. Energy 2018, 211,
113–125. [CrossRef]
Energies 2021, 14, 7910 25 of 25

10. Michailidis, I.; Kapoutsis, A.; Korkas, C.; Michailidis, P.; Alexandridou, K.; Ravanis, C.; Kosmatopoulos, E. Embedding autonomy
in large-scale IoT ecosystems using CAO and L4G-CAO. Discov. Internet Things 2021, 1, 8. [CrossRef]
11. Lawal, K.; Rafsanjani, H.N. Trends, benefits, risks, and challenges of IoT implementation in residential and commercial buildings.
Energy Built Environ. 2021, in press. [CrossRef]
12. Mariano-Hernández, D.; Hernández-Callejo, L.; Zorita-Lamadrid, A.; Duque-Pérez, O.; Santos García, F. A review of strategies
for building energy management system: Model predictive control, demand side management, optimization, and fault detect &
diagnosis. J. Build. Eng. 2021, 33, 101692. [CrossRef]
13. Alcalá, R.; Casillas, J.; Cordón, O.; González, A.; Herrera, F. A genetic rule weighting and selection process for fuzzy control of
heating, ventilating and air conditioning systems. Eng. Appl. Artif. Intell. 2005, 18, 279–296. [CrossRef]
14. Wright, J.A.; Loosemore, H.A.; Farmani, R. Optimization of building thermal design and control by multi-criterion genetic
algorithm. Energy Build. 2002, 34, 959–972. [CrossRef]
15. Pourzeynali, S.; Lavasani, H.; Modarayi, A. Active control of high rise building structures using fuzzy logic and genetic algorithms.
Eng. Struct. 2007, 29, 346–357. [CrossRef]
16. Moon, J.; Jung, S.; Kim, Y.; Han, S. Comparative study of artificial intelligence-based building thermal control methods—
Application of fuzzy, adaptive neuro-fuzzy inference system, and artificial neural network. Appl. Therm. Eng. 2011, 31, 2422–2429.
[CrossRef]
17. Liu, G.; Liu, M. A rapid calibration procedure and case study for simplified simulation models of commonly used HVAC systems.
Build. Environ. 2011, 46, 409–420. [CrossRef]
18. Youssef, A.; Caballero, N.; Aerts, J.-M. Model-based monitoring of occupant’s thermal state for adaptive HVAC predictive
controlling. Processes 2019, 7, 720. [CrossRef]
19. Dai, C.; Zhang, H.; Arens, E.; Lian, Z. Machine learning approaches to predict thermal demands using skin temperatures:
Steady-state conditions. Build. Environ. 2017, 114, 1–10. [CrossRef]
20. Ma, Y.; Borrelli, F.; Hencey, B.; Coffey, B.; Bengea, S.; Haves, P. Model predictive control for the operation of building cooling
systems. IEEE Trans. Control Syst. Technol. 2012, 20, 796–803. [CrossRef]
21. Aftab, M.; Chen, C.; Chau, C.; Rahwan, T. Automatic HVAC control with real-time occupancy recognition and simulation-guided
model predictive control in low-cost embedded system. Energy Build. 2017, 154, 141–156. [CrossRef]
22. Sira-Ramírez, H.; Luviano-Juárez, A.; Ramírez-Neria, M.; Zurita-Bustamante, E.W. Active Disturbance Rejection Control of Dynamic
Systems: A Flatness Based Approach; Elsevier: Amsterdam, The Netherlands, 2017; ISBN 978-0-12-849868-2. [CrossRef]
23. Michailidis, I.; Korkas, C.D.; Kosmatopoulos, E.B.; Nassie, E. Automated control calibration exploiting exogenous environment
energy: An Israeli commercial building case study. Energy Build. 2016, 128, 473–483. [CrossRef]
24. Baldi, S.; Michailidis, I.; Ravanis, C.; Kosmatopoulos, E.B. Model-based and model-free “plug-and-play” building energy efficient
control. Appl. Energy 2015, 154, 829–841. [CrossRef]
25. Fanger, P.O. Calculation of Thermal Comfort: Introduction of a Basic Comfort Equation. ASHRAE Trans. 1967, 73, III4.1–III4.20.
26. Kosmatopoulos, E.B.; Michailidis, I.; Korkas, C.D.; Ravanis, C. Local4Global adaptive optimization and control for system-
of-systems. In Proceedings of the 2015 European Control Conference (ECC), Linz, Austria, 15–17 July 2015; pp. 3536–3541.
[CrossRef]
27. Kurazumi, Y.; Sakoi, T.; Nyilas, A. The Influence of the Solar Radiation Absorptivity up on the Outdoor Thermal Environment
Evaluation Index and the Thermal Sensory Perceptions. Am. J. Clim. Chang. 2018, 7, 204–217. [CrossRef]
28. Fanger, P.O. Thermal Comfort. Analysis and Applications in Environmental Engineering; Danish Technical Press: Copenhagen,
Denmark, 1970; p. 244.