Energy & Buildings 278 (2023) 112664

Contents lists available at ScienceDirect

Energy & Buildings

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

Optimizing space cooling of a nearly zero energy building via model

predictive control: Energy cost vs comfort
Fabrizio Ascione a, Rosa Francesca De Masi b, Valentino Festa b, Gerardo Maria Mauro b,⇑,
Giuseppe Peter Vanoli c
a
Università degli Studi di Napoli Federico II, DII – Department of Industrial Engineering, Piazzale Tecchio, 80 80125 Napoli, Italy
b
Università degli Studi del Sannio, DING – Department of Engineering, Piazza Roma 21, 82100 Benevento, Italy
c
Università degli Studi del Molise, Department of Medicine and Health Sciences-Vincenzo Tiberio, via G. Paolo II contrada ‘‘Tappino”, 86100 Campobasso, Italy

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

Article history: The space conditioning of buildings is liable for more than 10 % of World final energy uses and related
Received 19 September 2022 CO2-eq emissions. Such share must be drastically reduced to pursue sustainability by optimizing both
Revised 8 November 2022 energy design and devices control. In this frame, space cooling is assuming an increasing weight owing
Accepted 10 November 2022
to climate change. Accordingly, this study applies a simulation- and optimization-based framework for
Available online 12 November 2022
the model predictive control (MPC) of space cooling systems. The case study is a nearly zero energy build-
ing located in Benevento – Southern Italy, Mediterranean climate – featuring an efficient air-source
Keywords:
multi-split system for cooling. The framework is envisioned to provide optimal values of setpoint temper-
Zero energy buildings
Space cooling
atures on a day-ahead planning horizon to minimize energy cost and thermal discomfort, based on
HVAC systems weather forecasts. Accordingly, a Pareto multi-objective approach is applied considering different dis-
Model predictive control comfort indicators to compare the Fanger theory with the adaptive one of ASHRAE 55. The optimization
Thermal comfort problem is solved by running a genetic algorithm – variant of NSGA II – under MATLABÒ environment.
Multi-objective genetic algorithm The objective functions are assessed via the coupling between MATLABÒ and EnergyPlus, using a vali-
dated building energy model. The multi-criteria decision-making is performed by setting a limit to dis-
comfort to pick an optimal Pareto solution. The framework is tested addressing a typical day of the
cooling season and using monitored weather data to simulate weather forecasts. Different optimal solu-
tions are provided to fit different comfort categories. Compared to a reference control at fixed setpoint –
26 °C – the proposed solutions with similar comfort performance ensure cost savings around 28 %.
Besides the proposed hypothetical implementation, the framework can be integrated in automation sys-
tems for real-time MPC. The novel contributions of this study lie in the methodology to combine MPC
with different thermal comfort models as well as in the results, which provide deeps insights about
the application of MPC for the space cooling of nearly zero energy buildings in a balanced climate.
Ó 2022 Elsevier B.V. All rights reserved.

1. Introduction and literature review tioning [4]. In this frame, globally, HVAC systems are liable for over
10 % of final energy uses and related CO2-eq emissions. Thus,
All over the World, the energy consumption for heating, venti- enhancing the energy efficiency of such systems plays a pivotal
lation, and air conditioning (HVAC) of buildings is constantly rising role in pursuing the carbon-neutrality by 2050, target set out by
because of population growth, climate change and higher comfort the European Green Deal to fight climate change and to ensure sus-
needs due to new lifestyles [1]. In the European Union (EU), build- tainable development [3]. Through research projects and grants
ings account for around 40 % of energy consumption and 36 % of policies, the EU is investing in smart meters and digital tools to
greenhouse gas emissions [2], which are linked to construction, optimize the control of energy systems. About such a research
use/operation, retrofit and demolition [3]. Similarly, at the World field, nowadays, there are advanced management strategies for
level, in 2020, the building sector accounted for about 36 % of HVAC systems, e.g., model predictive control (MPC) ones.
energy consumption, with a share of around 35 % for space condi- MPC is becoming an established technology in control and
automation of HVAC plants [5]. It is based on a building model to
achieve energy load predictions as a function of weather forecasts
⇑ Corresponding author. and building use. Such predictions are handled to optimize the
E-mail address: [email protected] (G.M. Mauro).

https://doi.org/10.1016/j.enbuild.2022.112664
0378-7788/Ó 2022 Elsevier B.V. All rights reserved.
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

Nomenclature

Symbols Acronyms
CC cooling cost ASHRAE American society of heating, refrigerating and air-
COP coefficient of performance conditioning engineers
CV(RMSE) coefficient of variation of the root mean squared error BNZEB nearly zero energy building of Benevento, i.e., case study
DHashrae discomfort hours according to ASHRAE 55 GA genetic algorithm
EER energy efficiency ratio HVAC heating, ventilating and air conditioning
gen generations of the genetic algorithm MCDM multi-criteria decision-making
MBE mean bias error MPC model predictive control
PMV predicted mean vote (Fanger theory) NSGA-II non-dominated sorting genetic algorithm-II
PPD predicted percentage of dissatisfied (Fanger theory) nZEB nearly zero energy building
Tout outdoor temperature
Tsp setpoint temperature Subscripts
el electricity
max maximum daily value
mean average daily value
ref reference control strategy (fixed 26 °C)

control actions in order to minimize/maximize one or more objec- optimal use in buildings. A novel economic MPC setup was devel-
tive functions, while taking into account comfort, cost and techno- oped by Knudsen et al. [17] and applied to a highly insulated
logical constraints within a systematic and flexible framework single-family house in Norway. The setup was able to achieve sig-
[6,7]. A proper implementation of MPC requires consistent defini- nificant demand response potentials, related to the exploitation of
tions of control/planning horizon, prediction horizon and sampling building thermal mass. De Lorenzi et al. [18] proposed an innova-
time. These factors can highly affect computational efforts, rejec- tive smart controller prototype for small-scale district heating net-
tion capability of disturbance factors, stability and setting time of works, which was tested on real case studies. Compared to
controllers [8]. The most popular applications employ determinis- standard rule-based control approaches, energy savings up to
tic models – i.e., based on physical laws to simulate building energy 34 % were attained while keeping the same comfort level. Freud
performance – because this approach ensures good trade-off et al. [19] applied a MPC framework in a building section hosting
between computational complexity and reliability, even if it may the two local ministries for energy & environment and urban
imply cost degradation [9,10]. Model predictive controllers can development in Hamburg, Germany. During a three-month testing
be a pivotal solution to increase energy efficiency for both from February to April 2020, they found that heating energy con-
newly-built and existing buildings. Notably, thanks to continuous sumption in the model-predictive-controlled section was lower
technological development and digital transition, MPC is becoming by around 30 % compared to the traditionally-controlled sections,
a market-ready and technically mature solution to promote the while keeping the same level of thermal comfort. Salakij et al.
energy renovation process of existing HVAC systems, implying [20] applied a finite horizon linear quadratic tracking to optimize
huge energy savings – up to 50 % in particular cases – with low setpoint temperatures, achieving thermal energy savings up to
payback times [11]. 43 % compared with a fixed setpoint temperature control. Accord-
The recent scientific literature proposes sundry studies unveil- ingly, Ławryńczuk et al. [21] compared MPC with a fixed setpoint
ing the advantages of MPC in building applications, as shown by temperature and MPC with a setpoint optimizer, finding that the
the comprehensive review of Yao and Shekhar [8]. The available setpoint decrement from 21 °C to 18.5 °C leads to reducing the
studies usually indicate that the adoption of this technique can energy demand by around 1430 kWh during the heating season.
provide energy savings and improvement of thermal comfort. Yang Bianchini et al. [22] proposed a MPC system integrating a central-
et al. [12] experimentally tested a MPC system in the campus of ized HVAC plant, photovoltaics, thermal and electric storage,
Nanyang Technological University, Singapore. They achieved a achieving energy cost savings of around 25.5 % compared to a stan-
decrease in cooling energy demand of around 58.5 % for an office, dard rule-based control. Aftab et al. [23] developed an automatic
and a decrease in HVAC electricity consumption around 36.7 % HVAC control framework integrating real-time occupancy detec-
for a lecture theatre, compared to the original rule-based control tion, dynamic occupancy predictions, and simulation-based MPC,
strategies. Mork et al. [13] numerically found that distributed which yielded energy savings higher than 30 % while ensuring
MPC can reduce thermal discomfort by 65 % and energy consump- acceptable comfort levels. Kumar et al. [24] compared stochastic
tion by 7.6 % compared to a rule-based control. Lee et al. [14], and deterministic MPC for a university campus, showing that the
addressing a residential unit in the Republic of Korea via simula- former provides total cost savings up to 7.5 %, reducing electricity
tions, showed that a MPC framework, compared to a simple on/ (by 6.9 %) and natural gas consumption (by 8.6 %), as well as peak
off thermostat, can achieve heating energy savings up to 12 %, or loads (by 29.8 %). Avci et al. [25] proposed a cost-effective and
reduction of comfort violation less than 0.5 °C. Hou et al. [15] pro- energy-efficient MPC strategy under real-time electricity pricing,
posed an error model to improve MPC prediction performance unveiling substantial energy savings compared with a fixed set-
under uncertainty in weather forecasts. The model was tested by point strategy. Recent studies combined MPC with advanced
simulations on a university building located in Norway, yielding multi-objective optimization algorithms in order to maximize
energy cost savings around 3.4 % and a reduction of temperature energy efficiency and thermal comfort. For instance, Du et al. [26]
violation numbers around 73 %, compared to a standard rule- proposed a novel regulation and control method integrating MPC
based control. Li and Wang [16] provided a comparative assess- and a multi-objective genetic algorithm to attain reasonable
ment of alternative MPC strategies and interesting insights for dynamic operation strategies of HVAC systems for large multi-

2
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

regional public buildings in Hohhot, Inner Mongolia, achieving payback times because the application of such strategies could
energy savings higher than 10 % without penalizing thermal require not-expensive software and hardware equipment. The
comfort. major investment lies in the development of effective MPC frame-
Definitely, MPC potentials are huge for both newly-built and works, easy to be applied to real buildings. Several case studies
existing buildings. In the former cases, efficient and smart con- have been addressed for different building types, using different
trollers are fundamental to ‘‘not waste” the optimized energy numerical methods for making simulations/predictions and solv-
design of the building-plants system, and to really achieve low- ing the optimization problems. In this frame, the present work pro-
or zero-energy performance. In the latter case, potentials are even poses a hypothetical implementation of a simulation- and
larger, since often existing buildings operate with rude HVAC reg- optimization-based framework for the MPC of space cooling sys-
ulation strategies, such as at fixed setpoint, implying substantial tems. The case study is an existing nearly zero energy building
energy savings. Definitely, the potential energy savings depend (nZEB) located in Benevento – South Italy. The framework imple-
on the baseline, and can vary between 10 % and 50 % at equal com- ments a Pareto multi-objective approach to minimize energy cost
fort conditions. and thermal discomfort, considering different comfort indicators.
However, the prediction of energy loads is a critical issue when The optimization problem is solved by running a genetic algorithm
it comes to develop controllers for heating and cooling systems, – variant of NSGA II [35] – under MATLABÒ environment.
especially as regards MPC, which needs reliable predictions of As shown by the reviewed studies, simulation- and
building performance. A solution can be the implementation of optimization-based MPC to minimize both energy consumption
simplified building energy models. Accordingly, Cholewa et al. and discomfort is not a novelty. However, as per the authors’
[27] proposed an original simplified but accurate method for knowledge, there are research gaps concerning the application of
short-term forecasting of heat power demand as a function of local such MPC frameworks to space cooling systems, investigating dif-
outdoor weather conditions (i.e., temperature, wind speed and ferent thermal comfort models within a Pareto optimization
solar radiation). The method is based on multiple linear regression framework. This can be crucial to answer important research ques-
applied to monitored data, and provides coefficients of determina- tions when implementing multi-objective MPC of space cooling,
tion around 0.9 for three real buildings in Lublin, Poland, with dif- i.e.,:
ferent use destinations (i.e., a multifamily residential building, a
health clinic, and a supermarket) and heating systems. The method
Which are the potential energy and cost savings provided by
was updated to take into account also occupant behavior as input optimal trade-off (Pareto non-dominated) solutions for differ-
variable [28], and integrated in an easy and widely applicable MPC ent comfort categories?
system (forHEAT) for space heating operation of existing and new
Which are the differences in the results between the Fanger the-
buildings [29], based on the forecasts of weather parameters. This ory and the adaptive comfort one?
system was installed and tested in a multi-family building and an
Can a correspondence between such theories be found?
office building located in Lublin, Poland, reaching energy savings
Which is the comfort theory that enables to exploit MPC bene-
around 15.2 % and 24.1 %, respectively. fits the most?
Another cutting-edge and promising solution for load forecast-
ing within MPC frameworks can be machine learning, which This paper tries to answer such questions, thus the main inno-
implementation is rapidly escalating in the building filed thanks vation lies in the following points:
to the increasing opportunities in collecting and processing build-
ing data. Accordingly, Yang et al. [30] proposed an approximate
a novel regulation of controlled parameters is presented for the
MPC framework using a nonlinear autoregressive network trained cooling season. The optimized values of controlled parameters
with measured data, addressing the case studies of [12], i.e., an are obtained by considering and comparing two different
office and a lecture theater in Singapore. For the office, the MPC approaches to the thermal comfort: the well-known semi-
and approximate MPC cut 58.5 % and 51.6 % of cooling energy con- static theory proposed by Fanger and the adaptive one of ASH-
sumption, respectively, compared to the original control, while for RAE 55 [36];
the lecture theatre, the savings were about 36.7 % and 36.2 %,
the dynamic temperature setpoint regulation model on-
respectively. The adoption of an artificial neural network-based demand level is established based on a multi-objective opti-
model predictive control for a residential house located in Ontario mization method, which can reduce the cooling demand while
was investigated in [31] and compared to a fixed setpoint strategy, ensuring the thermal comfort and the minimum exercise cost.
showing that MPC was able to save operating costs between 6 % The multi-criteria decision-making is performed by setting a
and 73 % depending on the season. In the same vein, machine limit to discomfort to pick an optimal solution from the Pareto
learning was applied to occupancy-based HVAC control systems front, providing – as novel contribution – different trade-off
in residential buildings, showing the significant impact of occu- optimal solutions to fit different comfort categories related to
pancy forecast models on controller performance [32]. Very recent different comfort theories;
studies combined machine learning and multi-objective numerical
the model predictive control concept is applied to a nZEB in a
optimization to fully exploit MPC potentials, e.g., energy savings up Mediterranean climate for verifying the further possible effi-
to 58 % [33], and to extract interpretable building control rules ciency increase in a building designed to have very low energy
[34]. requirements in such climatic conditions. The building model is
calibrated and validated against monitored data and the appli-
1.1. Aim of this study cation of the MPC framework is simulated for a typical day of
the cooling season.
The literature review indicates that the development of MPC
strategies for reducing energy consumption and running cost of 2. Materials and methods
HVAC systems is a central topic of present and future research.
The potential energy savings can be higher than 50 % while main- This section elucidates the addressed case study, the building
taining or even improving thermal comfort. Thus, MPC can be a energy model, and the applied simulation- and optimization-
precious tool to optimize the energy performance of both newly- based framework for the model predictive control of the space
built and existing buildings, generally with low investments and cooling system.
3
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

2.1. Case study (PV) system with 16 monocrystalline silicon panels (peak power of
330 W for each one) is installed on the roof. The PV modules are
The BNZEB (Fig. 1) is a nearly zero energy building (nZEB) South-oriented with a tilt angle of 5° in order to achieve the best
designed and built in Benevento (BN) in the frame of the Italian trade-off of power/producibility, minimization of mutual shading,
project ‘‘SMARTCASE” (code PON03PE_00093_1) – European Regio- and visual impact. The system is equipped with a 6.5 kWh lithium
nal Development Found. Benevento is a middle-size city in South- battery.
ern Italy with a typical Mediterranean climate (Italian climatic The energy converted by renewables does not affect the results
zone C, 1316 heating degree days at baseline 20 °C). The last sum- of this study, reported in section 3, because the focus is on electric-
mer (2021) has been characterized by a maximum temperature ity demand for space cooling.
higher than 38 °C, with average daily values very often higher than
30 °C. The situation is not better during the ongoing summer 2022. 2.2. Building energy model: Development and validation
January and February are usually the coldest months with a mini-
mum temperature of around 0 °C. Ascione et al. [38] already presented the development and val-
A detailed description of the technical solutions concerning idation of the numerical model of the BNZEB considering the pack-
BNZEB design has been already proposed by Ascione et al. [37], aged air system for HVAC. Briefly, DesignBuilderÒ was used for the
however for improving the readability of the paper the main infor- graphical construction (Fig. 2) and the building model was simu-
mation is here summarized. lated via EnergyPlus [39]. Particularized data concerning thermal
The building net conditioned area is around 70 m2, the window/ envelope and windows, lighting systems, and air conditioning sys-
wall ratio is 22.5 % and the surface-to-volume ratio is 1.03 m1. The tems were included, by considering all features of fans and pumps.
structural frame is in cross laminated wood with two layers of Also, the renewable energy source systems were properly modeled.
fiber-wood insulation. The measured thermal transmittance (U) For the purpose of this paper, the model has been modified con-
values of thermal envelope components are 0.19 W/m2K for the sidering the real occupation, equipment pattern, weather data, and
external walls, 0.22 W/m2K for the roof, 0.79 W/m2K for the inter- the activation of the multi-split system for space conditioning,
nal walls. Windows are double-glazed with low-emissivity coat- without taking into account energy conversion from renewables.
ings and PVC frame, resulting in a U-value of 1.5 W/m2K. At this aim, a monitoring campaign was done during the period
The space conditioning system consists of two alternative July 21–28, 2021. During this week, using the BNZEB monitoring
options, installed for testing purposes: system detailed in [38], the following parameters have been
recorded with a continuous time-step of 15 min: electric consump-
i. an efficient air-source multi-split system, with heating and tion, disaggregated for type (lighting, air-conditioning, equipment),
cooling capacity of 3.50 kW, heating COP (coefficient of per- outdoor conditions, indoor air temperature and relative humidity.
formance) of 3.45, cooling EER (energy efficiency ratio) of The monitored weather data have been used to define a
4.78; weather file for EnergyPlus simulations in real conditions. About
ii. a packaged air system, with an air-to-air reversible heat the occupants’ behavior during the indicated week, from 9:00 to
pump, which can be used for heating, cooling, dehumidifica- 17:00, three students occupied the building. More in detail:
tion, and mechanical ventilation, with heating capacity and
COP equal to 3.18 kW and 3.83 respectively, cooling capacity
on 21st and 26th July, passive approach: the space cooling sys-
and EER equal to 2.14 kW and 2.95, respectively. 100 m long tem was always turned on with a fixed setpoint (26 °C); the
horizontal geothermal probes at a depth of 2.0 m are used to windows were closed without the adoption of indoor shading
pre-heat or pre-cool the ventilation air. devices, the artificial lighting system was turned off and the
internal loads consisted only of personal computers;
The first option is activated in this study because it represents
on 22nd and 27th July, semi-active approach: the space cooling
the most popular space cooling system for residential buildings, system was always turned on with a fixed setpoint (26 °C), free
without precise control of ventilation and latent heat load. opening of windows and adoption of shading devices, as well as
A 2.16 m2 solar thermal collector is used to produce domestic free schedule for artificial lighting system and electric equip-
hot water and to pre-heat the ventilation air. Finally, a photovoltaic ment (personal computers, television, kitchen devices);

Fig. 1. Case study: BNZEB.

4
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

Fig. 2. BnZEB model.

on 23rd and 28th July, active approach: similar to the previous Table 1
one but the setpoint temperature was dynamic and freely Passive approach (July 21): measured data vs simulated data.

selected by the occupants. July 21, 2021 passive space cooling consumption
approach
measured simulated
Two days for each approach have been considered for kWhel kWhel
redundancy. 07/21 09:00:00 0.213 0.364
The EnergyPlus models have been calibrated by acting on set- 07/21 10:00:00 0.311 0.342
point values and schedules of indoor temperature, infiltration 07/21 11:00:00 0.282 0.460
and ventilation (windows opening) rates, use of shading systems, 07/21 12:00:00 0.588 0.532
07/21 13:00:00 0.529 0.620
and internal loads due to electrical equipment. In order to perform
07/21 14:00:00 0.943 0.977
the calibration procedure, the outputs of EnergyPlus simulations 07/21 15:00:00 0.821 0.784
have been compared with monitored data (both consumption 07/21 16:00:00 0.647 0.669
and indoor parameters) following the approach proposed by ASH- 07/21 17:00:00 0.637 0.717
RAE 14 Guideline [40]. The considered indices have been: MBE sum 4.97 5.46 MBE = 9.94 %CV
mean 0.552 0.607 (RMSE) = 16.7 %
(mean bias error) and CV(RMSE) (coefficient of variation of the root
mean squared error); the admitted thresholds, when hourly data
are used, are ± 10 % and 30 %, respectively [40].
Table 2
The comparison between the measured hourly values of elec-
Semi-active approach (July 27): measured data vs simulated data.
tricity consumption for space cooling and the outputs of the cali-
brated EnergyPlus models is reported in Tab. I, Tab.2 and Tab. 3 July 27, 2021 semi-active space cooling consumption
approach
for the passive (21st July), semi-active (27th July), and active measured simulated
(28th July) approaches, respectively. The tables also report the val- kWhel kWhel
ues of MBE and CV(RMSE) which fall within the mentioned admit- 07/27 09:00:00 0.261 0.656
ted ASHRAE thresholds. Similar values of such indices – complying 07/27 10:00:00 0.633 0.576
with the thresholds – have been achieved also for the other days. 07/27 11:00:00 1.45 1.14
07/27 12:00:00 0.978 0.896
Therefore, the EnergyPlus models can be considered validated 07/27 13:00:00 1.15 1.17
and well-representing the real energy performance of the building. 07/27 14:00:00 1.03 1.02
Thus, they can be used in the MPC optimization framework as a 07/27 15:00:00 1.27 1.20
prediction tool. Notably, the passive approach is assumed in the 07/27 16:00:00 1.18 1.11
07/27 17:00:00 0.618 0.782
application of the framework (see section 3) for testing purposes.
sum 8.57 8.55 MBE = -0.23 %
In this regard, the results in Tables 1-3 show that – moving from mean 0.952 0.950 CV
the passive approach to the active one – electricity consumption (RMSE) = 19.2 %
for space cooling tends to increase, even if the weather conditions
in these days were very similar. This is mainly due to setpoint tem-
perature variation – based on occupants’ preferences –, free open- 2.3. Model predictive control optimization framework
ing of windows and free use of electric equipment (higher internal
loads). Therefore, MPC potentials can be even higher if compared to A simulation- and optimization-based framework for the model
a baseline characterized by an ‘‘active” occupants’ behavior. predictive control (MPC) of space conditioning systems is applied.
Accordingly, a passive approach is here assumed to model the The framework is an enhancement of the version proposed by
baseline for simplification reasons and for avoiding to overrate some of the authors in [41], and compared to this latter it offers
the benefits – in terms of energy savings – produced by the MPC. the following upgrades:

5
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

Table 3 automation system, the optimization procedure should be run

Active approach (July 28): measured data vs simulated data. every day (assessment day) to optimize the control strategy for
July 28, 2021 active Space cooling consumption the following day (application day). Thus, the size of the planning
approach
measured simulated horizon is fixed, i.e., 24 h, but it is receding since it progressively
kWhel kWhel moves from one day to the next. In this regard, the minimization
07/28 09:00:00 0.235 0.695
of the computational time required by the optimization process
07/28 10:00:00 0.166 0.739 is fundamental, and it will be addressed in future studies as dis-
07/28 11:00:00 1.53 1.12 cussed in the Conclusions. For instance, if such a time is lower than
07/28 12:00:00 2.17 1.6 1 h, the procedure can be implemented at 23:00 of the assessment
07/28 13:00:00 1.59 1.62
day, when the weather forecasts for the application day are more
07/28 14:00:00 1.40 1.38
07/28 15:00:00 1.52 1.50 accurate. Otherwise, if it is higher, the procedure must be imple-
07/28 16:00:00 1.58 1.57 mented earlier, and this can affect weather forecast reliability.
07/28 17:00:00 1.45 1.47 Notably, the use of a day-ahead planning horizon is suggested in
sum 11.6 11.7 MBE = 0.46 %CV
different studies, e.g., [42,43], because it provides a good trade-
mean 1.29 1.30 (RMSE) = 26.2 %
off between forecast reliability and computational burden [41]. It
is more effective than shorter ones as long as it enables to maxi-
mize the use of the building thermal mass [44,45], and to achieve

it is applied to space cooling systems addressing a real building; optimal solutions by constraining the maximum daily operation of

it investigates different thermal comfort indicators to compare HVAC systems [41], which is restricted in some countries (e.g., in
the semi-static Fanger theory with the adaptive one of ASHRAE Italy for space heating). On the other hand, it is better than longer
55 [36] in order to achieve different optimal solutions fitting horizons because it ensures more reliable weather forecasts. In
different comfort categories, providing deep insights about the addition, such a planning horizon is particularly effective when
correlation between thermal comfort and energy the thermal inertia of the building-HVAC system is not consider-
consumption/cost. able, e.g., in the addressed case study characterized by lightweight
construction and air-based cooling system, thereby impeding to
The framework – sketched in Fig. 3 – is envisioned to provide exploit thermal buffers for next days. Definitely, the methodology
optimal values of cooling setpoint temperatures (Tsp) on a day- needs to be enhanced to take into account unexpected changes in
ahead planning horizon to minimize energy cost and thermal dis- weather forecasts, but this point needs further investigation.
comfort, based on weather forecasts, using a Pareto multi- (i.e., The design variables are the 24-hourly values of Tsp in the
two-) objective approach. building during the application day. Each variable can vary in a dis-
As argued in [41], the adoption of a 24-hour day-ahead planning crete range, i.e., in this study the following one (8 values per vari-
horizon means that in real-time applications within a building able) is used, typical of cooling operation in residential buildings:

Fig. 3. Multi-criteria decision-making to pick a Pareto solution.

6
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

Tspð CÞ 2 ½24; 24:5; 25; 25:5; 26; 26:5; 27; 50ðsystem switched off Þ see how the Pareto front changes as a function of the number of
ð1Þ generations. Such an analysis can be useful to understand if fewer
generations can be sufficient to reach consistent suboptimal solu-
It is assumed that Tsp is the same in the whole building being a tions. In order to achieve a further decrease of the computational
small construction, but different Tsp values could be considered in times, the minimum running period (tmin) introduced in [41] is
different thermal zones when larger buildings are addressed. The used in EnergyPlus simulations. It represents the minimum num-
solution domain is therefore composed of 8^24 = 4.722e + 21 solu- ber of days that should be covered by dynamic energy simulations
tions. As clear, a brute-force (i.e., exhaustive) search is unfeasible for achieving reliable results. The tmin value achieved in [41] for a
from a computational viewpoint making unavoidable the imple- similar case study is equal to 10 days. Here the running period is
mentation of a proper optimization algorithm. The first objective set equal to 20 days to be on the safe side. Therefore, EnergyPlus
function is the daily energy cost for space cooling per unit of con- dynamic simulations do not start from 1st January but from
ditioned area, denoted with CC, while different discomfort indica- 20 days before the day of application of the MPC strategy. Adopt-
tors are used as second objective function: ing such an expedient, the simulation time for each solution, as
concerns the addressed case study, is around 18.4 s using a pro-

PPDmax: maximum daily value of the predicted percentage of cessor Intel(R) Core(TM) i7-9700KF. Therefore, the maximum
dissatisfied assessed according to the Fanger theory (PPD) – computational time required by the whole optimization process
optimization 1; is equal to:

PPDmean: mean daily value of PPD – optimization 2;

DHashrae: daily percentage of discomfort hours (over 24 h)
ðmaximumnumber of generationsÞxðpopulation sizeÞxðsingle simulation timeÞ
assessed according to the adaptive comfort theory of ASHRAE
55 [36] considering a 90 % acceptability status, as recommended ¼ 100x144x18:4s ¼ 73:6h
for high performance buildings; the reference is to ASHRAE ð2Þ
guidelines, being a well-accredited standard at international This value is higher compared to the planning horizon (24 h),
level – optimization 3. and therefore it needs to be further reduced for real applications
of the proposed framework in building automation systems, as dis-
For PPD assessment, the clothing thermal resistance is set equal cussed in section 4.
to 0.0775 m2K/W (=0.5 clo) as typical of the summer season [36], At the end of the optimization, the GA provides the Pareto front,
while the metabolic rate is set equal to 120 W/person (around which collects the non-dominated solutions. Hence, the multi-
1.15 met assuming a body mean surface area of 1.8 m2) that is typical criteria decision-making (MCDM) is performed by setting a limit
of sedentary activity and standing at rest. Notably, this latter value to the discomfort objective function and selecting the solution that
has been calibrated in order to achieve the best fitting between sim- complies with such a constant and minimizes CC (see Fig. 3).
ulated PMV values and occupants’ perceptions during the monitor- Different limit values are investigated in section 3 to provide
ing campaign cited in section 2.2. These two parameters are different optimal solutions, fitting the following comfort categories
considered constant throughout the day because the clothing resis- of SFS-EN 15251 [47]:
tance does not undergo significant daily changes in summertime
(while it can drastically increase during nights in wintertime
comfort category II (PPD  10 %): ‘‘Normal level of expectation
because of heavy blankets), and the metabolic rate does not undergo and should be used for new buildings and renovations”;
substantial daily changes in residential buildings, which generally
comfort category III (PPD  15 %): ‘‘An acceptable, moderate
host sedentary activity or resting/sleeping. Actually, during summer level of expectation and may be used for existing buildings”.
nights, the clothing resistance tends to slightly increase because of
light bed sheets, while the metabolic rate decreases, but it is The comfort category I of SFS-EN 15251 is not considered
assumed that such variations compensate each other. because it refers to particular cases, i.e., ‘‘High level of expectation
Different indicators are investigated to compare different com- and is recommended for spaces occupied by very sensitive and
fort theories and to achieve insights about thermal comfort. The fragile persons with special requirements like handicapped, sick,
optimization problem is solved for each couple of objective func- very young children and elderly persons”.
tions (all of them to be minimized) – i.e., CC vs PPDmax, CC vs Notably, the optimal scenarios reported in Table 4 are identified
PPDmean, CC vs DHashrae – by running a variant of the non- and investigated. Each scenario is associated with one of the men-
dominated sorting genetic algorithm-II (NSGA-II) [35], under tioned comfort categories:
MATLABÒ [46] environment. EnergyPlus [39] is used as building
simulation tool and it is automatically coupled with MATALABÒ
as detailed in [41]. The genetic algorithm (GA) is a numerical,
semi-heuristic, semi-stochastic, and population-based optimiza-
Table 4
tion method, which simulates the ‘‘Darwinian” evolution of a
Optimal scenarios identified and investigated.
population of solutions, called ‘‘individuals”, through a series of
iterations, called ‘‘generations”, via the processes of creation, two-objective scenario comfort category SFS-
optimization EN 15,251 [47]
selection, crossover and mutation. Details about the algorithm
are available in [41]. Each design variable is encoded by a string optimization 1: a) Pareto optimal solution with II
CC vs PPDmax PPDmax  10 %
of 3 bits, so that it can assume one of the 8 (=2^3) values reported
b) Pareto optimal solution with III
in (1). Therefore, each solution is encoded by 3*24 = 72 bits. The PPDmax  15 %
proper setting of GA parameters is fundamental to achieve a fast optimization 2: c) Pareto optimal solution with II
convergence, which is essential in MPC applications because the CC vs PPDmean  10 %
computational time required by the optimization process should PPDmean d) Pareto optimal solution with III
PPDmean  15 %
be lower than the planning horizon. In this regard, the population optimization 3: e) Pareto optimal solution with II
size is set equal to 144 (twice the number of bits) and the maxi- CC vs DHashrae = 0
mum number of generations equal to 100. Furthermore, a kind of DHashrae f) Pareto optimal solution with III
convergence analysis is implemented, as shown in section 3.1, to DHashrae  10 %

7
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

the two scenarios related to optimization 1 – i.e., a) and b) –
the use of a passive approach – instead of an active one with
strictly respect the associated comfort category because the variable setpoint – to model occupants’ behavior enables to
PPD is always lower than the limit (10 % and 15 % for categories not overrate the potential benefits of MPC, as discussed in sec-
II and III, respectively), given that the addressed discomfort tion 2.2.
objective function is the maximum daily value of PPD;

the two scenarios related to optimization 2 – i.e., c) and d) – In addition, it is noticed that using a fixed setpoint strategy does
refer to ‘‘relaxed” comfort categories, because the mean daily not mean that the cooling system is switched on throughout the
value of PPD is considered, and therefore the hourly PPD value whole day, but only when the indoor temperature (Ti) gets higher
can violate such a limit; than the setpoint, i.e., 26 °C in this case. Accordingly, Fig. 4 shows

as concerns the two scenarios related to optimization 3 – i.e., e) the daily trend of hourly Ti values during the examined day apply-
and f) – the considered 90 % acceptability status of ASHRAE 55 ing the reference control, outlining the hours when the cooling sys-
[36] generally corresponds to PPD  10 % [36] like the comfort tem is switched on and off, respectively. Notably, the cooling
category II of SFS-EN 15,251 [47]. Therefore, DHashrae = 0 is system is turned off during most night hours because the outdoor
associated with category II, while DHashrae  10 % presumably temperature is quite low, falling below 20 °C, thereby making
provides performance similar to category III. redundant the use of night-time set-back for the reference case.
The validated EnergyPlus model provides the following ref. per-
The chosen solution, day-by-day, can be implemented by an formance indicators:
actuator system acting on the operation parameters of space con-
ditioning units, e.g., mass flow rate of heat transfer fluids or fan
CCref = 2.93c€/m2day;
power.
PPDmaxref = 16.0 %;
Notably, the implementation of the Pareto approach is funda-
PPDmeanref = 12.6 %;
mental to investigate different optimal solutions produced by the
DHashraeref = 25 %.
same optimization process, depending on the users’ preferences
(i.e., after MCDM), ensuring high flexibility and offering compre- It is noted that the comfort indices PPD and DH are standard
hensive insights about thermal comfort. outputs of EnergyPlus simulations.
Hence, the MPC framework is applied for each couple of objec-
3. Results tive functions:

The application of the model predictive control (MPC) framework
optimization 1 ? CC vs PPDmax;
is simulated for the case study – BNZEB – referring to the day July
optimization 2 ? CC vs PPDmean;
21, 2021. The assumed electricity price – including all taxes – is
optimization 3 ? CC vs DHashrae.
0.22 €/kWhel, which was roughly the average value for Italian
households in 2020/2021 [48]. The monitored weather data are used In the days before the application day, the cooling system oper-
to generate the weather data file, simulating weather forecasts. ates according to the reference strategy.
Addressing only one day can appear not comprehensive, but the
aim of this study is proposing and testing the methodology for a typ-
ical day of the summer season to understand the potential benefits 3.1. Convergence analysis
as concerns energy/cost savings and thermal comfort, providing
insights on different comfort models. In real-time applications A kind of convergence analysis is carried out to investigate the
within a building automation system, the MPC framework needs sensitivity of the achieved Pareto fronts to the number of genera-
to be applied day by day to achieve a day-ahead optimization of tions, as shown in the Figs. 5–7 for optimizations 1, 2 and 3, respec-
HVAC control. In this theoretical study, performing optimization
for all days of the summer season is not feasible from a computa-
tional viewpoint and goes beyond the aims. However, future studies
will address the application to different climates and weather sce-
narios to fully characterize the potentiality of this framework.
The achieved optimal control strategies are compared with a
reference case (ref.), characterized by a fixed cooling setpoint tem-
perature (24 h) equal to 26 °C, assuming a passive approach as con-
cerns occupants’ behavior (see section 2.2). This reference case has
been chosen for the following reasons:

it is typical of Italian residential buildings, especially in the age

of COVID 19 pandemic, which forces people to spend most
time in their homes requiring 24 h space conditioning of build-
ings; the rising of outdoor temperatures due to climate change
(values around 40 °C have been reached in Benevento during
the ongoing summer 2022) worsens the situation making cru-
cial a careful and constant micro-climatic control to ensure
occupants’ comfort and health; 26 °C is the conventional cool-
ing setpoint temperature according to the Italian standards
[49];

as shown in section 1, several previous studies [20,21,25,31]
compared the proposed MPC frameworks with basic strategies
at fixed setpoint in order to quantify the potential benefits com-
pared to a typical, rudimental and not-efficient control; Fig. 4. Reference case: hourly values of indoor temperature.

8
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

tively. The figures represent the Pareto fronts achieved after 25

generations (gen.), 50 gen. and at the end of the GA running. For
optimization 3, convergence is reached before the maximum num-
ber of generations (i.e., 100).
The results demonstrate that:

in all three cases the optimal solutions are consistent since the
difference between consecutive Pareto fronts tends to decrease;
indeed, the fronts related to 50 gen. and maximum gen. are
almost overlapped;

accordingly, 50 gen. are sufficient to attain robust suboptimal
solutions;

in order to reduce the computational time of the optimization
process, 25 gen. are sufficient to achieve acceptable suboptimal
solutions, even if the results are clearly worse because the Par-
eto front is completely dominated by the fronts related to
higher numbers of generations. In this regard, using 25 gen.,
the computational time is reduced from 73.6 h to 18.4 h. There-
fore, it becomes smaller than the planning horizon (24 h), mak-
ing the framework feasible for real applications.

The Pareto fronts related to the maximum number of genera-

tions are used for multi-criteria decision-making (MCDM) in order
to show the full potential of the proposed MPC framework.
Fig. 6. Convergence analysis of Pareto optimization 2: CC vs PPDmean.

3.2. Pareto optimization 1: CC vs PPDmax

As concerns optimization 1, MCDM is carried out by constrain-

ing the value of PPDmax using two limits:

a) PPDmax  10 %, associated with the comfort category II of

SFS-EN 15251 [47]: ‘‘Normal level of expectation and should
be used for new buildings and renovations”;
b) PPDmax  15 %, associated with the comfort category III of
SFS-EN 15251 [47]: ‘‘An acceptable, moderate level of expec-
tation and may be used for existing buildings”.

With regard to scenario a), Fig. 8 shows the achieved Pareto

front highlighting the selected optimal solution and the reference
case (ref.), while Fig. 9 represents the hourly values of cooling set-

Fig. 7. Convergence analysis of Pareto optimization 3: CC vs DHashrae.

point temperature (Tsp) over the planning horizon provided by

such an optimal solution. This latter is characterized by the follow-
ing performance indicators, which are all better (i.e., lower) than
the reference ones:

CC = 2.71c€/m2day:

PPDmax = 9.7 %;

PPDmean = 7.6 %;

DHashrae = 0.

With regard to scenario b), the Pareto front and the optimal Tsp
values are represented in Figs. 10 and 11, respectively. The selected
Fig. 5. Convergence analysis of Pareto optimization 1: CC vs PPDmax. optimal solution features the following performance indicators,

9
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

Fig. 10. Pareto optimization 1: CC vs PPDmax. Multi-criteria decision making based

Fig. 8. Pareto optimization 1: CC vs PPDmax. Multi-criteria decision making based on PPDmax  15 (comfort category III [47]) (scenario b).
on PPDmax  10 (comfort category II [47]) (scenario a).

Fig. 11. Optimal solution setting PPDmax  15 (comfort category III [47]) for
optimization 1: CC vs PPDmax (scenario b).
Fig. 9. Optimal solution setting PPDmax  10 (comfort category II [47]) for
optimization 1: CC vs PPDmax (scenario a).

3.3. Pareto optimization 2: CC vs PPDmean

which are all better (i.e., lower) than the reference ones also in this As concerns optimization 2, MCDM is carried out by constrain-
case: ing the value of PPDmean using two limits:

CC = 2.17c€/m2day; c) PPDmean  10 %, associated with a ‘‘relaxed” comfort cate-

PPDmax = 14.5 %; gory II of SFS-EN 15,251 [47], because the mean daily value

PPDmean = 10.1 %; of PPD is considered, and therefore the hourly PPD value can

DHashrae = 16.7 %. violate such a limit;
d) PPDmean  15 %, associated with a ‘‘relaxed” comfort cate-
As expected, the optimal solution b) provides a lower thermal gory III of SFS-EN 15,251 [47].
comfort level compared to a) but reduced cooling cost, well repre-
senting the human dilemma between ‘‘spending less” and ‘‘feeling Again, with regard to scenario c), the Pareto front and the opti-
better”. mal Tsp values are represented in Figs. 12 and 13, respectively. The

10
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

Fig. 12. Pareto optimization 2: CC vs PPDmean. Multi-criteria decision making Fig. 14. Pareto optimization 2: CC vs PPDmean. Multi-criteria decision making
based on PPDmean  10 (relaxed comfort category II [47]) (scenario c). based on PPDmean  15 (relaxed comfort category III [47]) (scenario d).

Fig. 13. Optimal solution setting PPDmean  10 (relaxed comfort category II [47])
for optimization 2: CC vs PPDmean (scenario c).
Fig. 15. Optimal solution setting PPDmean  15 (relaxed comfort category III [47])
for optimization 2: CC vs PPDmean (scenario d).

selected optimal solution features the following performance indi-

cators, showing lower cooling cost than ref., with a similar comfort
level:

CC = 1.64c€/m2day;

CC = 2.11c€/m2day;
PPDmax = 42.9 %;

PPDmax = 17.7 %;
PPDmean = 14.9 %;

PPDmean = 10.0 %;
DHashrae = 37.5 %.

DHashrae = 25 %.
As expected, optimization 2, provides lower cooling cost but
With regard to scenario d), the Pareto front and the optimal Tsp worse thermal comfort compared to optimization 1, because the
values are represented in Figs. 14 and 15, respectively. The selected comfort category is ‘‘relaxed” and high hourly values of PPD can
optimal solution features the following performance indicators, occur. Therefore, this scenario fits cases with less rigorous comfort
showing significantly lower cooling cost than ref., with a worse requirements and higher flexibility of occupants’ behavior and
comfort level: expectations.

11
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

3.4. Pareto optimization 3: CC vs DHashrae e) DHashrae = 0 corresponding, as aforementioned, to comfort

category II of SFS-EN 15251;
As concerns optimization 3, the ASHRAE 55 adaptive comfort f) DHashrae  10 %, which presumably provides performance
model with 90 % acceptability status [36] is considered. Such a similar to comfort category III of SFS-EN 1525I.
model assesses the acceptability of indoor conditions as a function
of monthly mean outdoor temperature and indoor operative tem- Again, with regard to scenario e), the Pareto front and the opti-
perature. It is here used to compare Fanger comfort theory with the mal Tsp values are represented in Figs. 16 and 17, respectively. The
adaptive one. Accordingly, a 90 % acceptability status is considered selected optimal solution features the following performance indi-
because it generally corresponds to PPD  10 % [36] like the com- cators, showing lower cooling cost than ref., with a similar comfort
fort category II of SFS-EN 15251 [47]. Notably, the objective func- level according to the Fanger theory, but a higher level according to
tion is the percentage of hours not falling in the 90 % the adaptive one:
acceptability status, denoted with DHashrae (discomfort hours).
MCDM is carried out by constraining the value of DHashrae using
CC = 2.10c€/m2day;
two limits:

Fig. 16. Pareto optimization 3: CC vs DHashrae. Multi-criteria decision making Fig. 18. Pareto optimization 3: CC vs DHashrae. Multi-criteria decision making
based on DHashrae = 0 (comfort category II [47]) (scenario e). based on DHashrae  10 (comfort category III) (scenario f).

Fig. 17. Optimal solution setting DHashrae = 0 (comfort category II [47]) for Fig. 19. Optimal solution setting DHashrae  10 (comfort category III [47]) for
optimization 3: CC vs DHashrae (scenario e). optimization 3: CC vs DHashrae (scenario f).

12
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

PPDmax = 17.0 %;

DHashrae

23–24

PPDmean = 11.1 %;

25.5

26.5
26.5
22.9
16.7 %

37.5 %

25

DHashrae = 0.

8.3 %
25 %

25 %
0

22–23
With regard to scenario f), the Pareto front and the optimal Tsp

24.5
25.5
24.5

25.5
25.5
24.4
values are represented in Figs. 18 and 19, respectively. The selected
optimal solution features the following performance indicators,

21–22
25.5

24.5

25.5
25.5
showing lower cooling cost than ref., with a worse comfort level

25

25

26
according to the Fanger theory, but a higher level according to
the adaptive one:

20–21
25.5

25.5

25.5
25.5
27.8
26

CC = 1.98c€/m2day;

PPDmean

PPDmax = 52.6 %;

19–20
%
%
%

%
%
%
%

25.5
25.5
25.5
25.5
29.1

PPDmean = 13.6 %;

25
26
7.6
12.6

14.9
11.1
13.6
10.1
10.0

DHashrae = 8.3 %.

18–19
24.5
26.5

31.5
26
26
26
26
4. Discussion

17–18
The achieved results are summarized in Table 5 as concerns

26.5
26.5
25.5
26
26

34
performance indicators and Tsp values provided by the optimal
solutions in all scenarios investigated, denoted with the letters a)-f).

16–17
The reference case (ref.) is reported too for comparison Table 5.

26.5
26.5
26.5
25.5
25.5
34.9
24
In all cases, the optimal solutions determine a reduction of the

PPDmax
cooling cost compared to ref. case. Scenarios a) and b) – optimiza-

%
%
%
%

%
%

15–16
14.5
17.7
42.9

52.6
9.7
16.0

17.0

25.5
25.5

25.5
25.5
35.5
tion 1 – provide the lowest reductions but the highest comfort

25

27
levels, because they set stricter thermal comfort requirements,
especially as concerns scenario a) (comfort category II).

14–15

25.5
26.5

25.5
25.5
34.8
The adaptive comfort theory – optimization 3, scenarios e) and

25

27
f) – provide similar results to the Fanger theory with ‘‘relaxed”
comfort categories – optimization 2, scenarios c) and d) – showing

13–14

25.5
25.5
25.5
25.5
34.6
that the ‘‘relaxation” of the categories emulates an adaptive occu-

24
25
pants’ behavior. In this regard, scenarios c) and e) – which refer to

12–13
category II – determine thermal comfort levels similar to ref. with a

25.5
25.5

25.5
25.5
25.5
32.8
24
cooling cost reduction of around 28 %.
Cooling Cost [c€/m2a]

Furthermore, the optimal control strategies are consistent with

11–12

the following observations:

24.5

32.1
24

24
24
26
26

during nights and early mornings (0:00–9:00) the cooling sys-
10–11

tem is switched off because the outdoor temperature (Tout) is

24.5
24.5
24.5
24.5

30.4
2.93
2.71
2.17
2.11
1.64

1.98
2.10

25
lower, except for scenarios a)-c) that feature stricter comfort 25

requirements. Notably, scenarios b) and c) envisage the system

9–10

24.5

25.5
25.5

switched on from 0:00 to 1:00 with a very low Tsp in order to

24

25
25

27

drastically cool the indoor environment, thereby avoiding that

24.7
8–9

in the next hours the indoor temperature increases excessively.

Scenario a) envisages the system switched on from 0:00 to 3:00
to ensure higher thermal comfort (comfort category II). From
21.7
7–8

0:00 to 1:00 Tsp is lower for the same reason aforementioned

and because Tout is higher. Notably, in such cases, the cooling
17.6
6–7
d) optimal solution CC vs PPDmean with PPDmean  15 %
c) optimal solution CC vs PPDmean with PPDmean  10 %

f) optimal solution CC vs DHashrae with DHashrae  10 %

system is switched on even if Tout is lower than Tsp in order

b) optimal solution CC vs PPDmax with PPDmax  15 %
a) optimal solution CC vs PPDmax with PPDmax  10 %

e) optimal solution CC vs DHashrae with DHashrae = 0

to balance the load due to the heat stored in the building ther-
16.1
5–6

mal mass during the previous day;

Tsp tends to be lower when the cooling system is switched on
Tsp: cooling setpoint temperature [°C]

16.5
4–5

after the night break – i.e., from 9:00 to 10:00 – in order to dras-
tically cool the indoor environment, thereby ‘‘preparing” the
ref.) Reference: fixed setpoint at 26 °C

building to the high heat loads of the next hours, when Tout
17.6
3–4
Reference case and optimal solutions.

rises, reducing the peak load of the system and increasing the
efficiency. Accordingly, the control is optimized thanks to load
18.4
2–3
25

predictions over the planning horizon. This is the essence of

model predictive control.
19.3
1–2

*Outdoor temperature.
25

In summary, the proposed MPC framework can ensure energy/-

24.5

19.8

cost savings by maintaining or improving the thermal comfort

0–1
24

24

level of the reference case. The user keeps his role of main actor
because he/she can set the comfort requirements to be fulfilled.
Tout*
Table 5

However, for real applications, the framework should be enhanced

d)
b)
a)

e)
c)

f)

according to the following points:

13
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

Fig. 20. Potential application and upgrading of the proposed MPC framework.

the computational time required by the optimization process solutions based on artificial intelligence. Also, the controllers
should be reduced, e.g., by using machine/deep learning meth- should be enhanced to be more flexible in order to take into
ods as simulation tool; account unexpected changes in weather forecasts and/or building

the procedure should be integrated in a building automation use, using dynamic planning and application horizons. Further-
system for achieving a real-time optimized control of space more, for a real implementation, different hardware (e.g., sensors,
cooling systems. Thus, hardware and software equipment actuators) and software (e.g., monitoring tool, weather/occupancy
should be carefully designed, and a mobile app can be devel- forecasting tool, simulation tool, optimization tool) need to be inte-
oped to ensure the best interaction with the users. In this grated in a holistic framework to attain effective, digital and smart
regard, the forecasts of occupancy profiles can be integrated controllers able to feel and satisfy the occupants’ needs under the
to attain customized and more efficient control strategies, and banner of sustainability. Accordingly, model predictive controllers
the setpoint temperatures can be optimized for each thermal can couple the big revolutions of our times, i.e., energy and digital
zone based on such profiles. A vision of the potential application transitions. This is challenging also because often the main barrier
and upgrading of the proposed MPC framework is sketched in is the people mindset, which is hard to change. However, model
Fig. 20. predictive controllers can make it possible because the potential
energy, economic and comfort benefits are huge, with potentially
The framework can provide sentient space conditioning sys-
low investments (payback times). Future studies will try to address
tems able to feel and satisfy the occupants’ needs under the banner
such weak points to achieve a real implementation of the proposed
of sustainability. Substantial benefits can stem, concerning both
conceptual MPC framework.
private and public perspectives. Savings in costs and comfort
improvement can be ensured for the single individuals, decrease
in energy and carbon footprint of buildings can be ensured to the 5. Conclusions
collectivity. Accordingly, since building space conditioning is liable
for over 10 % of final energy uses and CO2-eq emissions at Euro- This study simulates the application of an optimization-based
pean and World levels [2,4], large-scale implementation of the framework for the model predictive control (MPC) of space cooling
framework, and generally of optimization-based model predictive systems to a real nearly zero energy building located in Benevento
controllers, can have a huge impact on energy transition toward – Southern Italy, Mediterranean climate. The framework provides
sustainability and carbon neutrality. This is why research is pivotal optimal values of setpoint temperatures on a day-ahead planning
to bring MPC into our homes. Definitely, the applicability of such horizon to minimize energy cost and thermal discomfort, based
controllers is not easy because there are some barriers/weak points on weather forecasts. Accordingly, a Pareto optimization is imple-
than need to be overcome, but research is struggling to promote a mented by coupling a genetic algorithm – working in MATLABÒ
widespread implementation of MPC to new and existing buildings. environment – and a validated EnergyPlus model. Different objec-
To achieve this goal, controllers need to be efficient, easy to install tive functions related to thermal comfort are investigated to com-
to different terminals (e.g., radiators, fan coils, heat pumps), smart, pare the Fanger theory with the adaptive one of ASHRAE 55. Thus,
flexible for different users and destinations, feasible for large-scale different two-objective optimizations (energy cost vs comfort) are
deployment. As concerns the proposed framework, the main tech- implemented, and optimal solutions are selected to fit different
nical issues can concern the building energy simulation/prediction comfort categories, i.e., II and III of SFS-EN 15251.
tool, which must be at the same time reliable and fast in predic- The proposed case study refers to a specific day of the summer
tions, i.e., requiring low computational burden. Machine/deep season, i.e., 21st July 2021. The monitored weather data are used to
learning methods can be a solution in this regard providing MPC simulate weather forecasts. The optimal MPC solutions are com-
14
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

pared with a reference control at the fixed setpoint of 26 °C. In this References
regard, the main conclusions can be drawn as follows:
[1] N. Asim, M. Badiei, M. Mohammad, H. Razali, A. Rajabi, L. Chin Haw, M.
Jameelah Ghazali, Sustainability of Heating, Ventilation and Air-Conditioning

the parameters of the optimization algorithm should be care- (HVAC) Systems in Buildings—An Overview, International Journal of
fully set in order to reduce the computational time, which is Environmental Research and Public Health 19(2) (2022) 1016.
essential in MPC applications. Accordingly, the results show [2] Data available online at: https://ec.europa.eu/info/news/focus-energy-
efficiency-buildings-2020-lut-17_en (last accessed on 27/04/2022).
that using 25 generations of the genetic algorithm – instead [3] C.I. Jareño Escudero, M. Navarro Escudero, C.D. Mifsut García, M. Flores Fillol, J.
of 100 – can be sufficient to achieve acceptable suboptimal M. Salmerón Lissen, Potential of energy savings in the public housing stock of
solutions, even if the outcomes are clearly worse; 50 genera- comunitat valenciana region by applying the MedZEB cost-optimal
methodology, Appl. Sci. 12 (1) (2021) 138.
tions provide the best compromise;
[4] IEA 2021 (International Energy Agency) Report. Data vailable online at:

the optimal control strategies are able to predict the cooling https://globalabc.org/sites/default/files/2021-10/GABC_Buildings-GSR-2021_
load, thereby optimizing the hourly values of setpoint temper- BOOK.pdf (last accessed on 26/12/2021).
atures as a function of weather forecasts; [5] F. Oldewurtel, A. Parisio, C.N. Jones, D. Gyalistras, M. Gwerder, V. Stauch, B.
Lehmann, M. Morari, Use of model predictive control and weather forecasts for

the optimal solutions always imply a reduction of the cooling energy efficient building climate control, Energy Build. 45 (2012) 15–27.
cost compared to the reference case; [6] W.H. Kwon, S.H. Han, Receding Horizon Control: Model Predictive Control for

the Fanger comfort theory with ‘‘relaxed” comfort categories – State Models, Springer Science & Business Media, 2006.
[7] J. Drgoň, J. Arroyo, I.C. Figuero, D. Blum, K. Arendt, D. Kime, E.P. Ollé, J. Oravech,
i.e., considering the mean daily value of PPD (predicted percent- M. Wetter, D.L. Vrabi, L. Helsen, All you need to know about model predictive
age of dissatisfied) – provide results similar to the adaptive control for buildings, Ann. Rev. Control 50 (2020) 190–232.
comfort one. In this regard, by referring to the solutions [8] Y. Yao, D.K. Shekhar, State of the art review on model predictive control (MPC)
in Heating Ventilation and Air-conditioning (HVAC) field, Build. Environ. 200
achieved for the comfort category II, the cooling cost is reduced (2021).
by around 28 % compared to the reference case maintaining a [9] L. Magni, R. Scattolini, Robustness and robust design of MPC for nonlinear
similar comfort level; discrete-time systems, in: R. Findeisen, F. Allgöwer, L.T. Biegler (Eds.), Lecture
Notes in Control and Information SciencesAssessment and Future Directions of

the optimal solution achieved for the comfort category II using Nonlinear Model Predictive Control, Springer Berlin Heidelberg, Berlin,
the Fanger theory determines a drastic improvement of thermal Heidelberg, 2007, pp. 239–254.
comfort compared to the reference case, reducing, at the same [10] D. Limon, I. Alvarado, T. Alamo, E. Camacho, Robust tube-based MPC for
tracking of constrained linear systems with additive disturbances, Process
time, the cooling cost by around 8 %.
Control 20 (3) (2010) 248–260.
[11] T. Cholewa, C.A. Balaras, J. Kurnitski, A. Siuta-Olcha, E. Dascalaki, R. Kosonen, ...,
Finally, the framework can ensure high energy and cost savings M. Cakan, Energy Efficient Renovation of Existing Buildings for HVAC
by maintaining or improving the thermal comfort level. Neverthe- professionals, REHVA GB No.32, ISBN 978-2-930521-31-2 (2022).
[12] S. Yang, M.P. Wan, W. Chen, B. Feng Ng, S. Dubey, Model predictive control
less, upgrades are required for the application in building automa- with adaptive machine-learning-based model for building energy efficiency
tion systems. In this vein, future studies will address the following and comfort optimization, Appl. Energy 27 (2020).
tasks: [13] M. Mork, A. Xhonneux, D. Müller, Nonlinear distributed model predictive
control for multi-zone building energy systems, Energy Build. 264 (2022).
[14] H. Lee, Y. Heo, Simplified data-driven models for model predictive control of
- use of machine/deep learning methods as a simulation tool in residential buildings, Energy Build. 265 (2022).
order to reduce the computational time of the optimization pro- [15] J. Hou, H. Li, N. Nord, G. Huang, Model predictive control under weather
forecast uncertainty for HVAC systems in university buildings, Energy Build.
cess, which is still excessive; 257 (2022).
- careful design of software and hardware equipment for a real [16] H. Li, S. Wang, Comparative assessment of alternative MPC strategies using
implementation of the MPC framework. real meteorological data and their enhancement for optimal utilization of
flexibility-resources in buildings, Energy 122693 (2021).
[17] M.D. Knudsen, L. Georges, K.S. Skeie, S. Petersen, Experimental test of a black-
Research is needed to open our doors to large-scale deployment box economic model predictive control for residential space heating, Appl.
of model predictive controllers, which can play a crucial role in the Energy 298 (2021).
[18] A. De Lorenzi, A. Gambarotta, M. Morini, M. Rossi, C. Saletti, Setup and testing
energy and digital transition of the building sector.
of smart controllers for small-scale district heating networks: An integrated
framework, Energy 205 (2020).
[19] S. Freund, G. Schmitz, Implementation of model predictive control in a large-
CRediT authorship contribution statement sized, low-energy office building, Build. Environ. 197 (2021).
[20] S. Salakij, N. Yu, S. Paolucci, P. Antsaklis, Model-based predictive control for
Fabrizio Ascione: Conceptualization, Methodology, Software, building energy management. I: energy modeling and optimal control, Energy
Build. 133 (2016) 345–358.
Data curation, Investigation, Writing – review & editing. Rosa [21] M. Ławryńczuk, P. Ocłoń, Model predictive control and energy optimisation in
Francesca De Masi: Conceptualization, Methodology, Software, residential building with electric underfloor heating system, Energy 182
Data curation, Investigation, Writing – review & editing. Valentino (2019) 1028–1044.
[22] G. Bianchini, M. Casini, D. Pepe, A. Vicino, G.G. Zanvettor, An integrated model
Festa: Conceptualization, Methodology, Software, Data curation,
predictive control approach for optimal HVAC and energy storage operation in
Investigation, Writing – review & editing. Gerardo Maria Mauro: large-scale buildings, Appl. Energy 240 (2019) 327–340.
Conceptualization, Methodology, Software, Data curation, Investi- [23] M. Aftab, C. Chen, C. Chau, T. Rahwan, Automatic HVAC control with real-time
gation, Writing – review & editing. Giuseppe Peter Vanoli: Con- occupancy recognition and simulation-guided model predictive control in
low-cost embedded system, Energy Build. 154 (2017) 141–156.
ceptualization, Methodology, Software, Data curation, [24] R. Kumar, M.J. Wenzel, M.N. ElBsat, M.J. Risbeck, K.H. Drees, V.M. Zavala,
Investigation, Writing – review & editing. Stochastic model predictive control for central HVAC plants, J. Process Control
90 (2020) 1–17.
[25] M. Avci, M. Erkoc, A. Rahmani, S. Asfour, Model predictive HVAC load control in
Data availability buildings using real-time electricity pricing, Energy Build. 60 (2013) 199–209.
[26] Y. Du, Z. Zhou, J. Zhao, Multi-regional building energy efficiency intelligent
regulation strategy based on multi-objective optimization and model
Data will be made available on request. predictive control, J. Cleaner Prod. 349 (2022).
[27] T. Cholewa, A. Siuta-Olcha, A. Smolarz, P. Muryjas, P. Wolszczak, R. Anasiewicz,
C.A. Balaras, A simple building energy model in form of an equivalent outdoor
Declaration of Competing Interest temperature, Energy Build. 236 (2021).
[28] T. Cholewa, A. Siuta-Olcha, A. Smolarz, P. Muryjas, P. Wolszczak, Ł. Guz, C.A.
The authors declare that they have no known competing finan- Balaras, On the short term forecasting of heat power for heating of building, J.
Cleaner Prod. 307 (2021).
cial interests or personal relationships that could have appeared
to influence the work reported in this paper.
15
F. Ascione, R.F. De Masi, V. Festa et al. Energy & Buildings 278 (2023) 112664

[29] T. Cholewa, A. Siuta-Olcha, A. Smolarz, P. Muryjas, P. Wolszczak, Ł. Guz, M. [39] US Department of Energy. Energy Efficiency and Renewable Energy Office,
Bocian, C.A. Balaras, An easy and widely applicable forecast control for heating Building Technology Program, EnergyPlus (version 9.4). Available online at:
systems in existing and new buildings: First field experiences. Journal of https://energyplus.net/ (last accessed on 27/04/2022).
Cleaner Production 352 (2022) 131605. [40] ASHRAE, 2014. ASHRAE Guideline 14-2014: Measurement of Energy and
[30] S. Yang, M.P. Wan, W. Chen, B.F. Ng, S. Dubey, Experiment study of machine- Demand savings. American Society of Heating, Refigerating and Air-
learning-based approximate model predictive control for energy-efficient Conditioning Engineers, Atlanta (2014).
building control, Appl. Energy 288 (2021). [41] F. Ascione, N. Bianco, C. De Stasio, G.M. Mauro, G.P. Vanoli, Simulation-based
[31] A. Afram, F. Janabi-Sharifi, A.S. Fung, K. Raahemifar, Artificial neural network model predictive control by the multi-objective optimization of building
(ANN) based model predictive control (MPC) and optimization of HVAC energy performance and thermal comfort, Energy Build. 111 (2016) 131–144.
systems: a state of the art review and case study of a residential HVAC system, [42] C.D. Corbin, G.P. Henze, P. May-Ostendorp, A model predictive control
Energy Build. 141 (2017) 96–113. optimization environment for real-time commercial building application, J.
[32] M. Esrafilian-Najafabadi, F. Haghighat, Impact of occupancy prediction models Build. Perform. Simul. 6 (3) (2013) 159–174.
on building HVAC control system performance: application of machine [43] G. Henze, C. Felsmann, G. Knabe, Evaluation of optimal control for active and
learning techniques, Energy Build. 257 (2022). passive building thermal storage, Int. J. Therm. Sci. 43 (2) (2004) 173–183.
[33] S. Yang, M.P. Wan, W. Chen, B.F. Ng, S. Dubey, Model predictive control with [44] D. Lazos, A.B. Sproul, M. Kay, Development of hybrid numerical and statistical
adaptive machine-learning-based model for building energy efficiency and short term horizon weather prediction models for building energy
comfort optimization, Appl. Energy 271 (2020). management optimization, Build. Environ. 90 (2015) 82–95.
[34] M.G. Yu, G.S. Pavlak, Extracting interpretable building control rules from [45] V.M. Zavala, E.M. Constantinescu, T. Krause, M. Anitescu, On-line economic
multi-objective model predictive control data sets, Energy 240 (2022). optimization of energy systems using weather forecast information, J. Process
[35] K. Deb, Multi-objective optimization using evolutionary algorithms, Vol. 2012 Control 19 (10) (2009) 1725–1736.
John Wiley & Sons, Chichester (2001). [46] MathWorks, MATLAB – MATrixLABoratory (version 2015). https://
[36] ASHRAE, 2010. ASHRAE Guideline 55- 2010: Thermal Environmental it.mathworks.com/products/matlab.html (last accessed on 27/04/2022).
Conditions for Human Occupancy. Atlanta: American Society of Heating, [47] SFS-EN 15251, 2007: Indoor environmental input parameters for design and
Refigerating and Air-Conditioning Engineers, Atlanta (2010). assessment of energy performance of buildings addressing indoor air quality,
[37] F. Ascione, N. Bianco, F. De Rossi, R.F. De Masi, G.P. Vanoli, Concept, design and thermal environment, lighting and acoustics (2007).
energy performance of a net zero-energy building in mediterranean climate, [48] Data available at: https://ec.europa.eu/eurostat/databrowser/view/
Procedia Eng. 169 (2016) 26–37. nrg_pc_204/default/table?lang=EN (last accessed on 12/05/2022).
[38] F. Ascione, M. Borrelli, R.F. De Masi, F. de Rossi, G.P. Vanoli, A framework for [49] Italian Standard UNI/TS 11300-1 ‘‘Prestazioni energetiche degli edifici – Parte
NZEB design in Mediterranean climate: design, building and set-up monitoring 1: Determinazione del fabbisogno di energia termica dell’edificio per la
of a lab- small villa, Sol. Energy 184 (2019) 11–29. climatizzazione estiva ed invernale” (2014). [in Italian].

16