Content uploaded by Marko Jelić
Author content
All content in this area was uploaded by Marko Jelić on Feb 24, 2023
Content may be subject to copyright.
Autonomous Demand Response Control using Heat
Pumps in Residential and Commercial Buildings
Dayanne Peretti Correa
IRUSE - Research Group
National University of Ireland - Galway
Galway, Ireland
dayanne.peretti@nuigalway.ie
Marko Jeli´
c
Institute Mihajlo Pupin
University of Belgrade
Belgrade, Serbia
marko.jelic@pupin.rs
Dea Puji´
c
Institute Mihajlo Pupin
University of Belgrade
Belgrade, Serbia
dea.pujic@pupin.rs
Shima Yousefi
IRUSE - Research Group
National University of Ireland - Galway
Galway, Ireland
s.yousefigarjan1@nuigalway.ie
Marcus Keane
IRUSE - Research Group
National University of Ireland - Galway
Galway, Ireland
marcus.keane@nuigalway.ie
Nikola Tomasevic
Institute Mihajlo Pupin
University of Belgrade
Belgrade, Serbia
nikola.tomasevic@pupin.rs
Abstract—The energy used for heating and cooling has signifi-
cant impact on the electricity bills of residential and commercial
buildings, and heat pumps have been installed as a solution to
reduce these costs. In Europe, around 11% of the buildings
already have heat pumps installed, but there is still a lack
of optimization of their usage profile to maintain the thermal
comfort of the buildings and the equipment efficiency. More-
over, buildings with photovoltaic (PV) energy generation have
additional flexibility that can be explored, but the operational
complexity also increases, which makes finding the optimal profile
to enhance self-consumption challenging. Techniques for energy
optimization and building modeling can help to identify the best
energy profile in an automated way, facilitated by IoT devices
and advanced communication infrastructure. The objective of
this paper is to provide a framework to perform autonomous
demand response control actions and demonstrate a use case for
improving the usage of heat pumps. This includes the data to
be collected for the simulation of thermal patterns and to create
the optimal curve of energy usage in two real scenarios. The
achievements of this study show that remote access to the system
data can allow for enhanced energy usage, through the utilization
of building modeling and electric energy optimization models.
Index Terms—Demand Response, Heat Pump, PV Production,
Building Simulation.
INTRODUCTION
Nowadays, heat pump optimization is studied by manu-
facturers with a focus on increasing the efficiency of new
products. However, according to the Federation of European
Heating, ventilation, and Air conditioning associations, in
2021 the European heat pump stock amounts to 14.86 million
units and, putting this into the perspective of between 115 and
120 million residential buildings in Europe, the heat pump
market share in the building stock is about 11% [1].
The purchase of renewable energy technologies has an
extensive decision-making process due to uncertainties related
The research presented in this chapter is partly financed by the European
Union (H2020 REACT project, Grant Agreement No.: 824395 and SINERGY
project, Grant Agreement No.: 952140) and the Ministry of Education, Science
and Technological Development of the Republic of Serbia.
to financial returns [2]. These technologies are not a ”casual
purchase” and, as one of the most expensive items in a
household, people that commit with these purchases want them
to work effectively. However, these systems need some ”fine-
tuning” after their installation, with periodic maintenance in
terms of maintaining the product efficiency and adjustments
for thermal comfort based on the domestic routine [3].
According to De Wilde’s study, insufficient evaluation af-
ter the installation of a retrofit measure by contractors was
identified as a concern [2]. Heat pumps without the correct
setting may use more energy than needed to maintain the
comfort level of a building. Users without specific technical
knowledge may face issues when changing the patterns set
during the installation, especially if additional equipment is
added to the control loop, such as PV generation, that rely on
weather conditions and vary across the day.
Building performance simulation has the potential to re-
duce the environmental impact of the built environment, thus
improving indoor quality, as well to facilitate innovation and
technical progress in construction [4]. Creating a simulated
environment that can predict the behaviour of a building in
terms of energy usage can help to improve real scenarios
with a less invasive approach. As an example, the COVID-19
restrictions brought new challenges for the installation sector,
where the access to buildings became more restricted. On top
of that, technological challenges also became evident as non-
technical users are not confident in installing equipment or
changing configurations in their own houses without technical
supervision. In this regard, remote access to the data and the
possibility of autonomous control of appliances can bring ben-
efits to end users, through a more optimized system operation.
This paper proposes a framework to allow a remote control
operation of heat pumps in two different pilot sites, consider-
ing elements for optimizing their profile of usage. The main
components to be depicted are the pilot characteristics, the
data requirements, and the cloud platform.
I. ELECTRIC ENERGY OPTIMIZATION
In the context of demand response (DR) integration as
explored by this paper, the main focus is placed on the electric
domain and the process of electrification of various types
of loads with the goal of providing a more efficient control
system. On the receiving end of the controls that are aimed to
be delivered by the described system are prosumer residential
units as well as community and commercial facilities. These
have a number of flexibilities provided by either on-site storage
systems or flexible loads that can be exploited in a number of
ways.
The optimization is posed as a mixed-integer linear pro-
gramming (MILP) problem where variables depict energy
flowing to or from different components of the system under
corresponding constraints. The energy balance between the
loads Pload, net power exchanged with the grid Pgrid and
battery Pbatt as well as production from the local photovoltaic
(PV) array PPV at time step tis maintained by an equality
constraint given by
(∀t)Pload(t) = Pgrid (t) + ηinvPPV (t) + Pbatt(t),(1)
where ηinv is the estimated linear efficiency of the inverter.
In general, each variable of the model is constrained to values
between a lower and upper bound. For example, for curtailable
PV production
(∀t) 0 ≤PPV(t)≤Pforecast
PV (t),(2)
i.e., the optimization is free to choose any non-negative value
smaller than a forecasted maximum provided by an external
service. Of interest is also the possibility to utilize an assumed
amount of load flexibility from appliances as given by
(∀t)Plb
load(t)≤Pload (t)≤Pub
load(t),(3)
with the bounding values being calculated based on a fore-
casted load curve (again, provided by an external service)
Pforecast
load (t)and an assumed amount of flexibility f(t). How-
ever, since the optimization is set to minimize the operating
costs Jbased on current energy import rates α(t)and export
revenue β(t)as given by
J=X
t
hα(t)Pimport
grid (t)−β(t)Pexport
grid (t)i,(4)
an integral load balance constraint must be added as given by
X
t
Pload(t) = X
t
Pforecast
load (t)(5)
such that all load decrease events are balanced out with
corresponding load increase events during a predefined horizon
(e.g., 24 hours).
The key output of the optimization algorithm is the sug-
gested set of load modifications given by
∆Pload(t) = Poptimal
load (t)−Pforecast
load (t).(6)
This curve provides the basis for further assessment of what
particular control actions over appliances, or in the particular
case discussed within this paper, heat pumps, can be selected
to best fit the desired consumption pattern.
II. BUILDING MODEL
The development of building models is becoming popular
due to their capacity of creating a virtual representation of a
real-world building. The models help to better understand the
thermal losses of any building type, such as residential units,
commercial facilities, and industries. One of the main benefits
is that operators can test control strategies in a simulated
environment, aiming to find an optimized usage profile of the
building assets, without affecting the normal operation. For
instance, if a specific building has a heat pump system and on-
site PV generation, control strategies can be tested to identify
the best profile that keeps the user’s comfort at the lowest
possible cost. Testing such strategies in real scenarios can lead
users to face discomfort (indoor temperatures too low or too
high) or expensive electricity bills due to wrong configurations,
as finding the best usage profile can be a complex task.
There are three different main approaches when designing
a building model, as illustrated in Fig. 1 from [5], and the
decision of the best one relies on a number of factors, such
as available data, accuracy, deployment time, and associated
costs.
Fig. 1. Building modeling approaches.
The first approach is called white-box model (left side of
the figure), in which a detailed model of the building is created
using the physical knowledge of the building, such as di-
mensions, building construction materials characteristics, sub-
models of other systems and equipment, and also additional
information about local climate conditions. White-box models
can be very accurate, but also time-consuming and costly, as
getting all the necessary information may be hard to find or
sometimes nonexistent, especially for old buildings. On the
right side of Fig. 1 there is the black-box model, which is
based on measured or historical data and uses statistical and
mathematical methods. This method does not require extensive
model inputs, thus requiring a shorter development time. The
outputs are given by estimates from a pre-trained model,
which can vary from simple statistical models to complex
machine learning techniques. Finally, the centre of the figure
shows the grey-box model, which combines characteristics
from both white and black-box approaches. It uses simplified
physics equations compared to a white-box, hence reducing
the requirement of extensive training sets and model input
parameters. In summary, grey-box models aim to achieve high
accuracy with a quick deployment and execution time.
Regardless of the model chosen, data collection is an
important part of the process, as it is mandatory for most grey
and black-box models and is also important for improving
the results of white-box models. An example can be seen
in experiments performed by authors in [6], where a hybrid
approach has been selected. First, a detailed white-box model
of a dwelling in the Aran Islands was developed mainly
considering the building thermodynamics and the heat pump
system. Then, a simplified grey-box was developed to add
new features to the model in a simpler manner by using
data collected from IoT devices, such as real PV generation,
energy consumption and indoor temperatures. Finally, the
control strategy used a machine learning black-box approach
to identify the best profile of usage.
III. USE CASES’FRAMEWORK
The use cases described in this section were designed to
be performed in a number of dwellings in Ireland and Italy,
that are part of the REACT project [7]. Due to climate
conditions, the Irish use case has the objective of improving
the heat pumps operation for heating, while the Italian is
focused on cooling. Moreover, both test cases consider the
utilization of the electric energy optimization (Section I) and
the information provided by the building simulator (Section
II), aiming to change the pattern of the heat pump usage profile
and guaranteeing the achievement of comfort levels with a
more efficient usage of the onsite produced energy.
These use cases will be applied in a similar way as the
simulated demand response test case 03, also in the Aran
Islands, as described by authors in [9]. In these simulations,
the results showed an improvement in renewable energy usage
of 39.14%. However, instead of using only PV production as
the optimization target, this research considers a more complex
electric energy optimization, as specified in Section I.
A. Pilot characteristics
San Pietro Island is an island 7 kilometres off the south-
western coast of Sardinia, in the province of South Sardinia,
Italy. With an area of 51.1 km2and almost all the inhabitants
of San Pietro concentrated in Carloforte, the reported popula-
tion in 2018 was of 6,173. Electricity is imported to the island
via two submarine 10 MW power cables from mainland. In
2016 the annual electricity consumption in San Pietro was
15,779 MWh. The residential sector represents nearly 59% of
the total electricity consumption [8] and, the buildings where
the use cases are due to be applied in Italy were mostly built
before 1960s and nowadays have efficiency issues.
The Aran Islands are situated in the Atlantic Ocean, 5–12
nautical miles off the west coast of Ireland, in county Galway.
The biggest island is called Inis M´
or, and has 31 km2. With
around 700 regular residents, it has the Aran Islands’ greatest
population. Electricity is imported to the islands via a 3MW
sub-sea cable and, in 2018 a total of 2,993 MWh of electricity
was imported into the Aran Islands’ reclosers. The residential
sector represents around 29% of the total energy usage in the
island [8].
The selected commercial and residential buildings of both
pilots are equipped with heat pumps and onsite PV generation
with some of them already available previously and others
deployed as part of REACT project works.
B. Data requirements
To perform the use cases and find the best operation
schedule that keeps the indoor thermal comfort using the
lowest energy cost, a two-step process is proposed. First,
the electric energy optimization creates the optimal energy
consumption curve, considering predicted energy production
and consumption, and energy tariffs. Then, this optimal curve
is used as input for the building model, which will simulate
heat pump actions (heating or cooling), aiming to fit them in
the periods suggested by the optimizer as most suitable, and
also respecting pre-defined temperature setpoints to keep the
users’ comfort. The scope of this paper comprises in the data
and infrastructure needed to enable the mentioned use cases
and methodology, thus the level of detail of the full control-
loop optimization and building simulations are restricted to the
necessary for this purpose. Each of the steps needs a different
set of data to be collected, as depicted below.
Since the optimization framework is based on an Energy
Hub layout, the model needs to be instantiated with static data
that defines the physical characteristics of the included assets.
This includes storage and generation (peak power) capacities
and power flow limitations (maximum charge and discharge
rates of storages as well as grid interconnection limitations).
Furthermore, the model requires dynamic timeseries data in
form of predicted appliance power consumption and an esti-
mate of the produced power in each time step. These curves are
supplemented with a limited flexibility as described in Section
I, such that the optimizer can choose the best possible value
to be suggested to the subsequent models for building thermal
performances.
Regarding building modeling, the first analysis showed that
the buildings chosen do not have complete information about
their construction due to their age. A white-box model is not
recommended for these use cases because a complete survey is
needed for each of the buildings, which can be costly and time
consuming. Moreover, this approach does not generalize well,
as in case of new buildings added to the project would also
mean new detailed models need to be developed. According to
the REACT project deployment, buildings are equipped with
Mitsubishi heat pumps, that received a MELCloud adapter
communication device to enable data to be sent to the REACT
cloud platform. The data available comprises indoor tempera-
tures, heat pump configurations, and equipment consumption
with a 15-min or higher resolution. Along with additional data
from IoT devices (building energy consumption) and external
datasets (weather forecast), a black-box model can be deployed
if enough data has been collected, or it can also be used to
calibrate a grey-box model. These models can be generalized
for other buildings, using a reduced amount of collected data
for the calibration. To simulate the best profile of usage, the
building model also receives the output optimal curve from the
electric energy optimization model, also available on REACT
cloud.
C. Cloud platform
In order to properly support the automatic execution of
the control loop, a decoupled approach is utilized whereby
each service is run independently in a virtualized Docker
container with the inputs sourced from and the outputs stored
into a centralized repository hosted via a cloud platform. The
cloud holds two key databases necessary to run the system:
a relational (MySQL) database utilized to store intermediary
outputs of different services and a NoSQL IoT database
(Influx) that is used to monitor the data coming from edge
devices.
Of particular interest for the case discussed in this paper is
the control actions table of the relational database. This repos-
itory is utilized to consolidate outputs from the optimization
service for direct inverter controls and building models for heat
pump controls. Each row of this table is designated as one
control action that could either specify an operating set point
of some of the devices or toggle their on/off state. For example,
a control could force the inverter to operate the battery and PV
in such a way which as to ensure a certain power amount is
drawn from or exported to the grid. Alternatively, a set point
for the battery (dis)charge rate could be directly specified, a
heat pump could be turned on or off, or a particular set point
(temperature and fan speed) could be specified.
A graphical representation of the control actions can be seen
in figure 2, from a simulated test considering one of the Irish
buildings. The black dotted line represents the estimated usage
from the optimization service, which is used as input for the
building model. Then, simulations are carried out aiming to
perform control actions during the best expected times (i.e.
low energy tariff, high PV production, etc.), and also keeping
the user’s comfort. As a result, the indoor temperature is kept
within values around 21oC (yellow line), and actions (grey
bars) are spread out over the day, matching the expected usage.
Fig. 2. Building simulator overview.
The control action table is constantly being monitored by
an persistently running orchestrator service which compares
the timestamp at which an action is scheduled and the current
time, and when needed, dispatches the encoded message to
the MQTT broker in charge of communicating with connected
gateways. Once the message has been received and the action
acknowledged from an edge node (gateway), the sent status
field is updated with a corresponding indicator.
IV. CONCLUSION
This research presented a framework for autonomous de-
mand response in residential and commercial buildings, con-
sidering three main aspects: pilot characteristics, data require-
ments, and cloud platform. The studies were developed for
two pilots in different locations, also providing a simulation
that showcases the optimization process of a heat pump use
case. Heat pump configuration may be a complex task for
non-technical users, and unexpected challenges such as the
COVID-19 restrictions degraded the way these systems are
operated. Additionally, a potential cost reduction from PV
systems can also be explored if demand and consumption
are properly coordinated, hence the need for a remote and
autonomous control framework.
Starting with the pilot characteristics, they give an important
view about the kind of environment the project will face, which
defines the type of demand response high-level scenarios. This
is related to weather conditions, electric network capacity, and
user profile. Secondly, data requirements are tied with the
devices available for measurement and control, they impact
the real possibility of a DR use case, and also the level of
information available for the type of building modeling and
the optimization service. Finally, the cloud platform is crucial
for orchestrating the services and autonomous control, closing
the process loop.
The test performed shows how the process is seen in the
cloud platform, including the generated optimal energy curve,
the simulated control actions, and the indoor temperatures
within the expected range over the day.
REFERENCES
[1] ”European heat pump market.” Federation of European Heating, ven-
tilation and Air conditioning associations. https://www.rehva.eu/rehva-
journal/chapter/european-heat-pump-market (accessed 2022).
[2] M. de Wilde, ”The sustainable housing question: On the role of
interpersonal, impersonal and professional trust in low-carbon retrofit
decisions by homeowners,” Energy Research & Social Science, vol. 51,
pp. 138-147, 2019, doi: 10.1016/j.erss.2019.01.004.
[3] M. de Wilde, ”“A Heat Pump Needs a Bit of Care”: On Maintain-
ability and Repairing Gender–Technology Relations,” Science, Tech-
nology, & Human Values, vol. 46, no. 6, pp. 1261-1285, 2020, doi:
10.1177/0162243920978301.
[4] J. H. editor. and R. L. editor., Building Performance Simulation for
Design and Operation London ; New York : Routledge, Taylor & Francis
Group, 2019.
[5] Madsen, H., Juhl, P. B. R. & Muller, J. K., 2015. Presentation:
Grey-Box Modeling; An approach to combined physical and statisti-
cal model building, Denmark : DTU Compute Insitut for Matematik
og Computer Science. Availible at: http://smart-cities-centre.org/wp-
content/uploads/grey box intro.pdf.
[6] P. Lissa, C. Deane, M. Schukat, F. Seri, M. Keane, and E. Barrett, ”Deep
reinforcement learning for home energy management system control,”
Energy and AI, vol. 3, 2021, doi: 10.1016/j.egyai.2020.100043.
[7] ”REACT project website.” https://react2020.eu (accessed 2022).
[8] S. J. (VEO), L. M. D. I. (VEO), and J. M. (VEO), ”REACT - Deliverable
1.1 Pilot technical characterisation and regulatory framework,” 2019.
[9] M. Jelic, D. Pujic, N. Tomasevic, P. Lissa, D. Peretti, and M. Keane,
”Case study of Aran Islands: Optimal demand response control of heat
pumps and appliances,” in Industrial Demand Response: Methods, best
practices, case studies, and applications, H. H. Alhelou, A. Moreno-
Mu˜
noz, and P. Siano Eds., 2022.