Professional Documents
Culture Documents
Machine Learning Applications For Intelligent Energy Management: Invited Chapters From Experts On The Energy Field 1st Edition Haris Doukas
Machine Learning Applications For Intelligent Energy Management: Invited Chapters From Experts On The Energy Field 1st Edition Haris Doukas
https://ebookmeta.com/product/machine-learning-and-iot-for-
intelligent-systems-and-smart-applications-1st-edition/
https://ebookmeta.com/product/internet-of-energy-for-smart-
cities-machine-learning-models-and-techniques-1st-edition-jindal-
anish/
https://ebookmeta.com/product/smart-energy-management-data-
driven-methods-for-energy-service-innovation-1st-edition-kaile-
zhou/
https://ebookmeta.com/product/advances-in-digitalization-and-
machine-learning-for-integrated-building-transportation-energy-
systems-1st-edition-yuekuan-zhou/
Intelligent Data Analytics for Power and Energy Systems
1st Edition Hasmat Malik
https://ebookmeta.com/product/intelligent-data-analytics-for-
power-and-energy-systems-1st-edition-hasmat-malik/
https://ebookmeta.com/product/hybrid-energy-systems-for-offshore-
applications-1st-edition-dincer/
https://ebookmeta.com/product/nano-catalyst-for-energy-
applications-1st-edition-rohit-srivastava/
https://ebookmeta.com/product/iot-machine-learning-applications-
in-telecom-energy-and-agriculture-with-raspberry-pi-and-arduino-
using-python-1st-edition-puneet-mathur/
https://ebookmeta.com/product/machine-learning-for-intelligent-
multimedia-analytics-techniques-and-applications-pardeep-kumar-
amit-kumar-singh-eds/
Learning and Analytics in Intelligent Systems 35
Haris Doukas
Vangelis Marinakis
Elissaios Sarmas Editors
Machine Learning
Applications
for Intelligent
Energy Management
Invited Chapters from Experts
on the Energy Field
Learning and Analytics in Intelligent Systems
Volume 35
Series Editors
George A. Tsihrintzis, University of Piraeus, Piraeus, Greece
Maria Virvou, University of Piraeus, Piraeus, Greece
Lakhmi C. Jain, KES International, Shoreham-by-Sea, UK
The main aim of the series is to make available a publication of books in hard copy
form and soft copy form on all aspects of learning, analytics and advanced
intelligent systems and related technologies. The mentioned disciplines are strongly
related and complement one another significantly. Thus, the series encourages
cross-fertilization highlighting research and knowledge of common interest. The
series allows a unified/integrated approach to themes and topics in these scientific
disciplines which will result in significant cross-fertilization and research dissem-
ination. To maximize dissemination of research results and knowledge in these
disciplines, the series publishes edited books, monographs, handbooks, textbooks
and conference proceedings.
Indexed by EI Compendex.
Haris Doukas · Vangelis Marinakis ·
Elissaios Sarmas
Editors
Machine Learning
Applications for Intelligent
Energy Management
Invited Chapters from Experts on the Energy
Field
Editors
Haris Doukas Vangelis Marinakis
School of Electrical and Computer School of Electrical and Computer
Engineering Engineering
National Technical University of Athens National Technical University of Athens
Athens, Greece Athens, Greece
Elissaios Sarmas
School of Electrical and Computer
Engineering
National Technical University of Athens
Athens, Greece
© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2024
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Over the past two centuries, the Industrial Revolution has led to a dramatic increase
in global energy consumption, a trend that clearly continues into the twenty-first
century. Indeed, global energy consumption is estimated to have risen from less than
6.000 TWh in 1800 to over 120.000 TWh in 2000 and to about 180.000 TWh in 2022
[1]. Moreover, projections show that energy consumption will continue to grow well
into the foreseeable future and possibly exceed 250.000 TWh by the year 2050 [2].
This rise in energy consumption is certainly a sign of significant advances in
human civilization and of improvement of living conditions worldwide. However, it
does come with side effects that may have dangerous consequences for humankind
and the entire planet. Indeed, carbon dioxide (CO2 ) emissions and other greenhouse
gasses have been clearly recognized as the main contributors to climate change,
temperature rise and global warming [3]. These emissions are mainly due to the use
of fossil fuels as energy source, and they have been constantly rising from about
28.000.000 tons in 1800 [4] to almost 40.000.000.000 tons in 2022 [5] and already
over 32.000.000.000 tons in the first three quarters of 2023 [2].
Governments around the World, including the European Union, are addressing
climate change with high priority [6]. Indeed, energy transition from energy produc-
tion based on fossil fuels (oil, natural gas, or carbon) to renewable and environmen-
tally friendlier energy sources (solar energy, wind energy, hydrogen-based energy
production, or use of lithium-ion batteries) has been set as a common target of many
states worldwide. However, about 83% of consumed energy still comes from fossil
fuels, while the use of environmentally friendlier sources of energy is rising at a
slower-than-required pace [2] raising concerns even from oil giants such as BP [7].
Fortunately, the development and availability of important streamlined technolo-
gies, including artificial intelligence, big data, internet of things and blockchain tech-
nologies, can provide powerful tools towards intensifying efforts, accelerating energy
transition and getting all energy stakeholders actively involved when decisions are
made regarding energy production, distribution and management.
As chief editor of the Learning and Analytics in Intelligent Systems Series (LAIS)
series of Springer, I am particularly happy to present the book at hand, on Machine
Learning Applications for Intelligent Energy Management, which is one of the very
v
vi Series Editors’ Foreword
first of its kind. The book has been edited by three outstanding researchers, namely
Haris Doukas, Vangelis Marinakis and Elissaios Sarmas, who are renowned for their
contributions to the fields of decision-making and policy designing in the energy
sector. Its publication in our series aims at filling a gap in the literature on the use of
the most advanced artificial intelligence, big data, internet of things and blockchain
technologies in the energy sector and to verse the scientific community in the most
recent relevant advances.
More specifically, the book discusses both artificial intelligence-empowered
analytics of energy data and artificial intelligence-empowered application devel-
opment. The book consists of a preamble and an additional seven chapters written by
leading experts. Topics covered include a presentation of the various stakeholders in
the energy sector and their corresponding required analytic services, such as state-
of-the-art machine learning, artificial intelligence and optimization models and algo-
rithms tailored for a series of demanding energy problems and aiming at providing
optimal solutions under specific constraints.
Overall, the book is very well written and constitutes a valuable guide for both
the experts in the field and the newcomers. The former will be updated on the most
recent advances in terms of challenges and solutions regarding energy transition and
the transformation of the energy sector into an environmentally friendlier human
activity. On the other hand, the newcomers will also benefit from this book, as they
will obtain knowledge and develop practical skills. Finally, the book will certainly
attract the interest of readers from other areas as well, who wish to get versed in this
significant scientific discipline.
As series editor, I welcome this monograph to the Learning and Analytics in Intelli-
gent Systems Series of Springer and present it to the research communities worldwide.
I congratulate the editors for their superb work, in confidence that their book will help
its readers not only understand, but also apply the proposed methodologies in various
energy transition problems. Finally, I encourage the editors to continue their research
work in this important area and keep the scientific communities appropriately updated
on their research results.
References
1. https://ourworldindata.org/energy-production-consumption
2. https://www.theworldcounts.com/challenges/climate-change/energy/global-energy-consum
ption
3. https://www.youtube.com/watch?v=ipVxxxqwBQw
4. https://ourworldindata.org/co2-emissions
5. https://www.energyinst.org/statistical-review
Series Editors’ Foreword vii
6. Paris agreement. Report of the Conference of the Parties to the United Nations Framework
Convention on Climate Change (21st Session, 2015: Paris). Retrieved December, volume 4,
2017 (https://unfccc.int/process-and-meetings/the-paris-agreement)
7. “There is a growing mismatch between societal demands for action on climate change and the
actual pace of progress, with energy demand and carbon emissions growing at their fastest rate
for years. The world is on an unsustainable path.”—Spencer Dale, BP chief economist, 2019.
Preamble: AI-Powered Transformation
and Decentralization of the Energy Ecosystem
During the last decade, the global society has confronted the challenge of climate
change. The escalating repercussions of climate shifts, coupled with the relent-
less surge in carbon dioxide (CO2) emissions, have compelled profound shifts on
both national and international fronts. This transformation stands as an imperative
response to combat the devastating impacts of climate change and drive the crucial
process of decarbonizing our energy sector.
Decarbonization, in essence, signifies the imperative to diminish our dependence
on carbon-centric energy sources. It necessitates a methodical transition toward a
pioneering energy paradigm, meticulously structured to curtail carbon emissions.
The urgency of this transition becomes undeniably clear when we consider that
as of 2019, global CO2 emissions exceeded the 1990 levels by a staggering 60%,
signifying the inception of earnest climate negotiations. These emissions continue
to be the chief driver of the relentless global warming phenomenon.
Significantly, 2019 etched a disconcerting milestone with record-breaking CO2
emissions. Yet, there was a glimmer of hope in 2020, a year fraught with the COVID-
19 pandemic, when stringent lockdown measures led to a noteworthy reduction in
emissions. It is universally acknowledged that the most direct route to decarboniza-
tion hinges upon a sweeping shift toward harnessing Renewable Energy Sources
(RES). This clarion call has found a formidable proponent in the form of the Inter-
national Renewable Energy Agency (IRENA). The agency has meticulously charted
a comprehensive roadmap, envisioning a significant reduction in carbon emissions
by 2050. Consequently, a multitude of nations have already embraced a substantial
share of renewable resources, such as wind and solar energy, to meet their burgeoning
energy demands.
Simultaneously, technological advancements over the past decade have empow-
ered energy consumers, facilitating the decentralization of the energy system. Decen-
tralization involves the placement of energy production facilities closer to consump-
tion sources, as opposed to the centralized energy generation model characterized
by large-scale power plants. This transition is underpinned by the adoption and
widespread use of RES to reduce dependence on fossil fuels, as discussed in the
previous sections. The current relevance of decentralization is paramount, leading
ix
x Preamble: AI-Powered Transformation and Decentralization …
Stavros Stamatoukos
CINEA Horizon Europe Energy
Contents
xiii
xiv Contents
Abstract Climate change and energy production and consumption are two inextri-
cably linked concrete concepts of great concern. In an attempt to guarantee our future,
the European Union (EU) has prioritized the addressing of both concepts, creating a
new social contract between its citizens and the environment. The dazzling progress
in its methodologies and applications during the recent years and the familiarization
of the public with its abilities indicate Artificial Intelligence (AI) as a potential and
powerful tool towards addressing important threats that climate change imposes.
However, when using AI as a tool, it is vital to do so responsibly and transparently.
Explainable Artificial Intelligence (xAI) has been coined as the term that describes
the route of responsibility when implementing AI-driven systems. In this paper, we
expand applications that have been previously built to address the problem of energy
production and consumption. Specifically, (i) we conduct a survey to key stakehold-
ers of the energy sector in the EU, (ii) we analyse the survey to define the required
depth of AI explainability and (iii) we implement the outcomes of our analysis by
developing a useful xAI framework that can guarantee higher adoption rates for our
AI system and a more responsible and safe space for that system to be deployed.
1 Introduction
Since the Paris Agreement [1, 2], there is increasing international concern and action
taken with regard to climate change and viability of Earth. Indeed, the Paris Agree-
ment focuses on the design and the application of viable, effective, socially acceptable
and fair policies to fight and possibly reverse climate change on a global level [3].
As a result, radical changes are effected on the energy sector, which affect all of its
stakeholders.
These changes are collectively referred to as “Energy Transition”, which is a term
implying the transition from energy production based on fossil fuels (oil, natural gas
or carbon) to renewable and environmentally friendlier energy sources (solar energy,
wind energy, hydrogen-based energy production or use of lithium ion batteries).
Energy Transition is not restricted to decarbonisation only, but it also incorporates
multiple social, technological and environmental targets [4]. In more detail, Energy
Transition is founded on four pillars, namely the, so-called, 4D’s of the energy
sector, which specifically stand for Decarbonization, Digitization, Decentralization,
and Democratization.
In more detail, Decarbonization refers to the reduction of the dependence on
carbon for energy production and its gradual replacement with other sources of
energy, such as Renewable Energy Sources (RES). Decarbonization is a high priority
in Energy Transition, as it will slow down both climate change and the rate of increase
in carbon dioxide (CO2) emissions [5]. Indeed, CO2 emissions in 2019 were over
60% higher than 1990 emissions. On the other hand, there was a significant decline
in CO2 emissions in 2020, which was due to the strict lockdowns enforced by many
governments around the World to contain the COVID-19 pandemic [6]. These two
facts clearly indicate CO2 emissions as a main cause for climate change. The ultimate
goal of Decarbonization is a World economy that does not produce CO2 emissions.
To measure progress towards this goal, the Paris Agreement has established several
actions and performance indices, which, however, are still far from being met [3].
On the other hand, Decentralization refers to the construction of energy pro-
duction systems that are physically located near energy consumers, as opposed to
large-scale, but centrally located ones [7]. Decentralization is commensurate with
Decarbonization as Decentralization is based on the extensive use of RES and reduc-
tion of the dependence on fossil fuels. Decentralization also depends on the existence
of active consumers who not only consume energy, but also possess the ability to, at
least partially, produce and store energy and manage energy demand [8].
Democratization comes as a consequence of Decentralization. Indeed, the decen-
tralized model is largely based on the ability of consumers to produce, store and
manage the energy they consume, for example when they have installed photovoltaic
equipment on their roof. The term “prosumer” (a combination of the two terms “pro-
ducer” and “consumer”) has been coined to indicate any energy consumer with their
own means to produce part of the energy they consume [9]. Energy Democratiza-
tion is expected to produce significant local economic value as it will allow both
prosumers and SMEs to have reliable access to low-cost energy.
Digitization refers to the use of electronic tools, systems, devices and resources
which produce, store and process data and extract meaningful and useful information
[10]. Data, in particular, are widely available and of very high quality and can be
used towards optimizing functions and processes at the household, building, work
environment, community or city levels [11, 12]. As with most big data applications,
energy data are characterized by the, so-called, 5V’s, namely, very high V olume, great
An Explainable AI-Based Framework for Supporting Decisions in Energy … 3
V ariety, high rate of new data collection (V elocity), increased reliability (V eracity),
and the ability of information technologies to add V alue to them.
Clearly, the transition from a producer/provider-centered system to a consumer/
household-centered system is a very challenging goal [13]. Fortunately, it coincides
with the development and availability of important streamline technologies, including
Artificial Intelligence (AI), Internet of Things and Blockchain Technologies. These
technologies offer a unique opportunity to intensify efforts and accelerate Energy
Transition.
Particularly important is the active involvement of all energy stakeholders and
their participation in decision making with regard to the production, distribution
and management of energy. Towards this, decision support systems (DSSs) need
to be made available to the various stakeholders, which are stakeholder-tailored
and provide decision support at various levels and in various forms. On one hand
predictive models and optimization algorithms need to be incorporated in them,
which rely on state-of-the-art Machine Learning (ML) and AI technologies [14, 15].
On the other hand and in order to be adopted by the various classes of their users and to
exclude unacceptable decisions, these DSSs need to incorporate explainable AI (xAI)
and responsible AI technologies, which provide the system users with justification
of the recommended actions [16, 17]. As explainability and responsibility of an
AI system are concepts related to an audience, i.e. to a class of users or even to
specific individual users [16], the various stakeholders (i.e. classes of users) need to
be identified first.
In this paper, we present a DSS for supporting decisions on energy management
and energy efficiency in buildings. The system is based on a set of state-of-the-art ML
and AI models which have been tailored for a series of demanding energy problems,
as well as on a set of optimization algorithms aiming at providing the optimal solu-
tion under specific constraints. The novelty of the DSS lies in the use of xAI modules
which have been developed based on conducting a survey among key stakeholders
of the energy sector in the EU and analyzing its results. The development and imple-
mentation of the outcomes of our analysis into a useful xAI framework can guarantee
higher adoption rates for our AI system and a more responsible and safe space for that
system to be deployed as it provides well-established justification of the recommen-
dations presented to its users. This is especially important for critical applications in
the energy sector, such as energy load management or RES production forecasting,
where the consequences of incorrect predictions can be quite significant. The pro-
posed xAI system can provide insights into how AI models make predictions (e.g.
about future energy demand or supply), assisting energy companies and regulators
understand the factors that drive energy demand and make informed decisions. In
comparison to existing DSSs that often lack transparency and interpretability making
it difficult for users to assess the reliability of the results, the proposed xAI system
provides a clear and understandable explanation of the predictions being made and
the factors that contribute to them.
More specifically, the paper is organized as follows: Sect. 2 is devoted to a state-
ment of various problems that are associated with energy production, distribution
and consumption, an outline of the various stakeholders and their relation to cor-
4 E. Sarmas et al.
Current challenges that researchers are facing and addressing in the energy sector can
generally be grouped into three general categories, namely (1) managing renewable
energy sources, (2) establishing and managing distributed energy resources and (3)
improving building energy efficiency. In the following, these challenges are briefly
analyzed further.
Managing renewable energy sources: Driven by European and international con-
cern about climate change and action taken to reduce Greenhouse emissions by 55%
by the year 2030 [18], energy production from RES, especially from those based on
wind and solar power, is increasing at a fast rate. Indeed, it is expected that, by the
year 2026, up to 95% of the increase in global power generation will come from RES
[19]. This translates to an increase of over 60% in electrical power generation from
renewable sources during the years 2021 to 2026, with construction of 25% more
new wind-energy parks than new win-energy parks constructed during 2015–2020
[19].
Clearly, new tools need to be developed and made available to optimize the energy
production from RES, as well as its distribution and management. ML and AI tech-
nologies seem to provide an indispensable tool towards this goal. More specifically,
challenges that need to be addressed by researchers include:
– Predicting energy production from RES in the very short- (up to 30 mins ahead),
short- (up to one hour ahead), medium- (up to one day ahead), and long- (more
than one day ahead) term, based on predictive methodologies that process such
measurements as humidity, temperature, cloudiness, wind speed and direct or
diffuse solar irradiance.
– Detecting faults and applying prognostic maintenance, in order to minimize out-
of-service time and reduce relevant cost.
– Optimizing the location of RES, especially photovoltaic panels and wind-energy
units, based on measurement of various parameters, such as solar irradiance or
An Explainable AI-Based Framework for Supporting Decisions in Energy … 5
wind potential, and taking into account other factors, such as environmental or
social.
Establishing and managing distributed energy resources: A distributed energy
resource consists of a number of small energy production units, such as photovoltaic
panels or wind-energy units, which reside on the consumer side. These distributed
resources, when combined with energy storage units and managed in a way that takes
flexible electric load demands [20], form smart (micro-)grids.
Several challenges arise with regard to distributed energy resources forming smart
micro-grids, that researchers need to address:
– Predicting energy demand, i.e. predicting upcoming energy load, either in the short-
(a few minutes, a day or a week ahead) or in the medium- (up to a year ahead) or
in the long- (several years or more ahead) term. This is a crucial requirement for
the reliable operation of a distributed energy resource [21].
– Scheduling demand response, i.e. asking consumers to reduce energy consumption
during specified time intervals in order to reduce the strain on the energy grid,
save on energy costs, reduce the use of fossil fuels for energy production, and
incorporate renewable energy resources into the grid [22].
– Optimizing flexible loads and demand side managing with practices that go beyond
demand response, attempt to alter consumer behavior and possibly do not require
intermediate energy storage. Such practices include peak shaving, valley filling
and load shifting [23].
– Developing more efficient energy storage systems, ranging from the develop-
ment of new materials to hybrid methods in which advanced algorithms may
be employed to monitor battery health, reduce battery discharge or improve the
frequency of required battery charging [24–26],
Clearly, AI, especially ML, and Optimization are fields that can provide powerful
methodologies and tools to address the above-mentioned challenges in efficient ways
and to provide solutions that can be made available and embraced by all stakeholders
involved in the energy sector.
Several stakeholders are involved in the energy sector, who require intelligent energy
management systems that provide them with different groups of analytics services.
In this subsection, we outline eleven (11) key stakeholders, while in the following
subsection we present eight (8) analytics services, which are associated with stake-
holders and have been fully implemented.
Figure 1 illustrates key energy stakeholders along with their corresponding impor-
tance. The description and the roles of these stakeholders is analyzed in the following.
1. KS01—Producers (PD): A producer produces the energy1 that will be con-
sumed by households and businesses. Conventional means of electricity produc-
tion include nuclear power plants, combined cycle gas turbines and coal plants,
while renewable electricity production comes from biomass power plants, hydro-
electric power stations, wind farms and solar parks. The latter producers can be
further categorized into individual park owners and aggregators who may
control several solar parks.
2. KS02—Suppliers (SP): A supplier sells energy to households and businesses,
without necessarily being a producer as well. Today, a large number of suppliers
are in operation due to the opening of the energy market to competition.
3. KS03—Balance Responsible Party (BRP): A BRP is a private legal entity that
overlooks the balance of one or multiple access points to the transmission grid.
4. KS04—Transmission System Operator (TSO): The role of the TSOs is to
carry electricity from power plants and the different delivery points either to
the distribution networks or straight to industrial customers. The TSOs are also
obliged to ensure the overall balance of the network, i.e., to continuously balance
production and consumption. The required investments for a TSO are colossal
which forbids opening this business to competition.
5. KS05—Distribution System Operators (DSO): These stakeholders form the
last link in energy delivery, as they construct and manage medium and low volt-
age/pressure networks to liaise between the transmission networks and private
dwellings. Moreover, in the event of problems with the energy meter or power
failure, the DSO rather than the supplier should be contacted.
6. KS06—Regulatory Bodies (REG): As some stakeholders enjoy a legal
monopoly status, bodies have to be created to control and regulate the energy
1 In this work, the term “energy producer” refers to an electricity producer, but, in more general
terms, the term may as well refer to a natural gas producer.
An Explainable AI-Based Framework for Supporting Decisions in Energy … 7
market. Their role is to ensure transparency and competition in the energy mar-
ket, defend consumer interests, advise the authorities on energy matters, and
certify the operation of energy markets.
7. KS07—Electricity Customers, Consumers, Residents of Buildings, Occu-
pants (BUI): These stakeholders constitute entities that consume energy at var-
ious scales.
8. KS08—Project Developers, Investing Funds, Governmental Institutions
(INV): These stakeholders constitute entities that are looking for the best
financing opportunities in investing in energy efficiency projects (e.g. renova-
tions/refurbishments in buildings) with the goal of reducing carbon emissions at
the building, district, or city level [37].
9. KS09—Aggregators (AGG): An aggregator is a new type of energy service
provider which can increase or moderate the electricity consumption of a group
of consumers according to total electricity demand on the grid. An aggregator
consists of a grouping of agents in a power system, which acts as a single entity
when engaging in the electricity market. The aggregator’s role is to gather flex-
ibility from the prosumers’ devices and sell it to KS05-DSOs, KS03-BRPs, and
(either directly or through a KS03-BRP) to KS04-TSOs.
10. KS10—Facility Managers, Building Operators (FMB): Building Operators
ensure that the heating, cooling, mechanical and electrical equipment of a build-
ing is running effectively. FMB duties may include inspecting the building for
safety hazards and regulation violations, performing repairs, and checking the
ability of the facility to operate successfully and in the most efficient manner.
8 E. Sarmas et al.
11. KS11—Energy Managers (ENM): Energy managers are responsible for han-
dling heating, ventilation and air conditioning (HVAC) systems of large, multi-
storey buildings, so as to ensure that thermal comfort levels are kept within
acceptable ranges.
Figure 2 is illustrative of the various analytic services towards intelligent energy man-
agement, developed previously in related work. In this study, we examine explain-
ability approaches to better communicate each service functionality to the stakehold-
ers described in Fig. 1, to increase adoption and optimise usability. Each service is
defined by its inputs and outputs.
In detail, AS01 [38, 39] focuses on photovoltaic (PV) production forecasting,
which involves the use of PV production time series, historical weather data, and
numerical weather predictions as input to estimate the energy generation of the PV
system. This estimation can be done for short-term (e.g., an hour ahead) or mid-term
(e.g., a week ahead) intervals.
AS02 [40] deals with consumption forecasting at the building level, utilizing
building consumption time series as input to estimate the future energy consumption
of a building.
AS03 [41] is concerned with load forecasting of the grid. It employs building
consumption time series, historical weather data, and numerical weather predictions
as input to estimate the future grid load.
3 AI Explainability
xAI refers to the level and depth that the decision process of a trained (ML) model is
explained and described. When the process is adequately explained via a framework
of textual, visual and tabular paradigms, the usually referred to as “black box” in
ML is disassembled. The users are prawn to increase adoption of systems, when
misconceptions about the underlying technologies have been unraveled.
The key characteristics of a xAI system are fairness, ethics, transparency, pri-
vacy, security, accountability and safety. When those characteristics are addressed,
then the explainable framework is referred to as responsible. Systems that can infer
intelligence are required to have an adequate framework, where the underlying tech-
nology that outputs decisions and recommendations can be adequately and properly
explained. Due to the increasing social impact of such intelligent systems, xAI is
considered a prerequisite. xAI is also an important feature that increases adoption.
Familiarity with technology and AI literacy should be considered in the xAI frame-
work, design process according to [45], where layers of scientific language and
simplified explanations should be combined and applied accordingly. For that pur-
pose cross-discipline knowledge and the mental capacity of the user should be taken
into account and served accordingly [46].
The main concern to be addressed is related to the reasons that would make a user
trust ML models to make predictions, automate classifications and support decision
making. If the proposed solutions that are served by a machine are aligned with user
expectations, a need for deeper explanations may be overlooked. In any case though,
it is the developers’ responsibility to offer sufficient description and a road map to
the decisions suggested by the system [47].
System transparency is a basic social demand and on par with the need for data privacy
and data security [48]. It is important that the user can identify the characteristics of
the models utilised as the backbone of the AI system. The algorithms deployed, that
lead to a decision, have to be adequately represented [49]. If these requirements are
met, the system is considered responsible, inclusive and transparent.
Global model explanation is the process that takes place in the training phase of
the ML pipeline. It involves deep data analysis, measurement of bias and semantic
grouping of inputs and outputs, the analysis of the adopted ML methodology, fea-
ture importance and feature dependency, the classification report and the confusion
An Explainable AI-Based Framework for Supporting Decisions in Energy … 11
matrix. The global explainability method is, essentially, a way to give the user the
clarifications required as per the inputs of a ML model and the expected outcomes
(output) when those inputs are analysed.
The technology acceptance model [50] (TAM) theorizes that acceptance of a new
computer system is related to two external variables, namely the perceived usefulness
and the perceived ease of use. Those two variables are considered measurements of
individual intention to use a specific technology and, thus, determine the likelihood
of adoption of said technology.
Perceived usefulness refers to the users’ belief that a certain technology can deliver
value. Perceived ease of use refers to the level of effort that a user would be required
to undertake to use the technology. TAM is extracted from the theory of reasoned
action [51], which assumes that to predict user behaviour, those two external variables
should be linked to specific intentions (Fig. 4).
We use TAM to align survey outcomes with application deployment require-
ments to ensure higher adoption rate and smaller learning curves. For that purpose,
12 E. Sarmas et al.
The questionnaire is split into three sections. The first one is related to the demo-
graphic characteristics of the participants, including age, gender, educational level,
occupation and employment space (public, private or both). The aim of the second
section is to define the AI literacy level (perceived usefulness) of the participants,
while the third section attempts to define the perceived ease of use of the provided
AI tools. In the second and third sections, both qualitative and quantitative ques-
tions were included. The quantitative questions were used as scoring components
and facilitated the clustering of the specialists that participated in the survey. On the
other hand, the qualitative questions were used as descriptors of the market trends
and the general attitude towards AI infusion in the energy sector. In Tables 1, 2 and 3,
the different questions are separated in a way that reflects the described process.
4.3 Demographics
Based on [52, 53] and since the goal is to address AI specific usability and design
concerns, a number of 20 to 40 participants would suffice to draw reliable and high
quality results. Thus, the survey was conducted among 20 specialists working in the
energy sector, with 30% of the participants being employed in a public company,
45% in a private company and 25% both in a public and private company. Their
average employment duration was 2 years with a minimum of 1 and a maximum of
5 years in this particular sector. 75% of the participants are identified as men and
25% as women and work or/and live in the EU (Fig. 5). The primary language of
the participants is mainly Italian(6), followed by Spanish(5). The average age of the
participants was 36.6 years with a minimum age of 22 years and a maximum age of
58 years. The highest educational level of the participants was that of Ph.D status (2
out of 20), with the majority holding a master’s degree (15 out of 20) and the rest
having acquired relevant certifications or equivalent to certifications degrees (3 out
of 20). The electronic questionnaire was anonymously filled in, no detailed personal
An Explainable AI-Based Framework for Supporting Decisions in Energy … 13
data were collected and the participants submitted their consent prior to filling in the
questionnaire.
Table 2 AI usability
Question Perceived usefulness
Q1 In the system use, how easily can you predict what results are given by the
functions of the system (e.g. what happens if you press a button)?
Q2 In the system use, how easily can you understand what the current
situation of the system is based on your previous actions in the system?
Q3 In the system use, how easily could you predict what actions you needed
to perform in similar situations within the system
Q4 In the system use, were menus and figures easy to understand in the whole
of the system?
Q5 In the system use, how easy was to undo some action when needed?
Q6 In the system use, to what extent could you see the results of each of your
actions?
Q7 Were the results compatible with other external sources and expertise?
Q8 Did you find the overall system useful?
Q9 Did you find the overall system easy to use?
Q10 Was the experience of using the system satisfying for you?
Q11 Would you use this system in the future?
and, subsequently, approximates each observation with that mean (centroid) [56].
Since the number of questions/observations are more than 2 and it would be difficult
to define the clusters in the multidimensional space, the PCA reduction technique was
used for simplification purposes. Principal component analysis (PCA) converts the
full features into fewer features based on their contribution in explaining a component
(variance explained) [57].
AI literacy score, clusters and perceived usefulness Using the proposed method-
ology, of . K -means clustering, the silhouette coefficients to define the number of
clusters and PCA technique to visualise those clusters, the answers of the questions
of Table 1 are analysed. The results can be seen in Fig. 6. According to the analysis,
An Explainable AI-Based Framework for Supporting Decisions in Energy … 15
there are two different clusters related to the corresponding participants’ AI literacy
level, as revealed by the silhouette coefficient and seen in Fig. 6 in which the highest
value corresponds to the number of clusters. The primary characteristics, based on
the average score of each answer, can be seen in Table 4. To plot the different clusters,
16 E. Sarmas et al.
we have used 3 components which, as can be seen in the principal component anal-
ysis of Fig. 6, explain more that 58% of the total variables (14). The average score
of the first cluster is 3.45, while that of the second cluster is 2.93. There is greater
separation (above 0.70 decimal points) between the two clusters in Questions 1, 3,
9, 13 and 14. Question 1 is related to the users’ competency in the English language.
Question 3 is related to the years of working experience in the sector. Question 9 cor-
responds to the required explainability of the DSS result and the need of additional
information that is perceived as missing. Question 13 measured the frequency that
the user validated results using external sources and finally Question 14 was related
to the degree of user satisfaction with the results provided by the system.
To summarize, two levels of explainability seem necessary, that correspond to the
characteristics of the two different users. Thus, at least years of working experience,
level of understanding of the English language, and potentially offering a localised
environment based on user language requirements should be taken into considera-
tion. Trialability is also an important factor as described in Questions 13 and 14,
where related literature should be additionally provided and which users should be
encouraged to study. In both clusters AI literacy is either average or below average
(Q2) and for that reason usefulness of the different explainability methodologies
(local—global xAI) should be provided and detailed. The experienced users felt that
system messages were adequate, but the second less experienced group of users
felt that some actions performed by the provided AI pipeline were not sufficiently
explained.
An Explainable AI-Based Framework for Supporting Decisions in Energy … 17
AI tools usability score, clusters and perceived ease of use. In this section, we
assess the perceived ease of use of the AI tools offered. Similarly to the previous
section, a segmentation methodology was used to define the optimum number of
clusters. Specifically, 5 clusters were identified, using the silhouette coefficients,
where 5 is indicated as the highest value of a total 6 cluster scenarios tested. Using
PCA, we reduced the plotted space from 11 components in total in the questions to
three components, i.e. we reduce the feature space down to a 3-dimensional one. The
components explain almost 60% of the variance and the corresponding . K -means
plotted clusters can be seen in Fig. 7. The five different clusters assume different
scores, which indicates different abilities and different perceptions as per the usability
of AI tools offered by our system. Cluster 3, within which gather most of the users,
have a slightly above average perception of 2.98 for the perceived ease of use of
the system AI Clusters 2, 4 and 5 which hold the highest average scores and also
include 35% of the total participants, score above 3.5 which should be considered a
strong indicator of a general acceptance of the offered tools, with room, though, for
improvements.
The questions with the lowest scores are Question 3 in Cluster 1, Question 1 in
Cluster 4 and Question 4 in Cluster 5. These questions are stated as follows:
– In the system use, how easily can you predict what results are given by the functions
of the system (e.g. what happens if you press a button)?
– In the system use, how easily could you predict what actions you needed to perform
in similar situations within the system
18 E. Sarmas et al.
– In the system use, were menus and figures easy to understand in the whole of the
system?
The above mentioned questions indicate a difficulty to understand and navigate
through the exported results of the AI pipeline. Again, here as in the previous section,
an explainability framework, that considers different levels of AI-related capabilities
can indeed offer a greater understanding of the system underlying value and thus
increase adoption via improving perception of usability of AI and a general usefulness
of extracted recommendations and automated suggestions related to the different
energy analytical services provided by the ML pipeline.
It is also important to note that the general sentiment, as reflected by Question 8
on the overall system usefulness, scores the highest for all clusters, indicating that
the needs of the energy sector are well served and identified (Table 5).
Market trends, towards an AI-driven energy sector. The answers provided in the
last portion of the survey can be seen in Fig. 8. The main concern of the participants,
with regard to AI incorporation in the energy sector, is related to data privacy (80%)
and security. The general consensus is that AI will further decentralise the sector
(80%) and will facilitate faster decarbonisation (65%) and increase energy savings
(65%). From the following questions, which are also highly suggestive to the previous
results analysed and also validate the clustering hypothesis, it appears that the most
likely cause that would impede AI incorporation is by a majority of 60% the lack of
explainability (30%) and lack of required training of users to the available software
(30%). The suggested level of explainability should be at a level of limited detail for
non-technical users (45%), followed by a level that corresponds to deeper technical
detail, suitable for more technical users.
An Explainable AI-Based Framework for Supporting Decisions in Energy … 19
The vast majority (45%), expect AI to transform the energy sector in the 30’s,
while 30% see this transformation taking place in the 20’s. Moreover, 10% of the
participants expressed the opinion that AI will never transform the energy sector and
15% that the transformation will occur beyond the 30’s. Again the results signify
two different kind of users in the participant pool.
According to the TAM, analysed in the previous sections, the perceived ease of use
(AI usability) and perception of usefulness (AI literacy level) are communicating
vessels, interconnected in the users’ consciousness. We have identified two different
clusters in our data [58], related to their analytical experience and have quantified
their AI literacy level. More clusters have been proposed based on the perceived ease
of use, where some key points were identified and should be addressed via design
modifications, either via global functionality or via personalised options. In Fig. 9,
the two main clusters can be seen. For each cluster different tools are proposed with
some tools overlapping, with regard to AI explainability. For Cluster 1, a technical
analysis of the ML models should be provided, using local and global xAI, bias
analysis and model architecture characteristics. The related literature should also be
provided. Cluster 2, which consists of less experienced users, should have access to
basic textual description of the process. A semantic grouping of inputs and output
[59] combined with cost benefit analysis and contribution to the circular economy
via AI applicability in the sector, should use provide incentives and build trust.
20 E. Sarmas et al.
To address the problem of perceived usefulness and perceived ease of use, fol-
lowing the proposed methodology outlined in Fig. 9, we offer a preview of screens
used to explain the predictive technology used for one the developed applications.
In Fig. 10, the model architecture of one of the system applications is analysed,
namely analytics service AS04. i.e. estimation of energy savings from renovations in
Fig. 2. An ensemble of key models are used to make predictions based on a probability
of a three class outcome, using as features-inputs (1) the energy consumption before
renovation, (2) the cost of renovation, (3) the planned CO2 reduction as detailed by
technical specifications, (4) the building age in years and (5) the total heating area
defined by the building envelope [14]. Next, the overall model outputs the proba-
bilities for each class. Based on the investment potential, which is the relationship
between energy consumption reduction and cost of investment, the model considers
three classes. For Class A, the potential for investment is optimum, while for Class B
it is medium and the project should be only partially financed. For Class C, the project
should not be financed. The different models are stacked and use linear regression as
the activation function to make predictions. Using an ensemble of models, an average
prediction is provided, ensuring optimised results for a given problem using differ-
ent algorithms, for better definition and a boosted solution overall. In the benchmark
section, the accuracy of each model is calculated, alongside the final result of the
stacked model. Finally, the correlation between energy consumption and investment
cost for each class is shown. In this section, we look into the proposed methods out-
lined in Fig. 10, where model architecture for Cluster 1 and benchmarks for Cluster 2
are touched upon.
An Explainable AI-Based Framework for Supporting Decisions in Energy … 21
Textual descriptions are comments for users to understand the reasoning behind
the development of an application and the personal and macro-economic benefits of
using it. Some key points that have been analysed in this paper are used as descriptors
of usefulness, such as the idea of circular economics and the semantic grouping of
applications and stakeholders. The relation between inputs and outputs [59, 60] is
also an important factor that can increase adoption, add trust to the system and clarify
the system results. In Fig. 11, the semantic relation of inputs and outputs is presented
[14]. In the middle section, the data structure of each input is introduced. Each
element has more information attached to it, for the user to explore, thus offering an
in depth look of how the system analyses data and outputs predictions.
In Fig. 12, selected graphs show how the different features contribute to each decision
for a specific observation from the analysed dataset. We have used shapley values to
determine the different outcomes and the SHAP python libraries [61, 62] to create an
explainability framework for the examined case [63]. According to the observation
analysed, the true actual value refers to the Class B or Class 1. The predicted array,
entitled as predicted in Fig. 12, shows the probabilities for each class, as they are
extracted by the stacked model. For Class 0, the probabilities are equal to 0.16
(0.0155), while for Class 1 they are equal to 0.924 and for Class 2 they are equal
22 E. Sarmas et al.
to 0.06 (0.0597). Our model has correctly predicted the class as 1. To see how
the different features contributed to the prediction, we have used SHAP force and
waterfall plots. In the force plot, the different interactions between the features and
the decision per class is outlined, arriving to the predicted solution-probability of
outcome, symbolised as f(x). On the other hand, the waterfall plots similarly show
the different responses of each parameter. In the bottom of the graph, the expected
An Explainable AI-Based Framework for Supporting Decisions in Energy … 23
output is shown, while the top of the graph indicates how the different features
contributed to the actual prediction extracted by the model. Since Class 1 was indeed
the correct output, in both graphs all features in Class 1 point to the right result and
the proposed investment should be only partially funded.
In Fig. 13, selected graphs show how the different features contribute to the final
prediction for each class. Again, shapley values have been utilised, alongside the
SHAP python libraries for explaining the results visually. Feature importance is
first analysed for each class for all inputs. Cost is the highest contributing factor in
all cases, but the contribution of other factors varies for the different classes. The
dependency plots are then utilised to summarise the dependence between the actual
cost and the actual contribution of the cost to the decision eventually made by the
model, labeled as shap value for cost on the y-axis of each scatter graph. The shap
value for the examined variable shows the extent to which knowing this particular
variable affects the prediction of the model. On the opposite side of the y-axis a
feature that is closely related to the examined variable is tracked and the extent of
the effect to the examined variable is differentiated in color.
Acknowledgements The work presented is based on research conducted within the frame-
work of the project “Modular Big Data Applications for Holistic Energy Services in Buildings
(MATRYCS)”, of the European Union’s Horizon 2020 research and innovation programme under
grant agreement no. 101000158 (https://matrycs.eu/), of the Horizon 2020 European Commission
project BD4NRG under grant agreement no. 872613 (https://www.bd4nrg.eu/) and of the Hori-
zon Europe European Commission project DigiBUILD under grant agreement no. 101069658
(https://digibuild-project.eu/). The authors wish to thank the Coopérnico team, whose contribution,
helpful remarks and fruitful observations were invaluable for the development of this work. The
content of the paper is the sole responsibility of its authors and does not necessary reflect the views
of the EC.
An Explainable AI-Based Framework for Supporting Decisions in Energy … 25
References
1. P. Agreement, Paris agreement, in Report of the Conference of the Parties to the United Nations
Framework Convention on Climate Change (21st Session, 2015: Paris). Retrived December.
vol. 4, HeinOnline (2015), p. 2017
2. J. Tollefson, K.R. Weiss, Nations adopt historic global climate accord: agreement commits
world to holding warming’well below’2 [degrees] c. Nature 582(7582), 315–317 (2015)
3. H. Doukas, A. Nikas, M. González-Eguino, I. Arto, A. Anger-Kraavi, From integrated to
integrative: delivering on the Paris agreement. Sustainability 10(7), 2299 (2018)
4. S. Carley, D.M. Konisky, The justice and equity implications of the clean energy transition.
Nat. Energy 5(8), 569–577 (2020)
5. E. Papadis, G. Tsatsaronis, Challenges in the decarbonization of the energy sector. Energy 205,
118025 (2020)
6. P. Friedlingstein, M. O’sullivan, M.W. Jones, R.M. Andrew, J. Hauck, A. Olsen, G.P. Peters,
W. Peters, J. Pongratz, S. Sitch et al., Global carbon budget 2020. Earth Syst. Sci. Data 12(4),
3269–3340 (2020)
7. A. Hope, T. Roberts, I. Walker, Consumer engagement in low-carbon home energy in the United
Kingdom: implications for future energy system decentralization. Energy Res. Soc. Sci. 44,
362–370 (2018)
8. S. Baidya, V. Potdar, P.P. Ray, C. Nandi, Reviewing the opportunities, challenges, and future
directions for the digitalization of energy. Energy Res. Soc. Sci. 81, 102243 (2021)
9. R. Zafar, A. Mahmood, S. Razzaq, W. Ali, U. Naeem, K. Shehzad, Prosumer based energy
management and sharing in smart grid. Renew. Sustain. Energy Rev. 82, 1675–1684 (2018)
10. P. Weigel, M. Fischedick, Review and categorization of digital applications in the energy sector.
Appl. Sci. 9(24), 5350 (2019)
11. V. Marinakis, H. Doukas, J. Tsapelas, S. Mouzakitis, Á. Sicilia, L. Madrazo, S. Sgouridis, From
big data to smart energy services: An application for intelligent energy management. Futur.
Gener. Comput. Syst. 110, 572–586 (2020)
12. E. Sarmas, N. Dimitropoulos, S. Strompolas, Z. Mylona, V. Marinakis, A. Giannadakis, A.
Romaios, H. Doukas, A web-based building automation and control service, in 2022 13th
International Conference on Information, Intelligence, Systems & Applications (IISA). (IEEE,
2022), pp. 1–6
13. A. Esmat, M. de Vos, Y. Ghiassi-Farrokhfal, P. Palensky, D. Epema, A novel decentralized
platform for peer-to-peer energy trading market with blockchain technology. Appl. Energy
282, 116123 (2021)
14. E. Sarmas, E. Spiliotis, V. Marinakis, T. Koutselis, H. Doukas, A meta-learning classification
model for supporting decisions on energy efficiency investments. Energy Build. 258, 111836
(2022)
15. E. Sarmas, E. Spiliotis, V. Marinakis, G. Tzanes, J.K. Kaldellis, H. Doukas, Ml-based energy
management of water pumping systems for the application of peak shaving in small-scale
islands. Sustain. Urban Areas 82, 103873 (2022)
16. A.B. Arrieta, N. Díaz-Rodríguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garcia, S.
Gil-Lopez, D. Molina, R. Benjamins, R. Chatila, F. Herrera, Explainable artificial intelligence
(xai): concepts, taxonomies, opportunities and challenges toward responsible ai. Inf. Fusion
58, 82–115 (2020)
17. C. Meske, E. Bunde, J. Schneider, M. Gersch, Explainable artificial intelligence: Objectives,
stakeholders, and future research opportunities. Inf. Syst. Manag. 39(1), 53–63 (2022)
18. E. Council, Fit for 55: The eu’s plan for a green transition (2020)
19. IEA: Renewables 2021: Analysis and forecasts to 2026 (2021)
20. P. Skaloumpakas, E. Spiliotis, E. Sarmas, A. Lekidis, G. Stravodimos, D. Sarigiannis, I.
Makarouni, V. Marinakis, J. Psarras, A multi-criteria approach for optimizing the placement
of electric vehicle charging stations in highways. Energies 15(24), 9445 (2022)
21. M.Q. Raza, A. Khosravi, A review on artificial intelligence based load demand forecasting
techniques for smart grid and buildings. Renew. Sustain. Energy Rev. 50, 1352–1372 (2015)
26 E. Sarmas et al.
22. P. Bradley, M. Leach, J. Torriti, A review of the costs and benefits of demand response for
electricity in the UK. Energy Policy 52, 312–327 (2013)
23. C.W. Gellings, W.M. Smith, Integrating demand-side management into utility planning. Proc.
IEEE 77(6), 908–918 (1989)
24. R. Xiong, J. Cao, Q. Yu, Reinforcement learning-based real-time power management for hybrid
energy storage system in the plug-in hybrid electric vehicle. Appl. Energy 211, 538–548 (2018)
25. J. Duan, Z. Yi, D. Shi, C. Lin, X. Lu, Z. Wang, Reinforcement-learning-based optimal control
of hybrid energy storage systems in hybrid ac-dc microgrids. IEEE Trans. Ind. Inf. 15(9),
5355–5364 (2019)
26. G. Zsembinszki, C. Fernández, D. Vérez, L.F. Cabeza, Deep learning optimal control for a
complex hybrid energy storage system. Buildings 11(5), 194 (2021)
27. L. Fuhrimann, V. Moosavi, P.O. Ohlbrock P. D’acunto, Data-driven design: Exploring new
structural forms using machine learning and graphic statics, in Proceedings of IASS Annual
Symposia. vol. 2 in 1. International Association for Shell and Spatial Structures (IASS) (2018),
pp. 1–8
28. D. Nagy, D. Lau, J. Locke, J. Stoddart, L. Villaggi, R. Wang, D. Zhao, D. Benjamin, Project
discover: an application of generative design for architectural space planning, in Proceedings
of the Symposium on Simulation for Architecture and Urban Design (2017), pp. 1–8
29. P. Geyer, S. Singaravel, Component-based machine learning for performance prediction in
building design. Appl. Energy 228, 1439–1453 (2018)
30. M. Huang, J. Ninić, Q. Zhang, Bim, machine learning and computer vision techniques in under-
ground construction: current status and future perspectives. Tunn. Undergr. Space Technol. 108,
103677 (2021)
31. I.K. Brilakis, L. Soibelman, Shape-based retrieval of construction site photographs. J. Comput.
Civ. Eng. 22(1), 14–20 (2008)
32. Z. Zhu, I. Brilakis, Parameter optimization for automated concrete detection in image data.
Autom. Constr. 19(7), 944–953 (2010)
33. C. Fan, Y. Sun, K. Shan, F. Xiao, J. Wang, Discovering gradual patterns in building operations
for improving building energy efficiency. Appl. Energy 224, 116–123 (2018)
34. S. Lu, W. Wang, C. Lin, E.C. Hameen, Data-driven simulation of a thermal comfort-based
temperature set-point control with ashrae rp884. Build. Environ. 156, 137–146 (2019)
35. K. Yan, L. Ma, Y. Dai, W. Shen, Z. Ji, D. Xie, Cost-sensitive and sequential feature selection
for chiller fault detection and diagnosis. Int. J. Refrig 86, 401–409 (2018)
36. J. Granderson, S. Touzani, S. Fernandes, C. Taylor, Application of automated measurement
and verification to utility energy efficiency program data. Energy Build. 142, 191–199 (2017)
37. N. Dimitropoulos, E. Sarmas, M. Lampkowski, V. Marinakis, A quantitative methodology to
support local governments in climate change adaptation and mitigation actions, in International
Symposium on Distributed Computing and Artificial Intelligence (Springer, 2023), pp. 99–108
38. E. Sarmas, N. Dimitropoulos, V. Marinakis, Z. Mylona, H. Doukas, Transfer learning strategies
for solar power forecasting under data scarcity. Sci. Rep. 12(1), 14643 (2022)
39. E. Sarmas, E. Spiliotis, E. Stamatopoulos, V. Marinakis, H. Doukas, Short-term photovoltaic
power forecasting using meta-learning and numerical weather prediction independent long
short-term memory models. Renew. Energy 216, 118997 (2023)
40. E. Sarmas, N. Dimitropoulos, V. Marinakis, A. Zucika, H. Doukas, Monitoring the impact
of energy conservation measures with artificial neural networks, in In ECEEE Summer Study
(2022)
41. E. Sarmas, S. Strompolas, V. Marinakis, F. Santori, M.A. Bucarelli, H. Doukas, An incremental
learning framework for photovoltaic production and load forecasting in energy microgrids.
Electronics 11(23), 3962 (2022)
42. E. Sarmas, E. Spiliotis, N. Dimitropoulos, V. Marinakis, H. Doukas, Estimating the energy
savings of energy efficiency actions with ensemble machine learning models. Appl. Sci. 13(4),
2749 (2023)
43. C. Tsolkas, E. Spiliotis, E. Sarmas, V. Marinakis, H. Doukas, Dynamic energy management
with thermal comfort forecasting. Build. Environ. 237, 110341 (2023)
An Explainable AI-Based Framework for Supporting Decisions in Energy … 27
Keywords Big data value chain · Big data energy analytics · AI-enabled
decision-making services · Big data architecture
Nomenclature
Term Abbreviation
Internet of things IoT
Machine learning ML
Neural network NN
Support vector machine SVM
Online analytical processing OLAP
Deep learning DL
Online transactional processing OLTP
Atomicity, consistency, isolation and durability ACID
Hadoop distributed file system HDFS
International dataspaces IDS
Artificial neural network ANN
Multiple linear regression MLR
Multiple layer perceptron MLP
Building management system BMS
Building information management BIM
Infrastructure as a service IaaS
Platform as a service PaaS
Software as a service SaaS
Identity access management IAM
Generic enabler GE
Context broker CB
Machine to machine M2M
Linear model LM
K-nearest neighbour KNN
1 Introduction
During the Covid pandemic several measures have been proposed by governments
globally to prevent congestion in public places like large scale lockdowns, and
remote working and education was further embraced [1]. Those measures assisted
in reducing the energy demand significantly. It is actually the case that the electricity
generation within 16 European countries in April 2020 had dropped by 9%, whereas
fossil energy generation decreased by 28%, while nuclear energy decreased by 14%
[2]. Despite the decline in the electricity generation, the building sector is responsible
The Big Data Value Chain for the Provision of AI-Enabled Energy … 31
for one fourth of the energy consumed world-wide and 40% of the global carbon
dioxide emissions [3]. The EU has recognized the need for reducing the energy
consumption and tackling its environmental harm and issued the Paris agreement
for achieving carbon neutrality by 2050 [4]. It is therefore made clear that efficient
energy planning is of the outmost importance for reducing energy consumption and
preserving the already overstretched energy sources. Energy analytics can play a vital
role in efficient organization, planning and reduction of energy consumption. Energy
analytics are applied on various fields like IoT [5] and machine learning (ML) [6]
and regard the development of computing techniques for providing high-level user-
driven services such as energy consumption forecasting. Such computing techniques
may include AI techniques like Neural Networks (NN’s), support vector machines
(SVM) and gradient boosting, that can be used for model training in order to predict
the energy consumption under circumstances like particular weather conditions, such
as temperature and humidity [7–9].
Since data collected from sensors on the energy sector is arriving in batches and
is increasing rapidly in size, the need for big data analytics has arouse, taking the
problem of energy analytics to an entirely new level. Big data is hard to conceptualize,
the fact is that except the quality of data there are also other dimensions that need to
be defined like the volume, velocity, variety veracity, value known as the 5 V’s [10]:
Volume indicates the amount of the collected data and since smart meters produce
enormous numbers of data, efficient storing of such amounts is becoming a substantial
problem; velocity indicates the speed to which data is produced and transferred and
the problem is that each meter has its own production and transfer rate; variety
indicates there is a huge range of available data sources which may differ in structure,
so there can be structured data sources where data is stored in specific and strict
format, unstructured, where data is stored in a combination of types maintaining its
original structure and semi-structured, where data is not unstructured but still contains
some structure schemas; veracity indicates the quality of the collected data, in such
vast amounts of data it is highly possible that empty values and errors will be included
affecting the quality and results of analytics; value indicates that the collected data
should be useful for user-driven services and decision-making processes. With the
introduction of big data on AI analytics services of the energy sector the need for
dividing tasks into multiple subtasks executed on a large number of computing nodes
through parallel programming has become compelling to assist in effective storage,
processing and analysis of such huge volumes of data.
Artificial intelligence (AI) aims to develop intelligent machines able of performing
complex tasks not easily performed by human and is used in a plethora of domains
including among others text translation, speech recognition, healthcare, search and
recommendation engines [11]. AI allows the delegation of such complex problems
and contributes to the velocity of data by establishing rapid and accurate decisions.
AI can further allow users of big data to perform and automate complex descriptive
analytical tasks that would be extremely intensive and time consuming to perform
for humans [12]. Employing AI methods on big data can have a huge impact on the
role data plays on various aspects of daily and economic life. The market and public
32 K. Touloumis et al.
bodies adopt new methodologies to exploit and stay competitive by developing AI-
enabled tools and applications. Up to now, AI algorithms have been implemented in
single machines but nowadays with the introduction of big data, the need to increase
the scalability of AI algorithms, making them run on multiple nodes in a parallelized
manner has rapidly emerged [13]. Thus, companies and enterprises are looking for
new and innovative ways of implementing highly distributed AI methodologies that
will be able to receive, process and analyse big data assisting in solving complex
computation problems on energy analytics [14].
Large-scale smart grids consist of thousands of microgrids, with smart meters
being an integral part of them. Smart meters produce densely populated data thus
transforming the problem of “energy analytics” to “energy big data analytics” [15].
Big data methods for AI can prove very beneficial for energy analytics purposes [16,
17]. By exploiting big data and AI, energy analytics can assist in improving smart
meter and smart grid management from a high-level point of view, the accessibility
of users in consumption data by performing statistical analysis and online analyt-
ical services (OLAP) [18]. Energy analytics can also assist on energy forecasting,
analysing the consumer behaviour patterns on energy consumption and estimating
the demand of energy [19]. Furthermore, analytics can analyse historical data on
consumption and estimate energy production [20]. Load prediction to prevent exces-
sive energy consumption under parameters and unnecessary transmission of exces-
sive data can further be performed by exploiting Deep Learning (DL) techniques [6,
21]. Among the included services of big data analytics are services for the predic-
tion of power system failures and sensor downtime increasing the robustness and
efficiency of the power grid systems and meters. Energy analytics can further assist
in tracking the energy production from different sources and taking decisions on
prioritizing consumption demands [22].
Of course, in order to provide efficient energy analytics services, complex big data
technologies and infrastructure should be utilised. In particular, continuous flows of
real-time data from multiple data sources should be ingested, processed, harmonized
and stored to a big data infrastructure, to facilitate efficient, near real-time analytics
services. In this publication, we provide a thorough overview of all the technologies
and techniques that are required to facilitate efficient, scalable and effective energy
analytics AI services. Furthermore, we propose a high-level architecture, indicating
and analysing all the proposed technologies selected for each functionality along
with some alternative technologies that could be used instead.
The Big Data Value Chain for the Provision of AI-Enabled Energy … 33
2 Background
Value chains have been employed as decision support tools by organizations in order
to model the chain of activities they have to perform to deliver a product or service to
the market [23, 24]. The purpose of a value chain to is to categorize the activities of an
organization for them to be better performed and optimized. The value chain consists
of a chain of systems where each one receives an input, performs an operation and
produces an output. The term value chain is used to describe the steps needed in
order to produce valuable information from the collected data. With the exponential
growth of the collected data from sensors and smart meters on the energy sector the
data value chain is transformed to big data value chain which describes the processes
needed to efficiently store, process and retrieve valuable statistics from big data of
the energy sector. The big data value chain is depicted in Fig. 1.
Data acquisition describes the process of collecting, processing and storing data in
a data warehouse. Data acquisition has become one of the most challenging tasks on
collected data on the energy sector due to their huge volume and requires low latency,
high throughput and execution of queries on distributed environments. To achieve its
purpose of retrieving data from distributed sources and storing them effectively in
data warehouses many protocols are used for continuous flow of data like AMQP and
MQTT [25]. Such protocols can be used across different industries, their messages
can be easily encrypted, communication can be achieved through different protocols
(TCP, UDP), and further ease manageability and applicability on various systems
and hardware devices.
Data pre-processing describes the process of cleaning and homogenising data
from various data sources. Collected data from an increasing number of sources in
huge volumes is not going to be clean and will contain noise and anomalies that
will affect the quality of future steps in the value chain and especially data analysis.
Data pre-processing involves a set of steps [26, 27]. Specifically, filtration describes
the process of treating corrupted data; extraction describes the process of trans-
forming incompatible data to proper format; transformation describes the process
of adapting data scaling attributes to improve their quality; validation describes the
process of managing semantic structures and removing invalid data; cleaning refers to
the process of processing inaccurate data; fusion refers to the process of merging data
III. Judicial
Samnites in 315.[1501]
In reviewing the cases said to have been brought by tribunes
before the comitia centuriata it is surprising to find the period from
the institution of the office to the trial of Q. Fabius, 390, swarming
with such prosecutions, whereas for the century intervening between
that date and the Hortensian legislation comparatively few cases are
recorded and those of little significance.[1502] These circumstances
tend to prove that the cases assigned to the earlier and less known
period either belong mostly to the jurisdiction of the duumviri or of
the quaestors rather than of the tribunes, or are in great part
mythical, and that the tribunes, therefore, exercised no extensive
capital jurisdiction before the enactment of the Hortensian law.[1503]
We are led thence to the conclusion that either by an article of the
statute of Hortensius or at least as a recognized consequence of the
high place in the government assured the tribunes by it, the
jurisdiction of these magistrates in political cases was freed from
every restraint. At this time they succeeded wholly to the place of the
duumviri. The cases of which the tribunes had cognizance were
thereafter exclusively political, whereas the questorian jurisdiction
was confined to murder and other common crimes. This distinction
was not a limitation upon the power of the tribunes, who if they
chose might have superseded the quaestors as easily as they had
superseded the duumviri. It was rather a division of functions
adopted by the tribunes themselves in view of their own political
character and on the basis of the relative dignity of the two offices.
The chief judicial function of the tribunes, accordingly, was to hold
officials responsible for their administration, though occasionally they
called private persons to account for their conduct as citizens. All
grades of officials were within their jurisdiction, but most of the cases
were against the higher magistrates.
The first tribunician case of the kind after the Hortensian
legislation, and the first which is absolutely free from historical doubt,
is that brought against P. Claudius Pulcher on the ground that as
consul, 249, he fought the naval battle off Drepana contrary to
auspices, thereby losing his fleet. After the comitia had been
interrupted by a storm, the intercession of colleagues against the
resumption of the trial saved him from the death penalty. As the
result of a new trial before the tribes, however, he was fined 120,000
asses, 1000 for each ship lost.[1504] His colleague, L. Junius, by
suicide escaped condemnation on a charge of perduellio.[1505] In
212 two tribunes of the plebs prosecuted M. Postumius Pyrgensis, a
publican, before the tribes for fraud, setting the penalty at 200,000
asses; but the accused with his friends violently broke up the
assembly, whereupon the tribunes, dropping the original charge,
prosecuted him for perduellio,[1506] we should suppose before the
centuries.[1507] Among the complaints urged against him by the
consuls in the senate were that “he had wrested from the Roman
people the right of suffrage, had broken up a concilium plebis, had
reduced the tribunes to the rank of private persons, had marshalled
an army against the Roman people, seized a position, and cut the
tribunes off from the plebs, and had prevented the tribes from being
called to vote.” Specifically the crime must have been perduellio.
[1508] Before the day of trial he withdrew into exile. In his absence
the plebs on the motion of Sp. and L. Carvilius decreed that he was
legally in banishment, that his property should be confiscated, and
that he should be interdicted from fire and water. In this connection it
should be noticed that whereas the banishment of a citizen by lex or
iudicium was the exclusive right of the centuries,[1509] the tribes
were competent to decree him an exile after his voluntary retirement.
[1510] Some of the coadjutors in the violence of the publican above
mentioned left their bail and followed him into exile; others were
imprisoned to await capital trial, with what result the historian does
not inform us.[1511]
In the same year Cn. Fulvius, a praetor, met with military reverses
through gross cowardice,[1512] and in the following was prosecuted
in a finable action by a tribune of the plebs for having corrupted his
army by the example of his unsoldierly habits. Finding in the course
of the trial that the fault of the magistrate was far more serious than
had been imagined, and that the people were in a temper to vote the
extreme penalty, the prosecutor changed the form of accusation to
perduellio on the ground that such cowardly conduct in a
commander threatened the existence of the state. In this instance,
too, the accused avoided trial by withdrawing into exile.[1513] In 204
by a decree of the senate a special commission, consisting of the
praetor for Sicily with a council of ten senators,[1514] was appointed
for the trial of a legate of Scipio, Q. Pleminius, on the charge that he
had robbed the temple of Persephone in Locri and had violently
oppressed the Locrians.[1515] The commission brought him and his
accomplices in chains to Rome and cast them in prison to await their
trial for life before the centuries.[1516] The day of trial was continually
deferred, till finally Pleminius, now charged with the instigation of a
plot to burn the city, was put to death in prison.[1517] The fate of his
accomplices is unknown.[1518] Livy[1519] remarks that while
Pleminius was languishing in jail the wrath of the populace gradually
changed to sympathy, to such an extent doubtless as to convince the
authorities of their inability to secure a popular verdict in favor of the
death penalty. In fact since the death of M. Manlius Capitolinus, 384,
no example of the execution of a death sentence pronounced by the
assembly is recorded in history.[1520] But the magistrate probably
often inflicted corporal punishment in violation of the third Valerian
law. To put an end to this abuse, and at the same time to embody in
legal form the popular feeling against the application of the death
penalty to citizens, a Porcian law absolutely forbade the scourging or
slaying of a citizen under the imperium domi, the article prohibiting
the sentence of death being afterward reënforced by other
enactments.[1521] There has been much discussion as to the
authorship of this law; probably it was the work of M. Porcius Cato
the Elder in his praetorship, 198.[1522] Another Porcian law, probably
of P. Porcius Laeca, praetor in 195, extended the right of appeal to
Roman citizens who were engaged in the affairs of peace outside
the city, in Italy and the provinces, and were therefore under the
military imperium.[1523] According to this law the citizen who
appealed was sent to Rome for trial by the appropriate civil
authorities. Still later the third Porcian law, which Lange[1524]
conjecturally assigns to L. Porcius Licinus, consul in the year of the
elder Cato’s censorship, 184, seems to have been passed for the
benefit of Roman soldiers. We learn from Polybius,[1525] who wrote
later than the date last mentioned, that the military tribunes were
accustomed in court-martial to condemn common soldiers for
neglect of sentinel duty and that the condemned were cudgeled and
stoned, often to death, by their fellow-soldiers. He also speaks of the
punishment of entire maniples by decimation. Under Scipio
Aemilianus, 133, the Roman who neglected duty was flogged with
vine stocks, the foreigner with cudgels.[1526] Cicero[1527] intimates
that in his own time there was no appeal from the judgment of
commanders; and in fact it is impossible to understand how
discipline could otherwise be maintained. Evidence to the contrary is
scant and uncertain. The person against whom an accusation of
desertion was brought before the tribunes of the plebs in 138 seems
to have claimed to be a civilian, and on that ground appealed to the
tribunes. When proved guilty he was flogged and sold as a slave,
probably by a judgment of the military authorities.[1528] In 122 Livius
Drusus proposed to exempt Latin soldiers from flogging.[1529] While
informing us that in 108 a commander had a right to scourge and put
to death a Latin official, Sallust[1530] intimates that he had less
authority over a Roman. In the time of the emperors, on the other
hand, soldiers were subject to the death penalty as in the time of
Polybius.[1531] All these circumstances may be best explained by
supposing that the third Porcian law permitted the infliction of
flogging and death on Roman soldiers by the judgment only of a
court-martial.[1532] This difficult subject is further complicated by the
statement of Cicero[1533] that the three Porcian statutes introduced
nothing new excepting by way of penalty. Interpreted in the light of
other information given by various authors, including Cicero himself,
these statutes simply extended the right of appeal by adapting the
Valerian principle to new conditions, and substituted exile in place of
scourging and death. In the relation between the accused and the
civil court the cry “civis Romanus sum” was thereafter a sufficient
protection from bodily injury.[1534]
In the period to which the Porcian laws belong falls the accusation
of perduellio brought by the tribune P. Rutilius Rufus against the
censors C. Claudius and Ti. Sempronius Gracchus, while they were
in office, 169. The charge against Gracchus was disregard of the
tribunician auxilium, against his colleague the interruption of a
concilium plebis (quod contionem ab se avocasset). The accused,
foregoing the privilege of their magistracy, consented to a trial, which
came before the comitia centuriata. Claudius narrowly escaped
condemnation, whereupon the case against Gracchus was dropped.
[1535]