Mobile Apps Meet The Smart Energy Grid: A Survey On Consumer Engagement and Machine Learning Applications

Received November 16, 2020, accepted November 28, 2020, date of publication December 7, 2020,
date of current version December 17, 2020.

Digital Object Identifier 10.1109/ACCESS.2020.3042758
Mobile Apps Meet the Smart Energy Grid:

A Survey on Consumer Engagement and
Machine Learning Applications
SPIROS CHADOULOS , IORDANIS KOUTSOPOULOS , (Senior Member, IEEE),
AND GEORGE C. POLYZOS , (Member, IEEE)
Department of Informatics, School of Information Sciences and Technology, Athens University of Economics and Business, 104 34 Athens, Greece
Corresponding author: Spiros Chadoulos (spiroscha@aueb.gr)
This work was supported by project InterConnect (Interoperable Solutions Connecting Smart Homes, Buildings and Grids), which has
received funding from European Union’s Horizon 2020 Research and Innovation Programme under Grant 857237.
ABSTRACT Consumers lie at the epicenter of smart grids, since their activities account for a large portion
of the total energy demand. Therefore, utility companies, governmental agencies, and various other entities
with environmental concerns aim at lowering and shaping energy consumption patterns to achieve peak
load reduction, load smoothing, and hence carbon emission curtailment. In this survey paper, we present
an overview of approaches for engaging smart grid consumers and for providing them with information,
motivation, and recommendations for energy efficiency through mobile apps. Our focus is to bring machine
learning approaches closer to smart grid mobile apps so as to optimally manage consumer flexibility and
enhance energy savings through detailed consumer profiling and modeling, since an increasing amount
of energy consumer data is becoming available. A novel survey and analysis of prior work in the area is
conducted in order to identify gaps from this perspective. We consider both recent research project outputs
and commercial products and we discuss various aspects of the designs, such as state-of-the-art technologies,
extrinsic and intrinsic motivation techniques, gamification, consumer profiling, and the role of machine
learning and recommender systems in this context. Furthermore, different mobile apps are presented and
compared based on the most important features that affect consumer energy efficiency and sustainability,
such as data visualization, gamification, flexibility, consumer profiling methods, feedback mechanisms,
recommendations, social media, and machine learning integration. The main goal of this work is to identify
how mobile apps incorporate these features to engage energy consumers in energy-efficient behavior, assess
the current state-of-the-art in the area, and highlight future research directions.
INDEX TERMS Mobile apps, smart grid, consumer motivation, energy consumption reduction, smart
homes, smart buildings, demand shaping, machine learning, recommender systems.
I. INTRODUCTION operation, as well as new and innovative services and prod-

The smart grid ecosystem is growing rapidly during the last ucts for customers [1]. Consumers comprise the main actors
decades, contributing to the efforts of environmental green directly involved in the electricity grid [2], while buildings
transition, energy efficiency, and reduction of carbon levels. are responsible for about one third of the global energy con-
At the heart of a smart grid lie Internet of Things (IoT) sumption [3], with residential buildings accounting for 22%
devices providing real-time monitoring and management of of the total energy demand worldwide [4]. Hence making
each network node, through two-way communication chan- end-users more energy-aware and motivating them to lower
nels between utilities and users. This approach is called electricity consumption can have an immense impact on
Advanced Metering Infrastructure (AMI) and supports self- energy efficiency and carbon emissions reduction. In smart
monitoring, self-healing in the case of errors, and automated grids, an approach called Demand-Side Management (DSM)
is adopted by utilities to dynamically adjust energy consump-
The associate editor coordinating the review of this manuscript and tion. DSM models include Demand Response (DR) programs
approving it for publication was Khursheed Aurangzeb . that aim at changing the power consumption profiles, and
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
219632 VOLUME 8, 2020
S. Chadoulos et al.: Mobile Apps Meet the Smart Energy Grid: A Survey on Consumer Engagement and Machine Learning Applications
shaping the load curves of occupants with dynamic pricing engagement and acceptance in the smart grid context is car-
schemes and incentives. Thus, it is essential for utilities to ried out in [2]. That work is essentially a guide for researchers
engage and educate consumers on energy efficiency, while and policy makers to help them understand the behavioral
also understanding their behavior [2]. characteristics of a smart grid customer. The study states that
This is where mobile apps, paired with smart home and the reduction of electricity bills and carbon footprint, as well
IoT infrastructures come into play. Mobile devices have more as the increase of user comfort, are the main factors that
computational capabilities than ever, while mobile app design affect consumers’ perspectives towards energy management
and development advancements lead to their adoption by programs. Furthermore, user acceptance can be increased by
the majority of the population. Hence, the penetration of ease of use, education regarding the benefits of smart grids,
smartphones together with the proliferation of mobile apps and implementation of pilots.
create effective means to interface with smart grid con- As far as smart grid mobile apps for residential energy effi-
sumers. Furthermore, the rapid progress of artificial intelli- ciency are concerned, the authors in [10] carry out a review
gence and machine learning can significantly help utilities of how mobile apps and especially social media mobile
to better understand the behavioral characteristics of their apps can further engage smart grid consumers. They discuss
consumers, and thus design effective incentive mechanisms how electricity providers and utilities can take advantage of
through mobile applications and advanced electricity con- mobile applications and social media to achieve customer
sumer profiling. loyalty and engagement. The authors recommend that this
Many survey papers related to smart grids, smart homes, can be achieved through personalization, exclusivity, and
and the IoT have been published in the past. For example, gamification.
in [5] the authors conduct a thorough review of IoT applica- In this survey, we present a review of existing work on
tions regarding smart homes. That survey covers a large area, mobile apps for the smart grid ecosystem and residential
from other review articles and studies examining smart home energy efficiency. Particularly, we conduct a novel survey and
applications of IoT, to proposals and attempts of developing analysis of those mobile app features that affect consumer
such applications. The study briefly touches on the usage of engagement and energy savings, and then further analyze
some mobile applications in smart homes for energy conser- the impact that machine learning approaches can have on
vation, primarily through automation and actuation. such mobile apps through optimal flexibility management,
Furthermore, some survey papers regarding the utilization detailed consumer profiling, and targeted personalized rec-
of gamification in smart grids have also been published. ommendations. Our survey helps to assess the current state-
A review of gamification and serious games for residential of-the-art in smart grid mobile apps, focusing on consumer
energy consumption reduction is carried out in [6]. That motivation, engagement, and the use of machine learning
study includes a variety of gamification applications, and and recommender systems. This class of mobile apps has
states that most of them incorporate features like social not been explored much, while the repercussions of machine
media integration, point systems, challenges, leaderboards learning and recommender systems and their impact on smart
and rewards. One other review of gamification and seri- grid consumers is a new emerging research direction. Hence,
ous game approaches aimed at engaging consumers towards we focus on electricity consumer engagement and motiva-
energy efficient behaviors is undertaken in [7]. The study tion, both from the research and the industry perspectives,
shows that such strategies are quite effective and are mostly while also examining how machine learning can be utilized
utilized for environmental education, consumption awareness to improve the impact and the results of such mobile apps,
and pro-environmental behaviors. In [8] a survey of Euro- through techniques such as detailed consumer profiling, clus-
pean research projects is presented, which delivers gamifica- tering, and classification. Literature surveys that discuss var-
tion and visualization solutions to enhance energy-efficient ious applications of machine learning in IoT and smart grids,
behavior of energy and water consumers. The authors derive e.g., [11], [12], and [13], have been conducted in the past.
and propose solutions for frequent challenges related to such However, our work differs from existing ones since we focus
projects, such as data acquisition, applying baselines, ensur- on smart grid consumer engagement and motivation through
ing consumer engagement, producing various motivation fac- mobile apps and behavioral traits recognition with machine
tors, user modeling, providing user guidance and education, learning.
privacy issues, model validation, achieving long-term impact, The remainder of this paper is organized as follows.
and averting rebound effects. In section II, we present the typical system architecture
Moreover, several papers have surveyed the area of smart of a smart grid mobile app, with state-of-the-art technolo-
grid consumer engagement and behavioral characteristics. gies and frameworks. In section III, a brief overview of
A literature survey of Information and Communications user engagement and motivation techniques with a focus
Technology (ICT) applications regarding home energy effi- on gamification is conducted. In section IV, existing work
ciency is conducted in [9]. The authors focus on occupant on mobile apps for energy efficiency is presented, regard-
behavioral aspects and show that the adoption of such ICT ing both research project outputs and commercial products.
solutions can reduce household consumption by up to 5%. In section V, the role and impact of machine learning and
A review of the factors that enhance electricity consumer recommender systems in such applications is studied, and
VOLUME 8, 2020 219633

FIGURE 1. Smart grid system architecture example consisting of IoT devices, an IoT gateway for communication, a back-end
platform for data processing and web services, and a mobile app frontend, all with state-of-the-art technologies.
relevant approaches are presented. In section VI, a thor- forecasting can lead to a substantial reduction of carbon
ough comparison of existing mobile apps for electricity con- emissions. For example, excessive generation can be stored
sumer motivation is carried out, regarding various features. with batteries, while residents can shift scheduled activities
In section VII, we discuss a series of open issues and future (e.g. a washing machine task) to midday time slots with high
directions, and in section VIII we present a summary of the solar generation capabilities.
paper, key points, and our conclusions. These sensors and smart meters usually communicate with
an IoT gateway, and transfer the aforementioned measure-
II. SYSTEM ARCHITECTURE AND RELEVANT ments through wireless short-range network protocols like
TECHNOLOGIES Bluetooth [14], Wi-Fi [15], Zigbee [16], or Z-Wave [17].
A feature-rich smart grid mobile app is most often utilized Zigbee is a personal area wireless communication proto-
along with other hardware and software components, like col providing low data rates with low power consumption,
IoT devices, smart meters, back-end servers, and Cloud in a simpler and cheaper way, compared to other wireless
databases. A typical example of the system architecture of protocols like Bluetooth and Wi-Fi. Z-Wave is a wireless
a smart grid mobile app, using state-of-the-art technologies communication protocol specially designed for smart home
and frameworks, is presented in Fig. 1, and discussed in this automation and low-power communication between appli-
section. ances. The IoT gateway is typically a hardware component
installed at the user’s premises and is essentially responsible
A. IoT INFRASTRUCTURE for the communication between the IoT devices, and the
Starting from the users’ premises, IoT sensors and smart back-end platform in a secure way. It uses protocols like
meters are deployed at houses and buildings to collect mea- MQTT [18] to communicate and exchange data with the
surement data. Such measurements include power, energy, back-end system in a publish-subscribe approach. This means
current, and voltage both at whole dwelling and individual that the IoT devices and the gateway act as data producers/
appliance level, as well as climate conditions (e.g. temper- publishers, i.e. publishing to different topics, while the
ature, humidity, pressure, etc.), illuminance, and detected back-end server acts as a data consumer by subscribing to
motion. Furthermore, IoT devices are utilized to measure specific topics. All communication is handled by a broker,
the performance and energy generation output of renewable i.e. when a new measurement is available by a publisher,
sources, which most often are solar panels. Solar generation it sends the appropriate message to inform the broker, which
monitoring is essential for smart homes and smart energy then distributes the data to the subscribers of the related
grids since residential energy production modeling and topic.
219634 VOLUME 8, 2020

B. BACK-END SERVER Transfer, defines a set of web service standards and guide-
The back-end system contains multiple software components lines for computer system communication through the inter-
responsible for different services related to the mobile appli- net with high interoperability.
cation, such as database storage, big data processing, data
analytics, machine learning computation, and REST APIs. III. CONSUMER MOTIVATION TECHNIQUES
Different consumer interfaces are used by mobile app design-
1) DATABASE ers to elicit certain behavior and motivate consumers towards
Smart grid IoT measurements are often stored in NoSQL specific actions. For instance, prior work has been conducted
databases like MongoDB [19] or InfluxDB [20], since they for the optimization of mobile crowdsensing campaigns,
are a better fit for time-series data. Such NoSQL databases where service providers have to choose the subset of targeted
are most often utilized for real-time applications with big users, the assigned tasks and incentives to fulfill them [27].
data, since they store data in different ways compared to tradi- Such motivation techniques most often include monetary
tional relational databases, like key-value stores, documents payments or in-app rewards, and their efficient allocation
or objects, which can provide immense speed and memory requires accurate user profiling in terms of behavioral char-
benefits depending on the application. MongoDB is a NoSQL acteristics and preferences. Furthermore, machine learning
database that uses documents to store data instead of tables, techniques can also be used to capture consumer profile
while InfluxDB is also a NoSQL database specialized in time information from historical data, e.g. past responses to similar
series data like IoT measurements. recommendations from the mobile app, in order to make a
more efficient task allocation between users in terms of costs
2) BIG DATA PROCESSING AND MACHINE LEARNING and impact [28]. In addition, such crowdsensing approaches
Issues regarding volume, velocity, and variety might arise can be generalized and applied to mobile applications for
when processing smart grid data for analytics purposes. Thus, energy efficiency, where energy-related tips and recommen-
technologies like HDFS [21], Kafka [22], and Spark [23] can dations are presented to the users to reduce their energy
be utilized for big data processing and analytics. HDFS is consumption [27].
a distributed file-system capable of running on commodity Two types of motivation can be found in the literature:
machines and is ideal for storing large datasets, such as extrinsic and intrinsic motivation. The former is based on
smart meter measurement data used for analytics. Kafka is external rewards such as money and some type of recognition,
a publish-subscribe stream-processing framework designed while the latter refers to user behavior without clear external
for real-time big data streaming and messaging that pro- rewards that is triggered due to personal beliefs and needs.
vides connectors with other interfaces for big data storage
and analytics. Kafka is used in smart grid data pipelines in A. MEANS FOR EXTRINSIC MOTIVATION
order to handle the large volume and frequency in which One of the most commonly used techniques for extrinsic
smart meters can generate and send measurements in real consumer motivation is gamification, which refers to tech-
time, without having data loss, latency, and scalability issues. niques and systems that focus on motivating and engaging
Spark is a big data processing framework optimized to run on users by utilizing game design principles [29]. Common gam-
computer clusters and includes a variety of functionalities, ification features like points, badges, and leaderboards, can
from classic SQL-like data queries to streaming analytics substantially increase user engagement and motivation. Con-
with Spark Streaming and graph-processing with GraphX. sumer motivation can also be achieved through competition.
Furthermore, Spark comes with a machine learning library Namely, gamification creates competition among consumers,
for big data called MLlib, which includes popular machine and they compete to gain badges, points, etc. Points can be
learning algorithms. Spark can be used to execute heavy used to determine user achievements and progress, while they
smart grid data analytics jobs with high speed, both for batch are most often converted to real rewards, like discounts and
and streaming measurement data from IoT devices and smart gift items. Badges act as virtual status symbols that separate
meters. Frameworks like TensorFlow [24], Keras [25], and the users that acquired them from the rest. Leaderboards serve
PyTorch [26] can also be utilized for deep learning models as a comparison tool for users, engaging them in a competitive
and neural networks. Some of the trained models can also way.
be deployed and run directly on the mobile device using Gamification has been shown to have a positive impact on
libraries like TensorFlow Lite. We will further discuss the role residential energy consumption reduction [6]. In [29], a the-
of machine learning for smart grid mobile apps in section V. oretical gamification framework for household consumer
energy efficiency engagement is proposed. This framework
3) WEB SERVER can be applied in any setting assuming smart meters are
Finally, a back-end web server is responsible for providing present. It first establishes a behavioral change model for
all the REST API endpoints necessary for the mobile app to energy efficiency, self-consumption, and demand response,
communicate with the aforementioned services and exchange with specific requirements like awareness, knowledge, pri-
data. REST, which stands for Representational State vacy, incentives (economic, environmental, and social), ease
VOLUME 8, 2020 219635

of use, and self-control. The gamification component design IV. MOBILE APPS FOR ENERGY EFFICIENCY
of [29] includes: In this section, we present existing smart grid mobile appli-
• Information provision: Statistics, messages, tips. cations that provide users with motivation and incentives for
• A Reward system: Discounts, virtual currency, prizes, energy efficiency and consumption reduction. Both research
coupons, etc. project outputs and commercial products are discussed, while
• Social networking: Competitions, user collaboration, most of them use motivation mechanisms like visualization,
energy communities. gamification, and simple notification tips. Some of the apps,
• A friendly user interface: Dashboards, leaderboards, utilize user profiling techniques to better understand con-
progress bars, notifications/messages, degree of control sumer behavior and provide customized services. Machine
for automation/actuation. learning models and recommender system approaches are
• User performance status: Points, badges, levels. also integrated in some of them. Some apps use feedback
Other gamification frameworks have also been developed, mechanisms like questionnaires, while most of the research
e.g. [30] and [31], aimed at aspects related to privacy and apps integrate social media. A thorough comparison of vari-
supplier-consumer interaction respectively. ous app features is conducted in section VI and is presented
In general, individual user participation and actions cannot in Table 1.
be taken for granted, and motivation models should always
take into account uncertainty regarding user behavior [32].
A. RESEARCH PROJECTS
Nonetheless, unpredictability should also be considered for
In this subsection, we present prior work on the development
the incentives themselves. Extrinsic motivation is a quite
of mobile apps for residential energy efficiency and smart
effective way to derive certain behavior and actions from
grid consumer engagement, regarding academic and collab-
users. However, such rewards might lose their engagement
orative research project results. In recent years, researchers
impact after some time, due to the fact that users anticipate
have developed numerous mobile apps targeted at different
them. When such expected rewards incorporate uncertainty,
aspects of residential energy efficiency. These apps share
users are more likely to use the app on a daily basis, due to
many common features, such as consumption visualization
their curiosity.
and forecast, and despite their differences, they all have the
Despite the fact that unpredictable and uncertain rewards
same goal of reducing energy waste.
are an effective strategy to motivate users, the overjustifica-
tion effect should be taken into account. This effect arises
when the app offers extrinsic rewards to users for an activity 1) ACADEMIC RESEARCH PROJECTS
that they were anyway going to engage in. Consequently, any One of the first attempts to develop a mobile application to
prior intrinsic motivation technique that encouraged the user encourage consumer energy-efficient behavior was the Power
in the first place, is weakened. Explorer game [34]. It was also one of the first mobile games
that incorporated real household consumption data through a
B. MEANS FOR INTRINSIC MOTIVATION game server connection. The researchers validated the energy
Intrinsic user motivation can be achieved with methods like efficiency impact of the app through interviews with users.
showing progress, simple task assignment, sending success The EnergyLife [35] mobile game was developed to
messages, and presenting environmental impact. These meth- enhance the energy awareness of consumers and to reduce
ods are explained below: household energy consumption. The app included a feed-
• Progress visualization can increase the probability back system presenting historical and real-time smart meter
of completion for a time-consuming activity. When consumption to the users up to appliance level. The gam-
realizing how far they have come, users feel a ification component of the EnergyLife app incorporated
sense of accomplishment that reduces the drop out points, quizzes, levels, contextualized personal tips, and
rates. recommendations.
• Task simplification is important to achieve app sus- In [36], researchers examined consumer incentive mech-
tainability and continuous engagement. A task that is anisms in the smart grid, through a system called Ener-
too difficult to complete might force even a highly gyCoupon. This included a mobile application that issues
motivated user to abandon it. To avoid this, multiple coupons to consumers in return for demand response actions.
methods of completing a task could be provided to the The users could then take part in lotteries and win prizes
users. with these coupons. The EnergyCoupon mobile app was
• Success messages give users a sense of achievement, also responsible for providing personalized tips, historic
while task completion can also be confirmed. and real-time consumption visualization to users, as well as
• Environmental impact presentation informs users with consumption goals/targets for the current day. The system
statistics and reports on how their activities affected the also utilized machine learning algorithms to conduct elec-
environment and other impact that their saved energy tricity price and baseline consumption (i.e. without demand
had, motivating them to engage with the app even more response) forecasting, in order to help the user motivation
in order to increase their contribution. process.
219636 VOLUME 8, 2020

TABLE 1. Smart grid mobile app features comparison.
A mobile application called EnergyWiz was developed by performance, the app shows a message of social
researchers to study the effects of comparative and commu- approval or disapproval.
nity based feedback on household energy efficiency [37], • One-on-one consumption challenges through Facebook.
[38]. The app included the following features: • Energy efficiency rankings of similar users, in terms of
• Live data visualization. characteristics like house size and type.
• Temporal self-comparison, i.e. through historical data. The MyEarth [39] app developed at the University of
• Comparison with neighbors, i.e. the user is com- Wisconsin-Madison, allows users to track their environmen-
pared with the average performance of two groups of tal impact using five categories: electricity, energy usage,
neighbors: efficient and inefficient. Depending on user recycling, travel, and food. MyEarth manages to boost user
VOLUME 8, 2020 219637

FIGURE 2. PEAKapp dashboard screens [33].
engagement utilizing animations like a polar bear on a shrink- • Comparison of each user’s consumption to the average
ing iceberg. When users take more energy-saving decisions, baseline.
the iceberg gets bigger and the polar bear becomes happier. • Ability to share progress and achievements through
social media.
2) COLLABORATIVE EUROPEAN RESEARCH PROJECTS Furthermore, the mobile app included a serious game where
The ecoGator project [40], funded by the European Union’s players were invited to estimate their future consumption as
IEE (Intelligent Energy Europe) Programme, developed a precisely as possible, by placing a bet. They gained more
smartphone application aimed at energy efficiency through points as their bets got lower and/or closer to the actual con-
gamification. The ecoGator app helped users to select the sumption, and lost all of their points if the actual consumption
most energy-efficient products in the market, and to embrace exceeded their bets. If their bets were higher than the actual
an energy-efficient lifestyle through energy-saving tips, chal- household consumption, the app prompted them to turn on
lenges, quizzes, points, competition with other users, leader- specific appliances and thus adding the exact missing amount
boards, social media integration, and real-world rewards. of kilowatt-hours.
The PEAKapp project [41] (Horizon 2020, 2016-2019) The ChArGED project [44], [45] (Horizon 2020,
focused on developing a fun app that encourages users to 2016-2019) included a gamification based mobile applica-
behave more energy efficiently and to make use of the most tion, targeted at stimulating consumer energy savings and
appropriate load shifts. The core of the PEAKapp project ensuring balanced load. The ChArGED app engaged users
was a dynamic pricing system responsible for reducing con- towards education of efficient use of resources, by utiliz-
sumption during peak demand hours. The effects of various ing challenges based on NFC swipes, e.g. a ‘‘Windows’’
electricity tariff schemes on different household types based challenge, where users were prompted to close the windows
on load profiles and socio-economic characteristics were also when the A/C is on, and then swipe the relevant NFC tag
investigated [42]. One of its primary goals was to connect to complete the challenge. The app was also supported
users with each other and facilitate their interaction with by social networking technologies, that allowed people to
the energy provider while providing consumer access to the team up and participate in competitions through the integra-
electricity spot market. The key features of the PEAKapp tion of leaderboards and badges (Fig. 3). The system was
mobile application included: designed to provide personalized motivation for participating
• Presentation of the main consumption indicators, meter players based on their progression in consuming resources
readings, and monthly balance sheet (Fig. 2). efficiently, e.g. energy saving tips. An hourly solar energy
• Environmental analysis, saving tips, and discount offers. generation forecasting module was included, and the users
• Comparison between renewable production and were prompted to utilize appliances during peaks of solar
consumption. production. The end-users were able to access their energy
• Forecast of household consumption, costs, and carbon profiles and hence monitor aspects like their consumption,
emissions. energy-saving devices and achievements.
219638 VOLUME 8, 2020

FIGURE 3. ChArGED mobile app screenshots [43].
The GAIA project [46] (Horizon 2020, 2016-2019) • Occupant awareness through tips for more efficient use
produced a gamified online learning platform on energy of systems like HVAC, lights, and elevators.
efficiency where students play, explore and spread their The GreenSoul mobile app prompted users to fill an anony-
learnings, and thus impact on energy usage at school. The mous socio-economic questionnaire when they interacted for
GAIA ecosystem included features like a building man- the first time. The mobile app had two versions, a social
ager application that provides access to all IoT data pro- recognition version and a collective treatment version. The
duced by the GAIA infrastructure (consumption, temperature, former rewarded users with points for reducing the consump-
humidity, luminosity, noise levels and presence/movement) tion of their personal computer, while the latter determined
and a recommendations module. Furthermore, data analyt- the cumulative points of the whole office, visualized the total
ics modules were included to provide end-users with use- office consumption, points and carbon emissions saved, and
ful insights like consumption anomaly detection, building rewarded the office with collective points. Also, the collective
clustering and statistics. The gamification aspect of the treatment version compared their rank in terms of points with
GAIA project incorporated: other office colleagues.
• Team missions and competitions.
The enCOMPASS project [48]–[50] (Horizon 2020,
• Knowledge missions, i.e. knowledge quizzes related to
2016-2019) integrated visualization of energy data col-
energy efficiency.
lected from smart sensors, user-generated information,
• Action missions, i.e. playful sparking of student action
context-aware collaborative recommendations for energy
in the school building aiming to increase energy
saving, intelligent control, and adaptive incentives. It included
efficiency.
a mobile and web application for:
• Raising awareness and playfully sustaining students’ as
well as the wider community’s interest, and engagement • Visualizing consumption impact (i.e. monetary, environ-
with project activities, through weekly GAIA Scavenger mental, hedonic) and setting personal goals.
Hunt hashtag games implemented in popular social • Playing social games and pursuing rewards, badges,
media. achievements, etc.
Moreover, GAIA aimed at achieving sustainability in terms of • Receiving consumption tips and customized recommen-
user energy-efficient behavior, through short and long term dations, based on user clustering (Fig 4).
sustainability plans, and development of more gamification • Predicting the presence of users in indoor environments
content, educational material, and new features. using machine learning, based on environmental fea-
The GreenSoul project [47] (Horizon 2020, 2016-2019) tures and energy consumption variables [51].
provided behavioral adjustment recommendations to office • Enabling energy disaggregation regarding specific
building occupants, using smart meter data. The Green- appliances [52].
Soul system included mobile and web applications that • Detecting the operation state of each appliance using
incorporate: machine learning, in order to achieve user-activity
• Device smart monitoring and control. recognition [53].
• Manual control affected by user behavior change. • Optimizing user comfort.
VOLUME 8, 2020 219639

achieving certain goals. The app provides energy consump-

tion reports, game statistics like progress and points, and
energy efficiency tips for different home appliances. Con-
sumers can be either collaborative, where users living in the
same city try to collectively reduce the city’s total consump-
tion and reach a specific goal, or competitive, where users try
to reduce the city’s total consumption compared to another
team. User surveys were conducted before and after adopting
the app and the results showed a significant improvement in
sustainable consumption behavior.
The Tribe project [63] (Horizon 2020, 2015-2018), focused
on developing a mobile game to enhance the energy effi-
ciency of public buildings. Building occupants and visitors
FIGURE 4. enCOMPASS mobile app tips and recommendations [56]. could play a campaign game where they try to balance money
and comfort by interacting with a virtual home, and thus learn
how to be more energy aware.
The enCOMPASS ecosystem incentivized users towards
The GreenPlay project [64] (Horizon 2020, 2015-2018)
energy efficiency through gamification, by developing a
included a serious game accessible from multiple platforms
hybrid game called FUNERGY consisting of a physical card
that allowed users to interact with their own virtual island
game and a mobile game [54]. The players could scan QR
ecosystem. When users reduced their real house energy
codes located on the cards and expand the game rules or just
consumption, they gained in-game rewards. The GreenPlay
use the mobile app to recognize objects around them using
system also included a platform that monitored household
computer vision and answer energy-related questions about
consumption and provided visualization, tips, challenges,
them [55].
and team consumption reduction contests to the users. All
The EnerGAware project [57] (Horizon 2020, 2015-2018)
the features of the GreenPlay system aimed at sustainable
focused on developing a serious mobile game that enhances
energy-efficient behavior.
the energy efficiency of household occupants. The game was
The InterConnect project [65] (Horizon 2020, 2019-2023)
directly linked to occupant energy behavior through smart
is an ongoing project that aims at developing and demonstrat-
meter data. The users could create their own virtual smart
ing an interoperable ecosystem of demand-side flexibility ser-
household and learn how to reduce their energy consumption
vices targeted at energy consumers, through seven large-scale
through in-game interaction with the virtual household. The
pilots. As part of the Greek pilot, a smart grid mobile
achieved savings, of both their real home and their app sim-
app to increase consumer engagement will be developed.
ulated household, contributed to completing missions and
The app will include data visualization, gamification, social
goals.
media integration, machine learning models for advanced
The ENTROPY project [58] (Horizon 2020, 2015-2018)
consumer profiling and behavior prediction, energy-saving
included a mobile application that provided personalized ser-
tips, load shift recommendations, appliance flexibility actu-
vices to building occupants, regarding energy efficiency. This
ation, and user feedback mechanisms.
app allowed users to view real-time data from IoT sensors and
get real-time recommendations to reduce their energy con- B. INDUSTRY EFFORTS
sumption. Furthermore, the ENTROPY app provided tips and Numerous mobile applications for energy efficiency in smart
quizzes aimed at energy awareness, both through the person- grids have also been developed by the industry sector. The
alized app and a treasure hunt mobile game. The ENTROPY main focus of industry efforts in developing smart grid mobile
system utilized machine learning algorithms for user classi- apps is on consumption data visualization and appliance
fication, clustering, and consumption forecasting, in order to level monitoring, as opposed to academic and collaborative
gain valuable behavioral insights, which were crucial for the research project mobile apps, which additionally focus on
rule-based personalized recommendation system [59]. It rec- consumer behavior aspects and gamification.
ommended energy-efficient actions based on rules/conditions Smappee [66] is an energy management system that
in a personalized manner, meaning that the recommendation includes a mobile app for residential consumers. The system
message was stated in a different way depending on the type collects data on solar generation, as well as electricity, gas,
of the user (e.g. philanthropist, socializer, or free spirit). The and water consumption. Smappee’s goal is to provide users
clustering inputs included energy consumption and demo- with all the information needed to monitor home energy
graphic data. consumption and costs. A clip-on sensor is provided to the
The Social Power Project [60]–[62] supported by the users, which can be quickly attached to a fuse or breaker
BREF-Social Innovation programme, aims at engaging con- box. Smappee provides three ways to monitor appliance level
sumers towards energy-efficient behavior through a mobile consumption:
application. The Social Power App includes challenges, • It can be used for submetering, by attaching current
tips and quizzes to incentivize users, who earn points for clamps to the distribution panel.
219640 VOLUME 8, 2020
• It can collect measurements from the power outlet with renewable generation forecasting, and malicious activity
an intelligent smart plug and a single-phase submeter. detection, with great success [74]. In this section, we focus
• It provides submetering from the power cable using on how machine learning can be utilized to make smart grid
Non-Intrusive Load Monitoring (NILM) energy disag- mobile apps more effective in terms of user engagement
gregation. and motivation, and thus achieve higher levels of energy
Furthermore, Smappee visualizes electricity, gas, and water efficiency. Machine learning and recommender systems can
consumption and costs per day, week, month, or year, as well be used for more targeted and personalized user experience
as solar generation. The app can also compare the household’s with the mobile app, while aiming at energy efficiency. Con-
electricity usage with other locations. Moreover, users can sequently, this can be mainly achieved through more accurate
control their devices (actuation) and automate energy flows user profiling, clustering, classification, and recommenda-
from the app, even when they are not at home. tions, while always depending on data collected from house-
The TED (The Energy Detective) [67] ecosystem includes hold smart meters.
a real-time energy monitor mobile application called TED Hence, the goal of such approaches is to extract high-level
Advisor. Hardware is installed at the building breaker panel features and patterns related to occupant behavior from
and measures energy demand, calculates costs, and sends low-level device readings, which can prove to be quite helpful
email or notification alerts to the user. These alerts can for providing users with personalized services and incentives.
include tariff changes, reaching limits and thresholds, exces- These features and characteristics are utilized to construct an
sive load events, high average demand for the location, and energy profile for each consumer/household, and understand
solar generation reports. App users can also study their their behavior and habits so as to target them for energy
historic energy consumption, and understand how they can efficiency and demand response actions.
improve their performance in terms of energy efficiency. An electricity consumer profile is essentially a vector of
The Neurio Home Energy Monitor app [68] (by Generac) attributes related to the household’s energy consumption,
coupled with the Neurio Sensor, which is installed in the generation, socio-demographic characteristics, and other
home’s breaker box, monitors household appliance consump- available features that can provide insights regarding energy
tion and generation in real-time. The app also notifies users behavior. An electricity consumer profile example is depicted
when they surpass their budget, while also comparing them in Fig. 5, and consists of attributes like household typical
to their neighbors in terms of energy consumption. load profile (i.e. average power consumption of each time
JouleBug [69] gives its users the ability to monitor their window during the day), consumption aggregates (i.e mean,
everyday habits and earn pins, badges, and points by carrying maximum, minimum consumption during different periods),
out energy-efficient actions. Moreover, users can sign up for consumption ratios (i.e. features calculated as the ratio of two
local and national challenges to compete with other users consumption aggregates, e.g. the average consumption during
from their community. the afternoon and during the evening can indicate when cook-
Constellation Connect [70] (by Constellation, an Exelon ing takes place), temporal features (i.e. the first time an event
company) is a home energy management system that allows happens, e.g. daily consumption going over a threshold), flex-
users to interact with their devices, like smart appliances, ibility factor,1 number of occupants, socio-demographic char-
security cameras, lights, smart locks, and motion detec- acteristics, and heating/cooking type (i.e. electrical or not).
tors. Consequently, this app engages and motivates users to Clustering and classification are two of the most
become more energy-efficient, while also protecting their well-known and studied machine learning techniques. Mobile
houses. applications for residential energy efficiency can benefit
The majority of mobile apps focus on energy consumption immensely by integrating consumer clustering and classifica-
recording, by informing users on how much energy their tion, based on smart meter consumption data and other input
devices are consuming. Some apps educate users on the features, to derive detailed energy behavior profiles for their
amount of money they could save if they reduced their energy users.
consumption, like Energy Cost Calculator [71], Energy Other machine learning models can also be applied in
Tracker [72] and Kill-Ur-Watts [73], which do not come with smart grid mobile apps for consumer profiling. In [75] for
hardware sensors that need to be installed, and users have to example, a linear regression-based algorithm that detects the
insert their data manually. base load of households is proposed, i.e. constant consump-
tion from always-on appliances like refrigerators/freezers.
V. THE ROLE OF MACHINE LEARNING IN SMART GRID The authors suggest and validate, through experiments with
MOBILE APPS real household data from smart meters (including temperature
During the last decade, researchers have managed to address readings), that consumers with high base loads and/or high
various problems using machine learning algorithms, e.g. temperature sensitivity of base load (which indicates ineffi-
computer vision and speech recognition, while also beat- cient refrigerators-freezers) are good candidates for energy
ing humans in games like Go and Atari. Similar techniques efficiency programs and should be targeted first.
have also been applied to smart grids at the grid system 1 Smaller flexibility factor means the user can undertake relatively less
level, for tasks like building automation, grid load prediction, portion of DR responsibilities and vice versa.
VOLUME 8, 2020 219641

new samples of that cluster. Convergence is achieved when

the clusters remain unchanged.
v
u n
uX
d(x, y) = t (xi − yi )2 (1)
i=1
Algorithm 1 k-Means
Input: number of clusters: k, set of data points: D
Output: k clusters
1: set k random data points as the initial cluster centroids
2: while clusters do not remain unchanged do
3: assign each vector to the closest cluster
4: recompute each cluster’s centroid
5: end while
FIGURE 5. Smart grid consumer profile example. 2) SMART GRID CONSUMER CLUSTERING APPLICATIONS
As part of our ongoing work on energy consumer profiling
In the sequel, we focus on prior work regarding smart grid and clustering, we have utilized the k-means algorithm to
consumer recommender systems, clustering, and classifica- cluster 653 households from the Pecan Street dataset [76].
tion, as well as hybrid models of these approaches, in order We constructed a profile similar to the one depicted in Fig. 5
to construct better user profiles and design energy efficiency for each house, consisting of typical normalized load profiles
strategies of higher impact. (daily, weekday, and weekend), consumption aggregates,
consumption ratios, and statistics for different time periods,
A. CLUSTERING resulting in a total of 85 features. The optimal number of
1) CLUSTERING BASICS clusters was found to be 5 using the elbow method, and
The goal of clustering is to group certain objects, so that the resulted average normalized daily load profile for each
those belonging to the same group/cluster are similar to each cluster (i.e. hourly average normalized consumption for each
other and dissimilar to those of other clusters, based on a cluster) is presented in Fig. 6. Normalization was applied to
set of predefined features using some similarity measure. the energy consumption data, by dividing the consumption
In the machine learning component of a smart grid mobile measurements with the household’s average consumption,
app system architecture, clustering algorithms can be trained so that the clustering is not dependent solely on the magnitude
and deployed to cluster consumers, households, or build- of each household’s consumption (i.e. bigger households with
ings. One advantage of unsupervised learning models, like more occupants would be assigned to the same cluster), but
clustering algorithms, is that the training data are unlabeled, mainly on the behavior of the occupants throughout the day.
meaning that smart grid measurements can be directly used All 5 clusters had a population between 102 and 165 house-
for clustering, without the need for time-consuming labeling holds, while it is evident from Fig. 6 that different consump-
procedures. Energy consumption measurements are used to tion behavior patterns are observed for each cluster, meaning
derive various input features, like statistical properties, for that diverse engagement strategies for energy savings should
the clustering models. The resulted clusters can be used to be adopted for each cluster of users. For instance, consumers
discover household consumption profiles, which are them- having low consumption during the day while consuming
selves utilized to determine similar users for recommendation excessive energy during afternoons, are ideal candidates for
systems, and to derive high-level consumer behavior patterns DR engagement actions.
and characteristics for personalized services and incentives. In [77], typical load profiles are derived by cluster-
For example, if a household belongs to a cluster that has ing of consumer load data. An algorithm called Affinity
a consumption profile with high consumption only from Propagation is utilized that, contrary to k-means, does not
19:00 to 23:00, we can assume that, most probably, it has need to take the number of clusters as an input. A simi-
all its occupants working outside the household during reg- lar model is proposed in [78], where a DR scheme using
ular business hours. One of the most studied and effective clustering of residential consumers is designed. Specifically,
clustering algorithms is k-means (Algorithm 1 below), where 10 features regarding each household, including square
starting with k vectors in the feature space, it iteratively footage, last month’s consumption and solar generation, num-
assigns each training observation to the cluster having the ber of appliances, and flexibility factor, are utilized as inputs
closest mean/centroid in terms of Euclidean distance (1), for the k-means algorithm.
where x, y ∈ Rn are feature vectors (e.g. of a consumer), and A more sophisticated approach is taken in [79], where
computes again each cluster’s centroid vector including the the authors use hourly energy consumption data from
219642 VOLUME 8, 2020

reduction technique of the original consumption data, before

applying the clustering algorithms. The results show that k-
means outperforms the other two models.
B. CLASSIFICATION
1) CLASSIFICATION BASICS
The goal of classification is to correctly determine the cat-
egories in which data points belong. Classification models
belong to the supervised learning paradigm, meaning that
such algorithms learn a function that maps input features to
output labels. A supervised learning algorithm uses train-
ing data containing output values for each input feature
FIGURE 6. Clustering use case example: Average normalized daily load vector x and tries to learn the correlation between them, thus
profile for each cluster. being able to make predictions for a given problem based
on unseen features. In order to accomplish this, supervised
220,000 households to derive consumer segments/clusters,
learning algorithms need labeled training data, e.g. spam/ham
and thus target users for demand response and energy effi-
emails, from which they learn a function that maps an input
ciency programs in a more effective way. They use an encod-
to an output feature vector. Some of the most popular clas-
ing system with a pre-processed load shape dictionary. Each
sification algorithms are Support Vector Machines (SVM),
household is then classified based on features derived from
Logistic Regression, k-Nearest Neighbors (kNN), Decision
the base load like entropy.2 Segmentation strategies for spe-
Trees, Random Forests, and Neural Networks.
cific targeting schemes like household targeting for energy
efficiency programs and recommendations for time-of-use
a: LOGISTIC REGRESSION
shifts are also proposed. The representative load shapes of
The logistic regression classifier takes its name from the
each cluster are derived using an adaptive k-means algorithm
logistic/sigmoid function (2).
and then summarized with a hierarchical clustering algo-
rithm. Five different segmentation strategies are derived, and 1
σ (x) = (2)
it is explained how these strategies can be implemented for 1 + e−x
different programs. For a binary classification problem, the probability of fea-
In [80], the authors use a clustering algorithm called Finite ture vector x belonging to class C is:
Mixture Modeling to cluster residential customers based on 1
behavioral characteristics extracted only from their energy P(C|x) = , (3)
1 + e−w·x
consumption patterns. Four key time periods during the day
are defined, i.e. overnight period, breakfast period, daytime where w is the weight vector learned during the training phase
period, and evening period. Seven statistical attributes are by minimizing the cross-entropy of the examples using an
derived only from the consumption data and are used as inputs algorithm like Stochastic Gradient Descent (SGD). As far as
for the clustering algorithm. multinomial logistic regression is concerned, also known as
In [81] the authors use the k-means algorithm to cluster softmax regression, for the multi-class classification problem
energy consumers based on one-minute resolution power with n separate classes, the probability of x belonging to
consumption data averaged hourly. Average seasonal curves class Ci is:
are calculated for each household, which are then used as ewi ·x
P(Ci |x) = Pn wj ·x , (4)
inputs for the clustering model. Two clusters were derived j=1 e
for each season leading to a total of 8 load profile clus-
ters. A regression analysis was also conducted with explana- where wi is the weight vector for class Ci .
tory features from survey data, to examine the correlations
b: K-NEAREST NEIGHBORS
between a user’s survey answers and the corresponding clus-
ter. The study showed that features like remote working, Another widely applied classifier is the k-Nearest Neigh-
television watch time per week, and education level had bors (kNN) algorithm. During the training phase of the
meaningful correlations with the derived cluster. Such kNN algorithm, the feature vectors of the training examples
insights can be very useful for utilities to design personal- are simply stored in memory. Afterwards, when an unseen
ized energy efficiency strategies. A similar approach is taken instance needs to be classified, the closest k training examples
in [82], where k-means, Expectation Maximization (EM), are observed and their majority class is returned. In order to
and Self-Organizing Maps are applied to household consumer achieve better results, distance weighting can be added to the
data to derive behavioral patterns. The representative load model by giving a weight to each neighbor, which decreases
curves are calculated for each home, as a dimensionality when the distance from the point being classified increases.
As a distance measure, Euclidean distance or cosine similar-
2 Amount of consumption variability. ity are most often utilized.
VOLUME 8, 2020 219643
c: SUPPORT VECTOR MACHINES weak learners is boosted via a voting scheme. A Random
The Support Vector Machine (SVM) algorithm tries to find a Forest essentially consists of many Decision Trees. For the
separating hyperplane between instances of different classes development of each tree, a subset of the training data is used
in a high-dimensional space (Fig. 7). Specifically, it tries to at random but with replacement (bootstrap sampling). If K is
obtain the best possible separating hyperplane, which is the the number of input features, we specify a constant number
one with the highest distance (functional margin) from the k < K so that for every node, k out of K features are selected
nearest training instance of any class (support vectors). When randomly (random feature selection). Every Decision Tree is
the training data points are not linearly separable in a low fully grown without pruning using a classic CART approach
dimensional space, the SVM algorithm maps them into a and after the construction of the forest, all the predictions are
higher dimension, using a kernel function, in order to find averaged in order to form the final output.
the separating hyperplane. Popular kernel functions include
linear kernels, polynomial kernels, and the sigmoid kernel. e: DEEP LEARNING
The building block of a Neural Network is the neuron which
consists of an input feature vector x, a weight vector w,
a bias term b and a non-linear activation function f (most
often tanh, sigmoid, or ReLU). Each neuron computes the
input weighted sum, adds the bias term, and passes the result
through an activation function producing the real-valued out-
put of the neuron. One of the most popular Neural Network
architectures is the Multi-Layer Perceptron (MLP). An MLP
consists of a given number of layers, the input layer, the hid-
den layers, and the output layer (Fig. 8). Each layer has a
number of neurons and each neuron uses all the outputs of
the previous layer neurons as an input:
hn = f (wn T x + bn ), (5)
where hn is the nth neuron of the hidden layer, x =
[x1 , . . . , xk ] is the input feature vector, wn is the weight
FIGURE 7. Support Vector Machine two-dimensional example, where x is vector, and bn is the bias.
the input feature vector, w is the weight vector, and b is the bias.
A Neural Network learns all the weights and biases from
training data using an algorithm called backpropagation [85],
In other words, the SVM algorithm aims at maximizing the which is responsible for distributing the error gradient across
margin between the support vectors and derived the separat- all the network nodes using the chain rule. The weights are
ing hyperplane, which is a quadratic programming problem, randomly initialized to small real values using for example
and can be solved with standard methodology. SVMs can a uniform or Gaussian distribution. Then in each epoch3
achieve remarkable classification performance, while deriv- and for each mini-batch4 of training examples, the output
ing an optimal global solution contrary to other machine of the network is computed and for each weight w, ∇E(w)
learning models like Neural Networks that find local min- is calculated, where E(·) is the loss function output for
ima. However, SVMs suffer from computational and memory this specific mini-batch. The gradients are calculated from
issues in nonlinear problems with large datasets. right to left and each weight is updated using an optimizer.
The most commonly used optimizer is Stochastic Gradient
d: DECISION TREES AND RANDOM FORESTS
Descent (SGD), where the weights are updated using the
The Decision Tree is a tree-like structure that simulates the following rule:
segmentation of the training data. Each node represents a
rule/question regarding a feature, each branch represents the w ← w − η∇E(w), (6)
answer to the question, and each leaf node represents an
output class label. The generated set of rules can then be used where η is the learning rate. More sophisticated optimiz-
for the prediction of new instances. Some of the most popular ers to minimize loss are RMSprop [86], Adagrad [87], and
algorithms that generate Decision Trees include CART [83] Adam [88], which are widely used to train Neural Networks.
and ID3 [84]. Decision Trees are simple to use and interpret, Another popular Neural Network architecture is the Recur-
but they have some major disadvantages, one of the biggest rent Neural Network (RNN). Unlike feed-forward Neural
being lack of robustness, meaning that a small alteration in the Networks, RNNs consist of nodes/blocks which form a
training data can result in a significant change in the overall directed graph (Fig. 9). Each node has its own state/memory,
architecture and the final predictions. Random Forests came 3 An epoch refers to iterating the whole training dataset just once.
to the rescue to solve this and other problems by applying an 4 A mini-batch is a small subset of the training data used to calculate the
ensemble learning approach where the performance of some error gradient and update the Neural Network weights.
219644 VOLUME 8, 2020

to provide the most appropriate incentives to each user, and

hence maximize the probability that a suggestion is carried
out, as well as the campaign’s overall impact.
In [90], deep learning is utilized to model occupant charac-
teristics, leading to the design of a more effective and sophis-
ticated gamification approach. More specifically, a deep
bi-directional RNN with Long Short-Term Memory (LSTM)
cells is used to predict on/off events of appliances like ceiling
fans, lights, and A/Cs, and thus learn the user’s behavior.
A social game is also introduced to incentivize occupants
to reduce their energy consumption by rewarding them with
points and potential prizes if their consumption during a day
is lower than their baseline. A similar approach was taken
in [53], using machine learning algorithms to predict the
ON/OFF state of household appliances, and thus recognize
human activities like cooking, and washing of clothes and
dishes. In [51], the authors use appliance and water consump-
FIGURE 8. Multi-Layer Perceptron (MLP) example that has an input layer tion data as input features to train a classifier that predicts
with k features, a hidden layer with n neurons, and an output layer with the occupancy of a household, i.e. the number of occupants
m neurons.
present in the house.
enabling RNNs to perform much better than feed-forward Several works have utilized consumption data as input fea-
Neural Networks with sequences of inputs and tasks like tures to classify households into socio-demographic classes
speech recognition, time-series forecasting, and natural lan- derived from questionnaires [91]–[94]. More specifically,
guage processing. in [93], the authors conduct household classification regard-
ing various properties like number of residents and floor area,
using 22 features extracted only from the consumption load.
They evaluate the model on real residential data, and verify
that a priori knowledge of certain features can increase the
classifier performance. In [92], the authors build on the work
of [93]. They extend the input feature space from 22 dimen-
sions to 88 and then perform feature filtering techniques to
select the most important features based on classification
performance. They also make changes to some of the property
definitions and manage to further improve the household
FIGURE 9. Recurrent Neural Network (RNN) example for a time window
of size 3, where xt , yt , and ht are the input feature vector, output feature classification results in most cases. In [91], previous works
vector, and hidden state vector respectively at time t . are extended by incorporating weather data as additional
input features to predict 19 classes, using the Random Forest
2) SMART GRID CONSUMER CLASSIFICATION classification algorithm. An application of deep learning for
APPLICATIONS consumer behavioral modeling is studied in [95]. A Convo-
Classification models can be used in smart grid mobile apps lutional Neural Network (CNN) is implemented to extract
to classify new consumers in terms of socio-demographic highly nonlinear relationships between household consump-
characteristics and energy-related behavior. These insights tion data and consumer socio-demographic characteristics
are quite important for utilities and service providers, in terms derived from surveys, like age, income, number of children
of accurate user profiling and behavior prediction, which then and floor area. The classification performance of the CNN
are used to develop recommendation and incentive strategies is enhanced by replacing the softmax output layer with a
with greater impact and lower costs. Support Vector Machine (SVM).
In [89], the authors present a use case where an energy
efficiency campaign manager wants to recruit household
occupants, based on their profile and preferences, through a C. HYBRID MODELS
mobile app. These preferences are learned for each consumer In this subsection, we present approaches where both clus-
using a logistic regression classifier, trained on historical tering and classification algorithms are used to model con-
data of demand reduction suggestions and corresponding sumer behavior and characteristics more precisely. Insights
user reactions, i.e. if the consumer accepted the suggestion derived from unsupervised learning methods like cluster-
or not. Different consumers have different preferences and ing, where more data can be provided since there is no
reactions to suggestions and incentives. Thus it is important need for labels, can then be used as inputs for classification
for the manager to derive such preference models, in order problems.
VOLUME 8, 2020 219645

Socioeconomic characteristics along with smart meter are widely used to make personalized recommendations of
household consumption data are used in [96] for household products and services like movies, music, and electronic
load pattern forecasting. After the electricity load data are devices to consumers, based on their preferences. In the basic
clustered and a number of clusters with representative load RecSys problem, there is a set V of users, and a set I of items,
patterns are derived, the authors select a subset of the socioe- which are candidates for recommendation. The usefulness of
conomic input features for each cluster, based on the corre- item i ∈ I to user v ∈ V is measured by a utility function
lation between each feature and the cluster centroid pattern. U , i.e. U : V × I → R, with R being a totally ordered
Then, they train a separate deep Neural Network for each set (e.g. 0-5 stars or a continuous set e.g. [0, 1]). The goal
cluster, with the selected features as inputs, that outputs the of a recommender system is to choose items i ∈ I which
probability that the consumer belongs to the specific cluster, maximize U for each user v ∈ V. The utility function can
resulting in a binary classification model. be learned from past user behavior, user and item similarity,
In [97], a hybrid model that clusters consumers based item context, etc.
on daily load curves and then classifies them to different The most well known RecSys techniques are Collaborative
categories using the k-Nearest Neighbors (kNN) algorithm Filtering (CF) and Content-Based, with the former relying
is proposed. Despite the fact that this study is conducted for on user similarity measures to recommend items to a user
non-residential data, it can be transferred to household use that users with similar profiles have liked, and the latter
cases as well. relying on item profiles (e.g. descriptions and other features)
A similar model is implemented in [98], where smart meter to recommend similar items to those that the user liked in the
and survey data are used to classify new electricity customers past.
into consumption profile categories. This model can help The CF paradigm mainly consists of two methodologies,
utilities find which consumer attributes are correlated to their user-user and item-item. In user-user CF, for each user v ∈ V
consumption behavior, and thus enhance their engagement the algorithm finds a set N of other users who have similar
and energy efficiency through appropriate actions like per- ratings with user v in terms of a chosen similarity measure
sonalized services. At first, smart meter data are processed to (e.g. Jaccard similarity, Cosine similarity, or Pearson cor-
derive representative demand profiles, which are then used to relation coefficient). Then, the ratings of user v are calcu-
cluster the consumers into groups. Afterwards, these group lated based on the ratings of users belonging to set N (e.g.
labels are the ground truth labels that the supervised machine by averaging), and the items with the top derived ratings for
learning algorithms use for the training procedure, while their v are recommended. In item-item CF, for each item i ∈ I
input features are survey data and a limited amount of smart the algorithm finds other similar items in terms of other user
meter consumption data. Based on the input features, a new ratings, and calculates the rating of item i based on the ratings
customer with few historic consumption data available can of the target user on these similar items.
be classified into one of the consumer groups. The survey In the Content-Based approach, a predictive model for each
input features include variables like sex, age, education level, user’s preferences is derived, based only on the past user
income, home age, heating type, and types of available behavior and actions (e.g. ratings, clicks, accepted sugges-
appliances. tions, etc.), as well as on the content of the items that the user
In [99] a similar approach is taken into account, where likes. Thus, similarity measures are used to find similar items
three clustering methods are evaluated based on household to those that the user has liked in the past.
consumption data, k-means, k-medoid and self-organizing Matrix factorization or latent factor models [100] comprise
maps, with the latter showing the best results during exper- a different category of CF algorithms, that utilize latent fac-
iments. Afterwards, each day was clustered independently so tors to describe user and item attributes and characteristics
that seasonal characteristics of the household are not lost, from item-rating patterns. Such models can include additional
i.e. a consumer might belong on a different cluster on dif- implicit user feedback and behavioral features (e.g. search
ferent days. Averaging the overall consumer consumption history, past purchases, visited pages, etc.) when explicit user
for each day in a cluster, resulted in the final vectors of feedback, such as ratings, is not available. In a basic matrix
each cluster-profile class (10 profile classes were derived). factorization model that uses latent factors of dimensional-
A multi-nominal logistic regression model was trained to ity f , each item i ∈ I is represented by a vector qi ∈ Rf ,
predict the probability of a consumer belonging to a profile and each user v ∈ V is represented by a vector pv ∈ Rf . The
class, based on consumer’s characteristics like house type, item-user utility function U or rating matrix is approximated
number of bedrooms, family type, social class, heating type, by calculating the dot product:
occupant age, and available appliances.
r̂vi = qi T pv , (7)
D. ENERGY EFFICIENCY WITH RECOMMENDER SYSTEMS and captures the utility of item i to user v. The goal of matrix
1) RECOMMENDER SYSTEMS BASICS factorization is to find latent factor vectors qi and pv that lead
The ever-growing amount of digital content available online to low error (e.g. Root Mean Squared Error - RMSE) between
has raised the need for personalized recommendation tech- the real utility matrix part used for training, and the predicted
niques now more than ever. Recommender Systems (RecSys) utility values of the model. These latent factor vectors can be
219646 VOLUME 8, 2020

derived with algorithms such as Stochastic Gradient Descent often generated by smart meters and IoT sensors installed at
(SGD), while regularization techniques can be used to avoid the user’s premises, and have to pass through a pre-processing
overfitting on the training segment of the utility matrix. To phase before being presented to the consumers in a mean-
capture the effect of various external factors on the resulted ingful way. This phase includes techniques like data cleans-
user-item utilities, bias approximation approaches can be ing, aggregation, and enrichment. A visualization module
integrated into (7), which becomes: should follow specific design guidelines to raise consumption
awareness and spark consumer behavioral change. According
r̂vi = µ + bi + bv + qi T pv , (8)
to [106] these guidelines include visualizing electricity con-
where µ is the average utility, bi is the item bias, and bv is the sumption in a comprehensible manner, presenting particular
user bias. For instance, the item bias term bi might represent aspects of consumption, and making comparisons to related
the fact that item i has a higher utility than average, while references.
user bias bv might represent a situation where user v tends to
experience lower utility compared to the average user. 1) CONSUMPTION VISUALIZATION
However, the utilization of RecSys in the smart grid sector Some of the surveyed mobile apps in this study, only visualize
is sparse, and only a few studies have tackled this area. energy consumption, which is the most important aspect
RecSys can be utilized in smart grids for scenarios like of electricity consumer behavior. Power Explorer [34] only
energy-saving appliance recommendation, utility retail plan visualizes the current total power level with a bar. The Green-
recommendation and load shift recommendation [101]. Smart Soul app [47] presents both total and device power consump-
meter consumption data along with household basic information to the users, while ChArGED [44], [45] also provides
tion and user preference data are processed by such systems historical data visualization. The Social Power App [60],
to derive better recommendations. except for total hourly consumption and history visualization,
also presents a comparison with the baseline consumption of
2) APPLICATIONS OF RecSys IN SMART GRIDS the user.
In [102], a system that makes recommendations to elec-
2) ADVANCED VISUALIZATION
tricity consumers based on their profiles is presented. This
Most of the surveyed apps present additional visualization
system recommends other cheaper tariff plans for specific
services to end-users, in order to inform them in a richer way.
users, as well as how much energy can be saved by shifting
EnergyLife [35], EnergyCoupon [36], EnergyWiz [37] and
specific appliance loads. In [103], a thorough description
GreenPlay [64] visualize additional metrics like energy sav-
of an electricity retail plan recommender system is con-
ings, costs, and CO2 emissions. PEAKapp [41] incorporates
ducted. More specifically, this recommender system uses a
all the visualizations mentioned above, as well as solar gen-
Collaborative Filtering algorithm, which utilizes consump-
eration graphs, consumption forecasts, and comparison with
tion data to offer similar retail plans to similar consumers.
other households. Similarly, the Neurio app [68] visualizes
A clustering phase takes place, to identify which consumers
current and historic consumption, costs, savings, solar gen-
are similar, and historical retail plan data are used to train
eration, and comparison with neighboring houses. GAIA [46]
the model. In [104] and [105], the authors use pattern
visualizes current total and appliance consumption, history
mining techniques to extract energy-saving patterns from
data, as well as other sensor metrics related to climate, lumi-
household smart meter (appliance level) data and create
nosity, noise levels, and occupant presence. Furthermore,
association rules. These rules are then used as part of a
it presents consumption anomaly detection events, building
recommender system that matches the incoming real-time
clustering, and statistics. The enCOMPASS [48] app presents
household data with the association rules pool, and if a pattern
the basic insights of total and device power consump-
is matched, the relevant action of the rule is recommended to
tion, and historical data, but it also visualizes personalized
the user.
consumer impact (i.e. monetary, environmental, hedo-
nic), and goal progress as a ‘‘charging battery’’ graph.
VI. COMPARISON OF EXISTING APPS
ENTROPY [58] shows real-time total consumption data,
The mobile applications presented in section IV all have a
household baselines, target/goal progression, and daily num-
common underlying objective, that of engaging and motivat-
ber of energy peaks. MyEarth [39] and JouleBug [69] only
ing electricity consumers to become more energy-efficient,
present CO2 and water savings related to user activities,
i.e. reduce their consumption, or shift their loads. Some of
which are not derived from smart meter data but directly
them use similar approaches to achieve this goal, while oth-
through the app interface. Furthermore, the Smappee [66]
ers propose new and innovative concepts. A comprehensive
application visualizes current total and appliance consump-
comparison of the existing apps is conducted in this section,
tion, historic data, solar generation, gas and water consump-
with Table 1 serving as a summary to guide the discussion.
tion, costs, aggregates/analytics, and comparison with other
A. DATA VISUALIZATION locations. Last but not least, the TED [67] app shows current
One of the most common features among the applications power consumption, costs, history, and total location grid
presented in this study is data visualization. Such data are demand.
VOLUME 8, 2020 219647

TABLE 2. Smart grid mobile app most common gamification features.
B. GAMIFICATION home automation systems do, which means that despite the
As we analyzed in section III, gamification can substan- fact that most apps provide incentives for certain actions
tially enhance electricity consumer engagement and provide (e.g. to turn off the A/C), the user has to physically perform
incentives for consumption reduction. For that reason, a great these actions. With apps like Smappee [66] and Constella-
number of mobile applications targeted at energy efficiency tion Connect [70], users can control/actuate their registered
incorporate gamification techniques like the ones we men- devices directly through a mobile app, as well as automate
tioned earlier in the survey. household energy flows. Constellation Connect also allows
Some mobile apps like Power Explorer [34], Ener- users to create triggers on certain events.
GAware [57], and Tribe [63] are essentially serious mobile
D. USER PROFILING
games that aim at reducing user energy consumption through
various graphic interfaces and gaming interactions. Others Constructing detailed user profiles based on electricity con-
like PEAKapp [41], ChArGED [45], GreenPlay [64], and sumption patterns and behavioral characteristics can help
MyEarth [39] include simpler serious games, alongside the utilities to better understand their consumers, and thus enact
rest of their features, mainly related to the users’ con- more targeted and personalized services to achieve greater
sumption and actions. However, as depicted in Table 2 the impact. Data acquired through a mobile application, like the
most common gamification features of the studied mobile ones studied in our work, can be effectively utilized for the
apps are challenges, points, levels, badges, rewards, leader- purpose of user profiling.
boards, competitions, goals, and quizzes. In addition, Ener- EnergyCoupon [36] analyzed the energy consumption
gyCoupon [36] uses coupons, ENTROPY [58] shows user profile of the users to validate the effects of the mobile
progression, and the Social Power App [60] has a competitive app in terms of demand reduction and savings. Green-
structure for competitive users, and a collaborative structure Soul [47] incorporates a profiling module, that not only col-
for users to reduce consumption collectively. lects socio-demographic data from an in-app questionnaire,
but also captures behavioral insights from smart meters and
C. FLEXIBILITY user interaction. The user profiles extracted are then used
Flexibility refers to dynamic consumption adjustment from as input for data analytics and forecasting. enCOMPASS
the consumer’s side, most often at an appliance level, for enhances the user profiling procedure by detecting appli-
energy efficiency and demand response purposes. Household ance states and thus recognizing user activity patterns [53].
occupants can provide flexibility either manually by shifting ENTROPY [58] categorized users based on questionnaires
energy-consuming activities and switching on/off devices, into profile types (i.e. philanthropist, socializer, free spirit,
or automatically by letting a third party (e.g. their utility or achiever, player, and disruptor), while also extracting energy
a mobile app) to directly control pre-specified appliances. load profiles with clustering algorithms.
Consumers providing flexibility can immensely help energy
utilities and grid managers to evenly balance electricity E. MACHINE LEARNING TECHNIQUES
demand, and thus avoid peaks that can cause blackouts and As discussed in section V, integrating machine learning
substantial emission increase. techniques into mobile applications for energy efficiency
Most of the studied mobile applications do not implement can substantially enhance end-user engagement with refined
features that can help the user with flexibility, e.g. controlling incentive mechanisms. However, only a few of the existing
devices directly from the mobile app like many existing smart apps use such approaches.
219648 VOLUME 8, 2020

EnergyCoupon [36] utilizes classification Decision Trees H. PASSING NOTIFICATIONS AND SUGGESTIONS
to predict if the future electricity price is going to be TO USERS
high. GAIA [46] uses the k-means algorithm to clus- The majority of mobile applications for consumer energy
ter buildings based on seasonal and yearly energy con- efficiency use tips, suggestions, and notifications to keep
sumption, as well as on building characteristics from their users engaged and communicate their incentives in an
their knowledge base. It also performs anomaly detection effective and user-friendly manner. However, such mecha-
with a Generalized Additive model. GreenSoul [47] and nisms should be designed with great care, since app users
enCOMPASS [51] use machine learning techniques for occu- can become irritated if they get overwhelmed with tips and
pancy prediction, while enCOMPASS also conducts user notifications throughout a short period of time.
activity profiling through appliance state prediction with Power Explorer [34] alerts its users when a high con-
models like Support Vector Machines (SVM) and Neu- sumption load is detected. Social Power App [60], Green-
ral Networks, and user clustering based on consumption Play [64], and Neurio [68] send general energy-saving tips
behavior with k-means. ENTROPY [58] also applies the and advice to their users. In addition, EnergyCoupon [36],
k-means algorithm for user clustering based on socio- ChArGED [44], and enCOMPASS [48] also provide person-
demographic data, household characteristics, and smart meter alized energy efficiency tips, based on the conducted user
measurements. Additionally, ENTROPY includes consump- profiling. EnergyLife [35] uses contextualized tips based on
tion forecasting using popular machine learning techniques specific user actions and triggers, in order to change their
like Support Vector Regression (SVR). behavior. ecoGator [40] also sends simple pieces of advice,
hints, and shopping tips for energy-efficient products to the
F. RECOMMENDER SYSTEMS INTEGRATION consumers, while PEAKapp [41] notifies them with sav-
Recommender System (RecSys) techniques can provide very ing tips, and discounts. GreenSoul [47] uses high-level rec-
important insights to utilities regarding user preferences and ommendations/advice and context-aware notifications, i.e.
thus help them to develop incentive techniques. Despite different message formats are used for different user types
that, most mobile apps providing incentives for energy effi- (e.g. generic, energy, financial, and environmental mes-
ciency do not take full advantage of the available RecSys sage formats). The ENTROPY [58] app also utilizes gen-
approaches. eral tips and personalized recommendations based on user
GAIA [46] and ENTROPY [58] use rule-based sys- types, and notifications with invitations for environmental
tems for their personalized recommendations. enCOM- awareness events. Furthermore, Tribe [63] educates con-
PASS [48] also makes customized recommendations based sumers on energy efficiency measures for public build-
on user clustering, while also utilizing collaborative fil- ings, while TED [67] sends alerts/messages for tariff
tering techniques to make similar energy-saving sugges- changes, excessive demand load for the user’s location, and
tions to similar users in terms of consumption profile, surpassing power consumption thresholds. JouleBug [69]
while also taking into account user feedback regarding these uses push-notifications for gamification, social activity, and
suggestions [50]. reminders, while Constellation Connect [70] alerts occupants
for events and triggers inside the house, mainly related to
G. USER FEEDBACK AND ELICITATION
security aspects.
User feedback, either in the form of direct feedback regarding
specific app features or in the form of in-app questionnaires, I. SOCIAL MEDIA INTEGRATION
can help build better user profiles, targeting strategies, and Social media have a great impact on our lives in recent
more accurate models to achieve higher levels of energy years, and are the main tool in the hands of advertisers and
efficiency. ecoGator [40] integrates feedback with user rat- engineers to incentivize end-users of various products. This
ings for the energy efficiency tips it provides. Similarly, is also the case in the smart grid sector, where many exist-
GreenSoul [47] gives its users the ability to express positive ing mobile apps aimed at energy efficiency integrate social
or negative feedback on the messages the mobile app sends. media as a central component of their consumer engagement
Furthermore, consumers can fill an in-app questionnaire mechanisms.
regarding socio-economic characteristics. enCOMPASS [48] EnerGAware [57] allows players to communicate with
also uses a questionnaire to determine what appliances the each other and read related news. EnergyLife [35] devel-
user owns, while occupants can give feedback on the energy oped an in-game community for users to interact with
efficiency tips (i.e. ‘‘Okay, will do,’’ ‘‘already doing this,’’ each other and share knowledge and experience related to
‘‘not for me’’). In ENTROPY [58] consumers can fill ques- consumption reduction. Moreover, PEAKapp [41], ecoGa-
tionnaires that include demographic data. Implicit feedback tor [40], GreenSoul [47], ENTROPY [58], JouleBug [69],
can be derived from the ENTROPY app (i.e. a tip has been EnergyWiz [37], and GAIA [46] give their users the ability
read by the user), while users can provide explicit feedback to share in-app accomplishments on popular social media
by accepting or declining a task. Smappee [66] includes an platforms. In addition, EnergyWiz [37] includes one-to-one
in-app survey where users declare parameters like building energy efficiency challenges and user comparison on Face-
type, family structure, and appliances. book, while JouleBug [69] integrates both sharing on social
VOLUME 8, 2020 219649

media platforms and on its own social activity section with a validation results regarding household users. Fig. 10 does not
trending feed, likes, comments, and follows. reflect a direct comparison between the impact of the apps
since they were validated under very different circumstances
J. MEASURED IMPACT in terms of number of users, location, and time frame, but it
Arguably, one of the most meaningful aspects of a mobile serves as a visual representation of the measured impact for
app aimed at consumer energy efficiency is its resulted the reader.
impact. In other words, the essence of each app’s objective
is the amount of energy consumption reduction and energy
efficiency its users accomplish. Thus, it is crucial to compare
the studied mobile apps in terms of validation outcomes
and results achieved, using appropriate Key Performance
Indicators (KPIs).
However, this is possible only for some of the apps we
present in this paper, which were developed and validated
through research projects and pilots. Information for the mea-
sured impact of commercial products and standalone mobile
apps is either partial or not existent. KPIs that reflect each
app’s impact include energy consumption (reduction in kWh
or as a percentage), costs of energy consumption, greenhouse
gas emissions, and various app interaction statistics. In this FIGURE 10. Measured impact of mobile apps used by residential users.
subsection, we focus on energy consumption since it is the As for the mobile apps that were deployed for building
most common KPI measured by most of the apps. occupants (e.g. offices, schools, and public buildings), greater
Nonetheless, a direct comparison of the energy efficiency savings were achieved since such buildings had a consider-
impact achieved by the apps is not possible since the val- able consumption footprint in the first place. ChArGED [45]
idation approaches and the pilots of the projects are very reported an average consumption reduction of approximately
different in terms of user types, number of users participating, 40% and GAIA [46] measured savings between 15%-20%
and duration. For example, some projects recruited residential for several school buildings. GreenSoul [47] reported average
users while others focused on public buildings and offices. savings of about 10% for a realistic scenario, while Tribe [63]
Furthermore, the validation procedure carried out by some stated a consumption reduction between 15%-20%. Finally,
of the residential projects only included a small number of ENTROPY [58] managed to reduce average occupant energy
houses, meaning that their results have not the same value consumption by approximately 16%.
compared to larger validation phases, even if the energy As it is evident both from Table 1 and Fig. 10, most of the
savings seem greater. studied apps achieved significant energy savings for diverse
More specifically regarding apps used by residential con- use cases, although there is still room for improvement
sumers, Power Explorer [34] achieved an average energy regarding substantial long term energy savings, especially
consumption reduction of 14% for a small sample of 15 users, for residential consumers. In addition, one can argue that
during a 7-day trial. On the other hand, EnergyCoupon [36] the integration of machine learning for residential consumer
focused on energy consumption shifting to off-peak hours modeling can significantly improve the resulted impact of a
instead of consumption reduction in general. The app’s smart grid mobile app in terms of household consumption
impact was measured through a 3-month validation period reduction, as we observed with the enCOMPASS [48] project,
with 10 users, resulting in an average energy shift of 5% which achieved greater energy savings compared to other
of the total consumption, while some users reached an aver- residential projects (with a large number of recruited houses)
age shift of up to 20%. Considerable residential impact was that did not utilize machine learning approaches.
also reported by EnerGAware [57] and the Social Power
App [60] with an average energy consumption reduction of VII. OPEN ISSUES AND FUTURE WORK
about 5% (for 44 households during 2 years) and 8% (for After conducting a thorough comparison among existing
108 households during 4 months) respectively. PEAKapp [41] mobile apps for smart grid consumer engagement, it is clear
and GreenPlay [64] conducted larger validation procedures that most of them use data visualization, gamification, and
with 811 (during 17 months) and 157 (during about one year) energy saving tips as their main motivation and engagement
users respectively. PEAKapp achieved average energy sav- mechanisms. However, as it is evident from Table 1, none of
ings around 1%, while GreenPlay [64] did not show substan- the existing mobile apps integrates all the features studied in
tial energy savings in practice. Finally, the enCOMPASS [48] this survey, and thus there is still room for improvement in
project conducted a validation phase with 223 households, order to achieve higher levels of electricity consumer engage-
that lasted 10 months and showed separate energy savings for ment and consumption reduction.
the different pilot sites that averaged around 10%. In Fig. 10 Future directions should also focus on the integration and
we illustrate the measured impact for the apps that provide deployment of machine learning and recommender systems,
219650 VOLUME 8, 2020

as well as architecture interoperability, big data, and security. of the recommendation message, overall past user response,
Furthermore, several of the studied mobile app features are recommendation type, etc.
interrelated, e.g. machine learning and RecSys integration One more related aspect is that of feature engineering,
can affect notifications and user profiling. Thus it is crucial to i.e. how should the recommendation features/parameters be
design these features having in mind potential dependencies tuned, to maximize the probability of adherence by a specific
that might affect them, in order to achieve a smoother user user, since each user has different profile characteristics. Such
experience. parameters can include the timing of the recommendation,
In the future, mobile apps should be a fundamental the incentive provided (e.g. monetary or ecological), and the
component of the smart energy grid ecosystem, constantly wording used. Hence, an additional open issue is the optimal
providing bidirectional real-time communication between matching of the generated incentives and recommendations
end-consumers and the smart grid. Users should have the to the most appropriate consumers based on the advanced
ability to monitor and control all the energy-related aspects consumer profiling conducted, in order to increase the rec-
of their lives in the palm of their hands, while constantly ommendation impact.
receiving crucial information about the current, past, and Furthermore, the integration of deep learning models in
future state of their energy-related activities, consumption, smart grid mobile app systems is still quite sparse, despite
generation (e.g. solar panels), and impact, through advanced the immense success such algorithms have achieved in other
and interactive visualization approaches. domains. More complex models like deep learning and rein-
At the same time, smart grid mobile apps should provide forcement learning could be tested and adopted by smart grid
advanced consumer flexibility and feedback mechanisms to mobile apps, in order to derive high-level behavioral charac-
the users, which still remains an open issue to this day. More teristics regarding electricity consumers, buildings, and the
specifically, demand response mechanisms can be optimized smart grid as a whole. More specifically, deep learning archi-
and achieve a higher amount of demand shifting, if they tectures like Convolutional Neural Networks (CNNs) and
utilize real-time flexibility schedules for multiple appliances Recurrent Neural Networks (RNNs) can be used to extract
and tasks, while user feedback regarding the quality and highly non-linear relationships between energy consumption
acceptance of the proposed demand response actions will measurements and consumer characteristics, as well as to
help the system to make adjustments. The integration of build better recommendation engines as in [107]. To achieve
machine learning and recommender system techniques in that, such models should be trained on large datasets consist-
such approaches can further improve the system’s perfor- ing of complex consumer profiles and smart meter data of
mance and increase energy savings, but still remains an open diverse nature and granularity, e.g. appliance, whole building,
issue. We envision to tackle such issues related to flexibility indoor/outdoor climate, occupancy, and other measurements.
and machine learning utilization as part of our future work Consequently, more accurate clustering and classification
during the InterConnect project. can be achieved, meaning that the mobile app will be able
A. MACHINE LEARNING AND RECOMMENDER SYSTEMS
to provide more targeted personalized feedback and recom-
mendations, thus achieving greater engagement and energy
Smart grid mobile apps such as the ones presented in this
efficiency.
survey, generate and consume a significant amount of elec-
Moreover, reinforcement learning agents can learn what
tricity consumer data, which if processed appropriately can
recommendations are useful to propose to users through feed-
provide crucial insights regarding user behavior, consump-
back, either explicit (e.g. user accepts/declines the recom-
tion patterns, and the smart grid in general. However, few of
mendation) or implicit (e.g. after a load shift recommendation
the existing apps utilize advanced user profiling techniques
the smart meters record a drop in consumption). The agent
and machine learning, and those that do are research projects,
will learn, from a set of possible fixed recommendations,
meaning that such approaches have not been adopted yet by
the probability that a user will follow what the recommen-
utilities and the industry.
dation suggests in a short time horizon.
The same is the case with recommender systems, since
just a few research projects take advantage of them to B. SYSTEM ARCHITECTURE AND BIG DATA
make more effective energy-related recommendations to con- One area where smart grid mobile app systems could
sumers. Open issues regarding recommender systems in be improved is architecture interoperability, meaning that
smart grid mobile apps include the design of specific recom- the various system software components should be eas-
mendation formats, soliciting explicit and implicit feedback ily replaced if needed so that more entities and services
from consumers to continuously learn, and leveraging rec- are involved. For instance, a new smart home device or a
ommender system theory and algorithms while identifying new electricity provider should not lead to major adjust-
differences and specifics to smart grid recommendations. ments in the system architecture. This can be achieved
For instance, detailed energy consumer profiles should be by utilizing guidelines, standards, and ontologies like
constructed based on various parameters, such as consump- SAREF [108], and REST APIs. The Smart Application REF-
tion features, socio-demographic characteristics, and past erence ontology (SAREF) constitutes a reference ontology
responses to recommendations. In addition, item profiles specially designed for the IoT ecosystem. SAREF incorpo-
should also be derived based on their context, e.g. the text rates a variety of classes, objects, and properties to make the
VOLUME 8, 2020 219651
communication with different IoT devices seamless, even if privacy, by collecting household IoT measurement data in
they utilize different standards and protocols from different a secure, robust, and privacy-preserving manner [112], and
manufacturers. The mobile app we are developing as part of provide access to it under specific and tight authorization
our ongoing work in the InterConnect project will take full rules [113].
advantage of the SAREF ontology to collect and process data E. FUTURE WORK
from various sources while ensuring interoperability. Other After identifying open issues in the area and highlighting
technical IoT and energy standards that should be considered future research directions, we present our envisioned mobile
in the system architecture design of a smart grid mobile app for energy consumer engagement, and overall residential
app include: the Universal Smart Energy Framework (USEF) consumption reduction, which is part of our ongoing work
[109] for communication with flexibility markets, the Open during the InterConnect project. The mobile app will incor-
Charge Point Interface (OCPI) protocol [110] for smart elec- porate all the consumer engagement features identified in
tric vehicle charging, and the IEC 61970 standards [111] for this survey, i.e. advanced visualization, gamification, flex-
energy management system APIs. ibility, detailed consumer profiling, feedback mechanisms,
Despite the recent advancements in mobile device per- sharing through social media, machine learning models,
formance, apps might face operational problems when pro- and tips/notifications with recommendations for energy effi-
cessing a large volume of smart grid measurement data with ciency utilizing recommender system methodologies.
high variety and velocity. Hence, smart grid mobile apps More specifically, our new architecture will focus on the
must be supported by appropriate backend systems with big integration of machine learning and recommender system
data processing capabilities that incorporate frameworks like components to achieve higher levels of engagement and
HDFS [21], Kafka [22], or Spark [23]. The integration of such energy savings. As it is evident from our survey, this is an
frameworks and approaches by smart grid mobile app sys- aspect that the majority of the previous apps for residential
tems has not been fully studied, thus it is important for future energy efficiency have not touched upon, and the ones that
efforts to adopt such big data processing mechanisms in order do so, do not take full advantage of the recent advancements
to provide richer services to end-users. Such services could in machine learning (e.g. deep learning) and recommender
include more complex and faster data analytics pipelines for systems.
better energy insights, as well as visualization with higher In our future work, we will construct detailed energy pro-
granularity and overall greater system reliability and fault files for each user/household (as described in section V)
tolerance. utilizing energy consumption data, socio-demographic data
from questionnaires, and other IoT measurements (e.g. cli-
C. APP SUSTAINABILITY
mate conditions, occupancy, and illuminance). Then, super-
It is critical for smart grid mobile apps to achieve long-term vised and unsupervised machine learning algorithms are
consumer engagement and sustainability in terms of energy going to be used to extract highly nonlinear correlations
efficient behavior, otherwise the app’s impact on consump- between consumer profile features and behavioral charac-
tion reduction and load shaping will be short-term and teristics (such as the clustering solution depicted in Fig. 6).
insignificant. Only a few of the aforementioned studied Such insights will be utilized to engage consumers based on
mobile apps included a thorough sustainability plan in order their behavioral characteristics in a targeted manner using
to retain user engagement throughout time. This can be advanced RecSys approaches, thus achieving greater energy
accomplished by continuous development of new content and savings compared to the state-of-the-art. Furthermore, our
features, in order to avoid user fatigue and eventual dropout. mobile app will enable flexibility from the consumer per-
The gamification content should be updated over time and spective, providing appropriate incentives for users who are
new rewards should be provided to consumers. Furthermore, involved in the flexibility procedure, thus further decreasing
new visual insights should be presented, so that the users the overall consumption. The app will be validated through
always learn something new about their energy consumption a pilot with consumers from multiple companies (i.e. energy
patterns and characteristics. In addition, mobile apps should retailers and IoT providers), to verify its generality and inter-
adapt to the behavior and characteristics of each user, in order operability capabilities.
to provide them with personalized content, thus keeping them
engaged throughout time. VIII. CONCLUSION
A thorough survey of existing mobile applications for inform-
D. SECURITY AND PRIVACY ing, educating, and incentivizing smart grid consumers was
One more aspect that we think should be better addressed by presented. Approaches for user engagement and motiva-
such applications is security since household energy data are tion were briefly discussed, with a focus on gamification,
sensitive, and in the wrong hands can raise serious issues. For as well as state-of-the-art mobile app system architecture
example intruders can derive the working hours of occupants approaches. Moreover, the role and importance of machine
from their energy data and break into their house. Advanced learning integration by such apps, for more effective user and
anonymization technologies, Distributed Identifiers (DIDs), building modeling was discussed, analyzing the application
Distributed Ledger Technologies (or blockchains), smart con- of techniques like classification and clustering. Furthermore,
tracts, and Verifiable Credentials (VCs) can help improve a presentation and comparison of existing mobile apps for
219652 VOLUME 8, 2020
energy efficiency was carried out, both for research projects [13] M. S. Mahdavinejad, M. Rezvan, M. Barekatain, P. Adibi, P. Barnaghi,
and industrial products, regarding features like data visu- and A. P. Sheth, ‘‘Machine learning for Internet of Things data analysis:
A survey,’’ Digit. Commun. Netw., vol. 4, no. 3, pp. 161–175, Aug. 2018.
alization, gamification, flexibility, user profiling, feedback [14] Bluetooth. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
mechanisms, tips or notifications, social media, machine bluetooth.com/
learning, and recommender systems integration. This paper [15] Wi-Fi. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
wi-fi.org/
highlights the positive impact that such mobile apps can have [16] Zigbee. Accessed: Nov. 16, 2020. [Online]. Available: https://
on electricity consumer engagement and overall smart grid zigbeealliance.org/
[17] Z-Wave. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
efficiency through consumption reduction and demonstration z-wave.com/
of improved flexibility or load shifting, and points out the [18] MQTT. Accessed: Nov. 16, 2020. [Online]. Available: http://mqtt.org/
features that are missing from existing applications in order [19] MongoDB. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
mongodb.com/
to achieve higher levels of energy savings, more ecological [20] InfluxDB. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
energy consumption profiles, and overall lowering environ- influxdata.com/
mental impact. [21] Hadoop File System. Accessed: Nov. 16, 2020. [Online]. Available:
https://hadoop.apache.org/docs/r1.2.1/hdfs_design.html
ACKNOWLEDGMENT [22] Kafka. Accessed: Nov. 16, 2020. [Online]. Available: https://kafka.
apache.org/
The authors would like to thank Ms. E. Faltaka for helping [23] Spark. Accessed: Nov. 16, 2020. [Online]. Available: https://spark.
with an initial exploration of available mobile apps. The sole apache.org/
responsibility for the content lies with the authors. It does [24] TensorFlow. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
tensorflow.org/
not necessarily reflect the opinion of DG CNECT or the [25] Keras. Accessed: Nov. 16, 2020. [Online]. Available: https://keras.io/
European Commission (EC). DG CNECT or the EC are not [26] PyTorch. Accessed: Nov. 16, 2020. [Online]. Available: https://
responsible for any use that may be made of the information pytorch.org/
[27] M. Karaliopoulos, I. Koutsopoulos, and L. Spiliopoulos, ‘‘Optimal
contained herein. user choice engineering in mobile crowdsensing with bounded rational
users,’’ in Proc. INFOCOM IEEE Conf. Comput. Commun., Apr. 2019,
REFERENCES pp. 1054–1062.
[1] X. Fang, S. Misra, G. Xue, and D. Yang, ‘‘Smart grid—The new and [28] M. Karaliopoulos, I. Koutsopoulos, and M. Titsias, ‘‘First learn then earn:
improved power grid: A survey,’’ IEEE Commun. Surveys Tuts., vol. 14, Optimizing mobile crowdsensing campaigns through data-driven user
no. 4, pp. 944–980, 4th Quart., 2012. profiling,’’ in Proc. 17th ACM Int. Symp. Mobile Ad Hoc Netw. Comput.
[2] O. Ellabban and H. Abu-Rub, ‘‘Smart grid customers’ acceptance (MobiHoc), 2016, pp. 271–280.
and engagement: An overview,’’ Renew. Sustain. Energy Rev., vol. 65, [29] T. AlSkaif, I. Lampropoulos, M. van den Broek, and W. van Sark,
pp. 1285–1298, Nov. 2016. ‘‘Gamification-based framework for engagement of residential customers
[3] C. Fan, F. Xiao, and Y. Zhao, ‘‘A short-term building cooling load pre- in energy applications,’’ Energy Res. Social Sci., vol. 44, pp. 187–195,
diction method using deep learning algorithms,’’ Appl. Energy, vol. 195, Oct. 2018.
pp. 222–233, Jun. 2017. [30] C. Rottondi and G. Verticale, ‘‘A privacy-friendly gaming framework in
[4] T. Abergel, B. Dean, and J. Dulac, ‘‘UN environment and international smart electricity and water grids,’’ IEEE Access, vol. 5, pp. 14221–14233,
energy agency (2017): Towards a zero-emission, efficient, and resilient 2017.
buildings and construction sector,’’ Global Status, United Nations Envi- [31] T. G. Papaioannou, V. Hatzi, and I. Koutsopoulos, ‘‘Optimal design of
ron. Programme, Paris, France, Tech. Rep. 978-92-807-3686-1, 2017. serious games for consumer engagement in the smart grid,’’ IEEE Trans.
[5] M. Alaa, A. A. Zaidan, B. B. Zaidan, M. Talal, and M. L. M. Kiah, Smart Grid, vol. 9, no. 2, pp. 1241–1249, Mar. 2018.
‘‘A review of smart home applications based on Internet of Things,’’ [32] P. Micholia, M. Karaliopoulos, and I. Koutsopoulos, ‘‘Mobile crowd-
J. Netw. Comput. Appl., vol. 97, pp. 48–65, Nov. 2017. sensing incentives under participation uncertainty,’’ in Proc. 3rd ACM
[6] D. Johnson, E. Horton, R. Mulcahy, and M. Foth, ‘‘Gamification and Workshop Mobile Sens., Comput. Commun. (MSCC), 2016, pp. 29–34.
serious games within the domain of domestic energy consumption: A sys- [33] PEAKapp D2.2. Accessed: Nov. 16, 2020. [Online]. Available:
tematic review,’’ Renew. Sustain. Energy Rev., vol. 73, pp. 249–264, http://www.peakapp.eu/wp-content/uploads/2018/02/D2.2_The_Smart_
Jun. 2017. phonetablet_app_GP_final.pdf
[7] L. Morganti, F. Pallavicini, E. Cadel, A. Candelieri, F. Archetti, and [34] A. Gustafsson, M. Bång, and M. Svahn, ‘‘Power explorer: A casual game
F. Mantovani, ‘‘Gaming for Earth: Serious games and gamification style for encouraging long term behavior change among teenagers,’’ in
to engage consumers in pro-environmental behaviours for energy effi- Proc. Int. Conf. Adv. Comput. Enterntainment Technol. (ACE), 2009,
ciency,’’ Energy Res. Social Sci., vol. 29, pp. 95–102, Jul. 2017. pp. 182–189.
[8] P. Fraternali, F. Cellina, S. L. Herrera Gonzales, M. Melenhorst, J. Novak, [35] L. Gamberini, C. Björkskog, M. Salo, P. Aman, N. Corradi, L. Zamboni,
C. Pasini, C. Rottondi, and A. E. Rizzoli, ‘‘Visualizing and gamifying M. Perotti, C. Cadenazzi, S. Mandressi, G. Jacucci, G. Tusa, and
consumption data for resource saving: Challenges, lessons learnt and a A. Spagnolli, ‘‘Saving is fun: Designing a persuasive game for power con-
research agenda for the future,’’ Energy Informat., vol. 2, no. S1, pp. 1–13, servation,’’ in Proc. 8th Int. Conf. Adv. Comput. Entertainment Technol.
Sep. 2019. (ACE), 2011, pp. 1–7.
[9] L. Bastida, J. J. Cohen, A. Kollmann, A. Moya, and J. Reichl, ‘‘Exploring [36] B. Xia, H. Ming, K.-Y. Lee, Y. Li, Y. Zhou, S. Bansal, S. Shakkottai,
the role of ICT on household behavioural energy efficiency to mitigate and L. Xie, ‘‘EnergyCoupon: A case study on incentive-based demand
global warming,’’ Renew. Sustain. Energy Rev., vol. 103, pp. 455–462, response in smart grid,’’ in Proc. 8th Int. Conf. Future Energy Syst.,
Apr. 2019. May 2017, pp. 80–90.
[10] A. Moreno-Munoz, F. J. Bellido-Outeirino, P. Siano, and [37] P. Petkov, F. Köbler, M. Foth, and H. Krcmar, ‘‘Motivating domestic
M. A. Gomez-Nieto, ‘‘Mobile social media for smart grids customer energy conservation through comparative, community-based feedback in
engagement: Emerging trends and challenges,’’ Renew. Sustain. Energy mobile and social media,’’ in Proc. 5th Int. Conf. Communities Technol.
Rev., vol. 53, pp. 1611–1616, Jan. 2016. (C&T), 2011, pp. 21–30.
[11] D. Zhang, X. Han, and C. Deng, ‘‘Review on the research and practice of [38] P. Petkov, F. Köbler, M. Foth, R. Medland, and H. Krcmar, ‘‘Engaging
deep learning and reinforcement learning in smart grids,’’ CSEE J. Power energy saving through motivation-specific social comparison,’’ in Proc.
Energy Syst., vol. 4, no. 3, pp. 362–370, Sep. 2018. Annu. Conf. Extended Abstr. Human Factors Comput. Syst. (CHI EA),
[12] A. Gharaibeh, M. A. Salahuddin, S. J. Hussini, A. Khreishah, I. Khalil, 2011, pp. 1945–1950.
M. Guizani, and A. Al-Fuqaha, ‘‘Smart cities: A survey on data manage- [39] Myearth App. Accessed: Nov. 16, 2020. [Online]. Available: https://
ment, security, and enabling technologies,’’ IEEE Commun. Surveys Tuts., play.google.com/store/apps/details?id=com.ionicframework.
vol. 19, no. 4, pp. 2456–2501, 4th Quart., 2017. myearthapp496614
VOLUME 8, 2020 219653

[40] M. Peham, G. Breitfuss, and R. Michalczuk, ‘‘The ‘ecoGator’ app: [63] Tribe Project. Accessed: Nov. 16, 2020. [Online]. Available: http://tribe-
Gamification for enhanced energy efficiency in europe,’’ in Proc. 2nd h2020.eu/
Int. Conf. Technol. Ecosyst. Enhancing Multiculturality (TEEM), 2014, [64] Greenplay Project. Accessed: Nov. 16, 2020. [Online]. Available: http://
pp. 179–183. www.greenplay-project.eu/
[41] PEAKapp Project. Accessed: Nov. 16, 2020. [Online]. Available: [65] InterConnect. Accessed: Nov. 16, 2020. [Online]. Available: https://
http://www.peakapp.eu/ interconnectproject.eu/
[42] V. Azarova, D. Engel, C. Ferner, A. Kollmann, and J. Reichl, ‘‘Exploring [66] Smappee App. Accessed: Nov. 16, 2020. [Online]. Available:
the impact of network tariffs on household electricity expenditures using https://www.smappee.com/
load profiles and socio-economic characteristics,’’ Nature Energy, vol. 3, [67] TED App. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
no. 4, pp. 317–325, Apr. 2018. theenergydetective.com/
[43] ChArGED D3.6. Accessed: Nov. 16, 2020. [Online]. Available: http:// [68] Neurio App. Accessed: Nov. 16, 2020. [Online]. Available: https://www.
www.charged-project.eu/sites/default/files/publications/ChArGED_ neur.io/
D3.6v1.0withInstr.pdf [69] Joulebug App. Accessed: Nov. 16, 2020. [Online]. Available:
[44] ChArGED Project. Accessed: Nov. 16, 2020. [Online]. Available: https://joulebug.com/
http://www.charged-project.eu/ [70] Constellation Connect App. Accessed: Nov. 16, 2020. [Online]. Avail-
[45] T. Papaioannou, N. Dimitriou, K. Vasilakis, A. Schoofs, M. Nikiforakis, able: https://constellationconnect.com/
F. Pursche, N. Deliyski, A. Taha, D. Kotsopoulos, C. Bardaki, [71] Energy Cost Calculator App. Accessed: Nov. 16, 2020. [Online].
S. Kotsilitis, and A. Garbi, ‘‘An IoT-based gamified approach for reducing Available: https://www.rapidtables.com/calc/electric/energy-cost-
Occupants’ energy wastage in public buildings,’’ Sensors, vol. 18, no. 2, calculator.html
p. 537, Feb. 2018. [72] Energy Tracker App. Accessed: Nov. 16, 2020. [Online]. Available:
[46] GAIA Project. Accessed: Nov. 16, 2020. [Online]. Available: http://www. https://www.npower.com/campaigns/online-account/energy-tracker/
gaia-project.eu/ [73] Kill-Ur-Watts App. Accessed: Nov. 16, 2020. [Online]. Available:
[47] GreenSoul Project. Accessed: Nov. 16, 2020. [Online]. Available: https://openei.org/wiki/Kill-Ur-Watts
[74] Y. Wang, Q. Chen, T. Hong, and C. Kang, ‘‘Review of smart meter data
http://www.greensoul-h2020.eu/
[48] enCOMPASS Project. Accessed: Nov. 16, 2020. [Online]. Available: analytics: Applications, methodologies, and challenges,’’ IEEE Trans.
http://www.encompass-project.eu/ Smart Grid, vol. 10, no. 3, pp. 3125–3148, May 2019.
[75] H. Liang, J. Ma, R. Sun, and Y. Du, ‘‘A data-driven approach for target-
[49] P. Fraternali, S. Herrera, J. Novak, M. Melenhorst, D. Tzovaras,
ing residential customers for energy efficiency programs,’’ IEEE Trans.
S. Krinidis, A. E. Rizzoli, C. Rottondi, and F. Cellina, ‘‘EnCOMPASS—
Smart Grid, vol. 11, no. 2, pp. 1229–1238, Mar. 2020.
An integrative approach to behavioural change for energy saving,’’ in
[76] Pecan Street. Accessed: Nov. 16, 2020. [Online]. Available:
Proc. Global Internet Things Summit (GIoTS), Jun. 2017, pp. 1–6.
https://www.pecanstreet.org/dataport/
[50] P. Fraternali, F. Cellina, S. Herrera, S. Krinidis, C. Pasini, A. E. Rizzoli,
[77] A. K. Zarabie, S. Lashkarbolooki, S. Das, K. Jhala, and A. Pahwa,
C. Rottondi, and D. Tzovaras, ‘‘A socio-technical system based on gamifi-
‘‘Load profile based electricity consumer clustering using affinity propa-
cation towards energy savings,’’ in Proc. IEEE Int. Conf. Pervas. Comput.
gation,’’ in Proc. IEEE Int. Conf. Electro Inf. Technol. (EIT), May 2019,
Commun. Workshops (PerCom Workshops), Mar. 2018, pp. 59–64.
pp. 474–478.
[51] T. Vafeiadis, S. Zikos, G. Stavropoulos, D. Ioannidis, S. Krinidis,
[78] B. Dai, R. Wang, K. Zhu, J. Hao, and P. Wang, ‘‘A demand response
D. Tzovaras, and K. Moustakas, ‘‘Machine learning based occupancy
scheme in smart grid with clustering of residential customers,’’ in Proc.
detection via the use of smart meters,’’ in Proc. Int. Symp. Comput. Sci.
IEEE Int. Conf. Commun., Control, Comput. Technol. Smart Grids
Intell. Controls (ISCSIC), Oct. 2017, pp. 6–12.
(SmartGridComm), Oct. 2019, pp. 1–6.
[52] C. Rottondi, M. Derboni, D. Piga, and A. E. Rizzoli, ‘‘An optimisation- [79] J. Kwac, J. Flora, and R. Rajagopal, ‘‘Household energy consumption
based energy disaggregation algorithm for low frequency smart meter segmentation using hourly data,’’ IEEE Trans. Smart Grid, vol. 5, no. 1,
data,’’ Energy Informat., vol. 2, no. S1, p. 13, Sep. 2019. pp. 420–430, Jan. 2014.
[53] A. Vafeiadis, T. Vafeiadis, S. Zikos, S. Krinidis, K. Votis, D. Giakoumis, [80] S. Haben, C. Singleton, and P. Grindrod, ‘‘Analysis and clustering of
D. Ioannidis, D. Tzovaras, L. Chen, and R. Hamzaoui, ‘‘Energy-based residential customers energy behavioral demand using smart meter data,’’
decision engine for household human activity recognition,’’ in Proc. IEEE IEEE Trans. Smart Grid, vol. 7, no. 1, pp. 136–144, Jan. 2016.
Int. Conf. Pervas. Comput. Commun. Workshops (PerCom Workshops), [81] J. D. Rhodes, W. J. Cole, C. R. Upshaw, T. F. Edgar, and M. E. Webber,
Mar. 2018, pp. 704–709. ‘‘Clustering analysis of residential electricity demand profiles,’’ Appl.
[54] S. Albertarelli, P. Fraternali, J. Novak, A.-E. Rizzoli, and C. Rottondi, Energy, vol. 135, pp. 461–471, Dec. 2014.
‘‘Drop and FUNERGY: Two gamified learning projects for water and [82] A. C. Varghese, P. V. G. Kumar, and S. A. Khaparde, ‘‘Smart grid con-
energy conservation,’’ in Proc. 11th Eur. Conf. Games Based Learn. sumer behavioral model using machine learning,’’ in Proc. IEEE Innov.
(ECGBL), 2017, pp. 935–938. Smart Grid Technol. Asia (ISGT Asia), May 2018, pp. 734–739.
[55] P. Fraternali and S. L. H. Gonzalez, ‘‘An augmented reality game for [83] L. Breiman, J. Friedman, C. J. Stone, and R. Olshen, Classification
energy awareness,’’ in Proc. Int. Conf. Comput. Vis. Syst. Thessaloniki, Regression Trees. Boca Raton, FL, USA: CRC Press, 1984.
Greece: Springer, 2019, pp. 629–638. [84] J. R. Quinlan, ‘‘Induction of decision trees,’’ Mach. Learn., vol. 1, no. 1,
[56] enCOMPASS D5.4. Accessed: Nov. 16, 2020. [Online]. Available: http:// pp. 81–106, 1986.
www.encompass-project.eu/wp-content/uploads/2020/04/enCOMPASS_ [85] P. J. Werbos, ‘‘Backpropagation through time: What it does and how to
D5.4.V.1.0.pdf do it,’’ Proc. IEEE, vol. 78, no. 10, pp. 1550–1560, 1990.
[57] EnerGAware Project. Accessed: Nov. 16, 2020. [Online]. Available: [86] G. Hinton. RMSprop: Divide the Gradient by a Running Average of
http://www.energaware.eu/ Its Recent Magnitude. Accessed: Nov. 16, 2020. [Online]. Available:
[58] ENTROPY Project. Accessed: Nov. 16, 2020. [Online]. Available: http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf
https://cordis.europa.eu/project/id/649849 [87] J. Duchi, E. Hazan, and Y. Singer, ‘‘Adaptive subgradient methods
[59] E. Fotopoulou, A. Zafeiropoulos, F. Terroso-Sáenz, U. Şimsȩk, for online learning and stochastic optimization,’’ J. Mach. Learn. Res.,
A. González-Vidal, G. Tsiolis, P. Gouvas, P. Liapis, A. Fensel, and vol. 12, pp. 2121–2159, Jul. 2011.
A. Skarmeta, ‘‘Providing personalized energy management and aware- [88] D. P. Kingma and J. L. Ba, ‘‘Adam: A method for stochastic optimiza-
ness services for energy efficiency in smart buildings,’’ Sensors, vol. 17, tion,’’ in Proc. 3rd Int. Conf. Learn. Represent., 2015. [Online]. Available:
no. 9, p. 2054, Sep. 2017. https://arxiv.org/abs/1412.6980
[60] Social Power Project. Accessed: Nov. 16, 2020. [Online]. Available: [89] I. Koutsopoulos, T. G. Papaioannou, and V. Hatzi, ‘‘Modeling and opti-
http://www.socialpower.ch/ mization of the smart grid ecosystem,’’ Found. Trends Netw., vol. 10,
[61] D. Wemyss, F. Cellina, E. Lobsiger-Kägi, V. de Luca, and R. Castri, nos. 2–3, pp. 115–316, 2016.
‘‘Does it last? Long-term impacts of an app-based behavior change [90] I. C. Konstantakopoulos, A. R. Barkan, S. He, T. Veeravalli, H. Liu, and
intervention on household electricity savings in Switzerland,’’ Energy C. Spanos, ‘‘A deep learning and gamification approach to improving
Res. Social Sci., vol. 47, pp. 16–27, Jan. 2019. human-building interaction and energy efficiency in smart infrastruc-
[62] D. Wemyss, R. Castri, F. Cellina, V. De Luca, E. Lobsiger-Kägi, and ture,’’ Appl. Energy, vol. 237, pp. 810–821, Mar. 2019.
V. Carabias, ‘‘Examining community-level collaborative vs. Competitive [91] K. Hopf, M. Sodenkamp, and T. Staake, ‘‘Enhancing energy efficiency in
approaches to enhance household electricity-saving behavior,’’ Energy the residential sector with smart meter data analytics,’’ Electron. Markets,
Efficiency, vol. 11, no. 8, pp. 2057–2075, Dec. 2018. vol. 28, no. 4, pp. 453–473, Nov. 2018.
219654 VOLUME 8, 2020

[92] K. Hopf, M. Sodenkamp, I. Kozlovkiy, and T. Staake, ‘‘Feature extraction SPIROS CHADOULOS received the B.Sc. degree
and filtering for household classification based on smart electricity meter in management science and technology and the
data,’’ Comput. Sci. Res. Develop., vol. 31, no. 3, pp. 141–148, Aug. 2016. M.Sc. degree in computer science from the Athens
[93] C. Beckel, L. Sadamori, and S. Santini, ‘‘Automatic socio-economic University of Economics and Business (AUEB),
classification of households using electricity consumption data,’’ in Proc. Greece, in 2018 and 2019, respectively, where he
the 4th Int. Conf. Future Energy Syst. e-Energy, 2013, pp. 75–86. is currently pursuing the Ph.D. degree in com-
[94] C. Beckel, L. Sadamori, T. Staake, and S. Santini, ‘‘Revealing household
characteristics from smart meter data,’’ Energy, vol. 78, pp. 397–410,
puter science. He is also a member of the Mobile
Dec. 2014. Multimedia Laboratory (MMLab) and works as a
[95] Y. Wang, Q. Chen, D. Gan, J. Yang, D. S. Kirschen, and C. Kang, ‘‘Deep Software Engineer and a Researcher for the Inter-
learning-based socio-demographic information identification from smart connect Project (Horizon 2020 Programme). His
meter data,’’ IEEE Trans. Smart Grid, vol. 10, no. 3, pp. 2593–2602, research interests include smart energy grid optimization, machine/deep
May 2019. learning applications for energy efficiency, energy generation and demand
[96] W.-J. Tang, X.-L. Lee, H. Wang, and H.-T. Yang, ‘‘Leveraging socioeco- forecasting, and electricity consumer engagement.
nomic information and deep learning for residential load pattern predic-
tion,’’ in Proc. IEEE PES Innov. Smart Grid Technol. Eur. (ISGT-Eur.), IORDANIS KOUTSOPOULOS (Senior Member,
Sep. 2019, pp. 1–5. IEEE) received the diploma degree in electri-
[97] Z. Jiang, R. Lin, and F. Yang, ‘‘A hybrid machine learning model for
cal and computer engineering from the National
electricity consumer categorization using smart meter data,’’ Energies,
Technical University of Athens (NTUA), Greece,
vol. 11, no. 9, p. 2235, Aug. 2018.
[98] J. L. Viegas, S. M. Vieira, R. Melício, V. M. F. Mendes, and in 1997, and the M.Sc. and Ph.D. degrees in
J. M. C. Sousa, ‘‘Classification of new electricity customers based on sur- electrical and computer engineering from the
veys and smart metering data,’’ Energy, vol. 107, pp. 804–817, Jul. 2016. University of Maryland, College Park (UMCP),
[99] F. McLoughlin, A. Duffy, and M. Conlon, ‘‘A clustering approach to College Park, MD, USA, in 1999 and 2002,
domestic electricity load profile characterisation using smart metering respectively.
data,’’ Appl. Energy, vol. 141, pp. 190–199, Mar. 2015. He was a Lecturer from 2005 to 2010 and an
[100] Y. Koren, R. Bell, and C. Volinsky, ‘‘Matrix factorization techniques Assistant Professor from 2010 to 2013 with the Department of Electrical and
for recommender systems,’’ Computer, vol. 42, no. 8, pp. 30–37, Computer Engineering, University of Thessaly, and an Assistant Professor
Aug. 2009. with the Athens University of Economics and Business (AUEB), Athens,
[101] F. Luo, G. Ranzi, X. Wang, and Z. Y. Dong, ‘‘Service recommendation
Greece, from 2013 to 2016. He has been an Associate Professor with the
in smart grid: Vision, technologies, and applications,’’ in Proc. 9th Int.
Conf. Service Sci. (ICSS), 2016, pp. 31–38. Department of Informatics, AUEB, since 2016. His research interests include
[102] J. E. Fischer, S. Reece, E. Costanza, T. Rodden, N. R. Jennings, the general area of control and optimization and on applications of machine
S. D. Ramchurn, M. Osborne, O. Parson, T. D. Huynh, M. Alam, learning, with application areas, such as mobile crowdsensing, wireless net-
N. Pantidi, S. Moran, and K. Bachour, ‘‘Recommending energy tariffs works, social networks, recommender systems, smart energy grid, and cloud
and load shifting based on smart household usage profiling,’’ in Proc. Int. computing systems. He received the Single-Investigator European Research
Conf. Intell. User Interfaces, 2013, pp. 383–394. Council (ERC) competition Runner-Up Award for the Project RECITAL:
[103] F. Luo, G. Ranzi, X. Wang, and Z. Y. Dong, ‘‘Social information Resource Management for Self-Coordinated Autonomic Wireless Networks
filtering-based electricity retail plan recommender system for smart from 2012 to 2015, a Marie Curie Fellowship from 2005 to 2007, a Fulbright
grid end users,’’ IEEE Trans. Smart Grid, vol. 10, no. 1, pp. 95–104, Scholarship from 1997 to 2003, and three best paper awards for research on
Jan. 2019. online advertising, network economics, and network experimentation.
[104] M. Zehnder, H. Wache, H.-F. Witschel, D. Zanatta, and M. Rodriguez,
‘‘Energy saving in smart homes based on consumer behavior: A case
study,’’ in Proc. IEEE 1st Int. Smart Cities Conf. (ISC2), Oct. 2015, GEORGE C. POLYZOS (Member, IEEE) received
pp. 1–6. the diploma degree in electrical engineering from
[105] D. Schweizer, M. Zehnder, H. Wache, H.-F. Witschel, D. Zanatta, and
the National Technical University in Athens,
M. Rodriguez, ‘‘Using consumer behavior data to reduce energy con-
sumption in smart homes: Applying machine learning to save energy Greece, and the M.A.Sc. degree in electrical engi-
without lowering comfort of inhabitants,’’ in Proc. IEEE 14th Int. Conf. neering and the Ph.D. degree in computer science
Mach. Learn. Appl. (ICMLA), Dec. 2015, pp. 1123–1129. from the University of Toronto, Canada.
[106] I. Micheel, J. Novak, P. Fraternali, G. Baroffio, A. F. Castelletti, and He was a Professor of computer science and
A. Rizzoli, ‘‘Visualizing and gamifying water & energy consumption engineering with the University of California, San
for behavior change,’’ in Proc. Fostering Smart Energy Appl. Workshop Diego, where he was the Co-Director of the Com-
(FSEA) Interact. Bamberg, Germany: Springer, 2015, pp. 1–4. puter Systems Laboratory, a member of the Steer-
[107] B. Hidasi and A. Karatzoglou, ‘‘Recurrent neural networks with top-k ing Committee of the Center for Wireless Communications, and a Senior
gains for session-based recommendations,’’ in Proc. 27th ACM Int. Conf. Fellow of the San Diego Supercomputer Center. He is leading the Mobile
Inf. Knowl. Manage., Oct. 2018, pp. 843–852. Multimedia Laboratory (MMlab), Athens University of Economics and
[108] L. Daniele, F. den Hartog, and J. Roes, ‘‘Created in close interaction with
Business, where he is currently a Professor of computer science and the
the industry: The smart appliances reference (SAREF) ontology,’’ in Proc.
Int. Workshop Formal Ontologies Meet Ind. Berlin, Germany: Springer,
Director of the M.Sc. Program in computer science. Under his leadership
2015, pp. 100–112. the MMlab has participated in a series of research projects funded by the EU
[109] Universal Smart Energy Framework (USEF). Accessed: Nov. 16, 2020. (PSIRP, PURSUIT, and POINT), the ESA (φSAT), and Greece (I-CAN) that
[Online]. Available: https://www.usef.energy/ co-developed Publish-Subscribe Internetworking, an Information-Centric
[110] Open Charge Point Interface protocol (OCPI). Accessed: Nov. 16, 2020. Networking architecture. PURSUIT received the Future Internet Award
[Online]. Available: https://evroaming.org/ocpi-background/ in 2013. His current research interests include the Internet of Things,
[111] IEC 61970 Standards. Accessed: Nov. 16, 2020. [Online]. Available: smart grid, security and privacy, Internet architecture and protocols, mobile
https://en.wikipedia.org/wiki/IEC_61970 multimedia communications, wireless networks, and systems performance
[112] N. Fotiou, I. Pittaras, V. A. Siris, S. Voulgaris, and G. C. Polyzos, ‘‘Secure
evaluation. He has served on journal editorial boards, as a special issue
IoT access at scale using blockchains and smart contracts,’’ in Proc. IEEE
guest editor, and on committees of many conferences and workshops. He has
20th Int. Symp. A World Wireless, Mobile Multimedia Netw. (WoWMoM),
Jun. 2019, pp. 1–6.
chaired the Steering Committee of the ACM SIGCOMM conference on
[113] N. Fotiou, I. Pittaras, V. A. Siris, S. Voulgaris, and G. C. Polyzos, Information-Centric Networking and is also on the Steering Committee of
‘‘OAuth 2.0 authorization using blockchain-based tokens,’’ in the Wireless and Mobile Networking Conference (IFIP TC6 WG 6.8) and on
Proc. Workshop Decentralized IoT Syst. Secur. (DISS) Conjunct the editorial boards of the IEEE TRANSACTIONS ON MOBILE COMPUTING and the
Netw. Distrib. Syst. Secur. Symp. (NDSS), 2020. [Online]. Available: Journal on Reliable Intelligent Environments.
https://arxiv.org/abs/2001.10461
VOLUME 8, 2020 219655

Mobile Apps Meet The Smart Energy Grid: A Survey On Consumer Engagement and Machine Learning Applications

Uploaded by

Copyright:

Available Formats

Mobile Apps Meet The Smart Energy Grid: A Survey On Consumer Engagement and Machine Learning Applications

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mobile Apps Meet The Smart Energy Grid: A Survey On Consumer Engagement and Machine Learning Applications

Uploaded by

Copyright:

Available Formats

Received November 16, 2020, accepted November 28, 2020, date of publication December 7, 2020,

date of current version December 17, 2020.

Mobile Apps Meet the Smart Energy Grid:

I. INTRODUCTION operation, as well as new and innovative services and prod-

VOLUME 8, 2020 219633

219634 VOLUME 8, 2020

VOLUME 8, 2020 219635

219636 VOLUME 8, 2020

TABLE 1. Smart grid mobile app features comparison.

VOLUME 8, 2020 219637

FIGURE 2. PEAKapp dashboard screens [33].

219638 VOLUME 8, 2020

FIGURE 3. ChArGED mobile app screenshots [43].

VOLUME 8, 2020 219639

achieving certain goals. The app provides energy consump-

VOLUME 8, 2020 219641

new samples of that cluster. Convergence is achieved when

219642 VOLUME 8, 2020

reduction technique of the original consumption data, before

219644 VOLUME 8, 2020

to provide the most appropriate incentives to each user, and

VOLUME 8, 2020 219645

219646 VOLUME 8, 2020

VOLUME 8, 2020 219647

TABLE 2. Smart grid mobile app most common gamification features.

219648 VOLUME 8, 2020

VOLUME 8, 2020 219649

219650 VOLUME 8, 2020

VOLUME 8, 2020 219653

219654 VOLUME 8, 2020

VOLUME 8, 2020 219655

You might also like