1 s2.0 S2666792420300068 Main

Advances in Applied Energy 1 (2021) 100006
Contents lists available at ScienceDirect
Advances in Applied Energy

journal homepage: www.elsevier.com/locate/adapen
Smart grid encounters edge computing: opportunities and applications☆

Cheng Feng a, Yi Wang b, Qixin Chen a,∗, Yi Ding c, Goran Strbac d, Chongqing Kang a
a
State Key Laboratory of Power Systems, Department of Electrical Engineering, Tsinghua University, Beijing 100084, China
b
Power Systems Laboratory, ETH Zürich, Zürich 8092, Switzerland
c
College of Electrical Engineering, Zhejiang University, Hangzhou 310027, China
d
Department of Electrical and Electronic Engineering, Imperial College London, London SW7 2AZ, UK
a r t i c l e i n f o a b s t r a c t
Keywords: Smart grids (SGs) are reforming towards utilizing massive data for operations and services. During this reform,
Smart grid the information and communication technologies (ICTs) play a critical role, especially for the computing model,
Edge computing which determines how data analytics in SG can be executed. Edge computing (EC), a novel computing paradigm
ICT solutions
innovation, has high potential to help with the digitization of SG. This paper seeks to provide a comprehensive
Data analytics
review of interdisciplinary research on EC applications in SG. First, an in-depth analysis of EC, including the
Sustainable computing
definition, architectures, characteristics, and key enabling technologies is conducted from the perspective of SG
application requirements. Afterward, this paper systematically explores application scenarios of EC in SG based
on its characteristics. Beyond that, the synergistic effect of the integration of EC and SG is also assessed: SG
not only benefits from EC but also supports EC to become sustainable. Challenges and open questions are then
investigated.
1. Introduction prove system scalability and availability, and protect data security and
privacy [3].
Massive volumes of data are generated every second in Smart grids Due to these advantages, EC has drawn considerable attention
(SGs). Advanced data analytic algorithms are required to transform the from academia and industry. In 2014, the European Telecommunica-
data into information and knowledge, which can be further utilized for tions Standards Institute (ETSI) established the Mobile Edge Computing
SG operations and services. Generally, these data analytics depend on (MEC) Industry Specification Group. In 2015, ARM, Cisco, Dell, Intel,
information & communication technologies (ICTs), which perform a crit- Microsoft, and Princeton University found the OpenFog Consortium.
ical role in data collection, transmission, and processing [1]. ACM and IEEE have jointly organized the ACM/IEEE Symposium on
Among the major functions of ICTs, computing determines how Edge Computing (SEC) since 2016 [4]. To date, several ICT companies
SG data analytics are executed and thus, it becomes the foundation have already launched the research & development of commercial EC
for SG operations and services. Centralized cloud computing prevails products. Conventional industrial electrical equipment providers in the
as a feasible solution for the SG computing paradigm [2]. In cloud power industry also announce that they will invest in EC technology on
computing, geo-distributed devices and equipment are connected to the development of related products [5].
cloud data centers, making centralized decisions and issuing control This promising future enlightens the potential of implementing EC
orders. Nevertheless, it suffers from several weaknesses, such as lim- in SG. Computing performed by ICT in SG is industrially oriented (for
ited bandwidth resources, heterogeneous environments, and privacy power grid operations such as controlling, dispatching, scheduling) and
concerns. Against this background, another solution, edge computing service-oriented (for energy services such as recommendation, negoti-
(EC), pushes the frontier of computation applications away from cen- ation, transactions). The application scenarios of EC in SG should be
tralized nodes to the communication network’s extremes. EC lever- explicitly specified. Thus, an insightful investigation on how to achieve
ages computing resources closer to sensors and users to carry out the integration of SG and EC is warranted. This paper introduces the
data analytics. It gains advantages for its ability to effectively re- interdisciplinary values to integrate EC and SG from the prospect of SG
duce system delay, lighten the burden of cloud computing centers, im- application requirements.
☆
This work was supported by Major Smart Grid Joint Project of National Natural Science Foundation of China and State Grid under Grant U2066205.
∗
Corresponding author.
E-mail address: qxchen@tsinghua.edu.cn (Q. Chen).
https://doi.org/10.1016/j.adapen.2020.100006
Received 16 September 2020; Received in revised form 10 December 2020; Accepted 17 December 2020
Available online 23 December 2020
2666-7924/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
C. Feng, Y. Wang, Q. Chen et al. Advances in Applied Energy 1 (2021) 100006
Table 1
The taxonomy of typical computing paradigms. The computing paradigm from the top the bottom successively corresponds
to the the computing paradigm (a)–(e) in Fig. 1.
Paradigms Characteristics
Local/ Decentralized Each computing resource computes purely locally, and there is no communication
among computation resources.
Distributed Each computing resource equally communicates with its neighbors and depends on
the others to diverge.
Hierarchical/ Coordinated Each computing resource computes and communicates with high-level computing
resources hierarchically and depends on a higher layer to diverge.
Federated/ Collaborated Each computing resource computes independently to obtain a local result. They
could also communicate with other resources for aggregating the results. Notice the
local result do not depend on others to diverge.
Centralized Each resource does not compute and only communicates with a centralized
controller and entirely depends on the controller to compute the result.
Fig. 1. The taxonomy of typical computing paradigms. (a) is the local/ decentralized computing paradigm; (b) is the distributed computing paradigm; (c) is the
hierarchical/ coordinated computing paradigm; (d) is the federated/ collaborated computing paradigm; (e) is the centralized computing paradigm.
1.1. Terminology tions in IoT and CPS are relevant to this review. IoT and CPS mentioned
in this review mostly focus on application scenarios in the power sys-
Other terminologies, such as cloudlets, mobile edge computing/ multi- tem. The IoT for SG emphasizes interconnecting all things in the phys-
access edge computing or fog computing, variously refer to the similar ical power grid, especially in the power distribution network; the CPS
meaning of EC. These terms have been proposed by different organiza- for SG emphasizes information exchange, where the system should give
tions, but overall, they share a common ground with trivial differences. feedback and control the physical devices, forming a closed-loop control
Generally, EC covers a broader scope, so this paper uses the terminology system[13].
EC [6]. For a detailed comparison among these terms, readers can refer
to other related reviews [7]. 1.2. Related works
In addition to the terms above proposed in the communication indus-
try, another term, the grid edge with a similar meaning to EC, has been Thus far, a large number of works have surveyed EC on different
popularized recently in the power industry. The scopes of these two topics and from different viewpoints. Not intending to provide a fully
terms partially overlap but are not equivalent. Compared with EC, grid exhaustive list, we summarize several related works in Table 2, which
edge only comprises the last mile of the power distribution network/the provide complementary perspectives to this paper:
low voltage power grid. It mainly refers to the electrical equipment but In addition, a growing body of studies has investigated how EC can
not computing resources [8]. EC refers to the edge of ICT resources in help different industries and services. To exhibit the outline of indus-
SG. trial applications of EC, this paper selects some works and lists them as
Notably, the differences among EC and some other computing follows:
paradigms frequently used in studies, such as the distributed comput-
• Transportation and logistics: Forming a vehicle network is a future
ing paradigm, may be ambiguous. The authors describe existing com-
trend in the transportation industry. However, acquiring compu-
puting paradigms mentioned in studies [9,10] referenced in Table 1 to
tational and communication resources during mobility are signifi-
clear ambiguity. The local, distributed, hierarchical, and federated com-
cant concerns. [24] directly utilizes vehicles as the EC infrastruc-
puting paradigms can all be launched with EC. EC resources can switch
tures for communication and computation to form vehicular net-
among these paradigms with task offloading technology. In this sense, EC
works. [25] proposes a vehicular network architecture assisted by
does not equal to the distributed computing paradigm. In a broad sense,
EC integrating different types of communication access. During mo-
it may be arguable that if computing resources are allocated in different
bility, the handover/transferring of data among different computa-
places, then it is ”distributed” computing. This paper uses geo-distributed
tional nodes is critical to guarantee the continuity of vehicle net-
computing to clear ambiguity.
works’ operation. [26] investigates how to achieve the handover
There are some related reviews which discuss EC’s applications
and the processing of large datasets generated by vehicular networks
in the Internet of Things (IoT) and cyber-physical systems (CPSs). IoT-
with EC.
based [11] and CPS-based [12] applications are widely discussed to
• City services: [27] reviews existing approaches that utilize EC for
assist with the reform of SG. Thus these discussions about EC’s applica-
sustainable smart cities. Smart city services are comprehensive use
2
C. Feng, Y. Wang, Q. Chen et al.
Table 2
Summary of state-of-the-art surveys about EC: contributions and scopes.
Reference Contribution Inspiration to EC in SG Scope
Key Motivations Architectures Features Enabling technologies Applications Practical Projects Challenges
F. Bonomi et al.[14] Pioneering work to discuss EC SG is treated as a typical application. ✓ ✓ ✓ × ✓ × ×

applications in IoT.
W. Shi et al. [3] Milestone comprehensive review of SG is treated as a typical application. ✓ ✓ ✓ × ✓ × ✓
EC, covering the definition,
applications, et al.
M. Chiang et al.[11] Comprehensive EC review in the The power system is treated as a ✓ ✓ ✓ × ✓ × ✓
networking context of IoT. typical application.
M. Chiang et al.[15] Q&A guide on EC, covering common Offer a pratical guide on EC ✓ × ✓ × ✓ × ✓
questions about EC. implementations.
P. Mach et al.[16] Describes major use cases and Complement materails on EC’s task ✓ ✓ × ✓ ✓ × ✓
reference scenarios of EC, especially offloading.
on computation offloading.
3
T Taleb et al.[17] A survey focusing on EC’s fundamental Complement materails on enabling ✓ × ✓ ✓ ✓ ✓ ✓

key enabling technologies. technologies.
D. Sabella et al. [18] A tutorial on EC from the perspective Illustrate the use case of vehicle × ✓ × × ✓ ✓ ✓
of the communication standardization communication.
group.
N. Abbas et al. [19] A thorough review of EC, covering SG is treated as a typical application. ✓ ✓ ✓ ✓ ✓ ✓ ✓
almost all the scopes of EC.
P. Hu et al. [20] A comprehensive review of EC, SG is treated as a typical application. ✓ ✓ ✓ ✓ ✓ × ✓
particulally elaborating on
architectures and features.
F. Liu et al. [21] Explores existing EC systems and Specifies the categories of EC ✓ ✓ × ✓ × ✓ ✓
introduces representative projects. applications.
B. Omoniwa et al. [22] A tutorial overview for EC and IoT in SG is treated as a typical application. ✓ ✓ ✓ ✓ ✓ × ✓
32 pages, including the protocols and
the simultation tools.
X. Wang et al. [23] A comprehensive introduction of edge Complement materails on the edge ✓ ✓ × ✓ × × ✓
Advances in Applied Energy 1 (2021) 100006

intelligence, the core of EC use cases. intelligence.
cases for ICT technologies, requiring coordination among different the context of the IoT, EC is described as ”the enabling technologies
resources in the smart city. [28] offers an EC model for dynamic allowing computation to be performed at the edge of the network, on
orchestration of networking, caching, and computing resources to downstream data on behalf of cloud services and upstream data on be-
improve the performance of smart city applications. [29] presents half of IoT services”. [3]
an EC-supported smart city network architecture, on which things Although the definitions may differ, they still share one thing in com-
jointly compute, route, and communicate with one another through mon: EC deploys and uses computing resources close to data sources.
the smart city environment. Specifically, in SG, “edge” refers to places where the data are generated
• Intelligent healthcare: [30] reviews on EC for healthcare informatics near electrical equipment and users [34]. Admittedly, the scope of the
and lists application-specific health tasks handled by EC. Intelligent edge in SGs contains “the last mile” part of power distribution/ grid
healthcare needs EC to process data collected from body sensors to edge but is not limited to it. In the power transmission network, places
achieve a rapid diagnosis. [31] designs a system for monitoring and close to data sources should also be included. Take system operators as
analyzing the physical health of users using delicately deployed EC an example. Although substations do not belong to the last mile in the
computation nodes. [32] uses gateways as EC nodes for diagnosis, power distribution network, they generate data through various sensors.
which exploits the strategic position of gateways at the edge of the They thus include the scope of the edge.
network to form a smart e-health gateway. The definition of EC should also clarify the functions of EC in SG.
• Energy industry: Some studies have addressed implementing EC in the 4C, namely computation, communication, caching, and control are all vi-
power industry. These works are discussed in the following chapters tal functions of EC [35]. Conventionally, part of these 4C functions is
in more detail. used to be executed in the core data center by default. Moving 4C func-
tions to the edge will not cannibalize cloud computing but complement
1.3. Contributions and Organizations it. Through EC, the resources of the cloud, the edge, and the electri-
cal equipment are pooled as a cloud-to-things continuum [11]. Differ-
This paper attempts to provide a comprehensive review of existing ent computing resources, whether they are hierarchically placed in the
research related to the interdisciplinary field of EC and SG. A thorough same layer or different layers, can collaborate. Typical ways to organize
analysis of potential application scenarios of EC in SG is offered. Draw- these computational resources with EC are summarized as follows [36]:
ing upon opportunities for the integration of EC and SG, this present
• Push from Core/Cloud: Services and computation are actively pushed
paper also seeks to illustrate SG and EC’s synergistic effect: SG not only
to the edge. They used to be de facto carried out in the core data
benefits from EC but also supports EC to become sustainable. To sum-
center/cloud. They are offloaded to the edge to leverage locality to
marize, the contributions of this paper are as follows:
reduce response time or improve performance.
• Conducting a comprehensive review of interdisciplinary works re- • Pull By Bottom/IoT: Due to physical limits in communication, appli-
lated to the integration of EC and SG, with an emphasis on EC po- cations, usually IoT applications, are prohibited from transmitting
tential application scenarios in future SGs. data to the core center. Services and computation must be passively
• Providing a comprehensive technology roadmap to implement EC in pulled from the faraway places to the near edge. These applications
SG. Specifically, this paper explores the definition, characteristic, ar- are typically run on embedded systems-on-a-chip.
chitecture, and enabling technology of EC, from the power industry’s • Hybrid Cloud-Edge Analytic: This category combines the characteris-
perspective. tics of the cloud and the edge. For instance, EC resources are utilized
• Designing a fine-grained taxonomy for application scenarios of EC to pre-process data to alleviate data traffic; the data are then trans-
in SG. This taxonomy covers both applications summarized from ex- mitted to the powerful cloud to run sophisticated analytics. In this
isting related studies and potential applications yet undiscovered in way, both optimal results and minimum response time can be facil-
existing associated studies. itated.
• Investigating how the deep integration of EC and SG can support EC
Arguably, EC is brand new for the communication industry, but not
to achieve sustainability goals. It is shown that SG not only benefits
for the power industry. Some existing electrical equipment has already
from EC but also promotes the sustainable development of EC in
been installed with similar functions with EC. Take a conventional relay
turn.
protection system as an example. Relay protection equipment can gather
The rest of this paper is organized as follows: Section 2 presents data from electrical meters (communication), compute according to its
a detailed overview of EC encompassing the definition, architectures, relay setting (computation and caching), and then issue the control or-
characteristics, advantages, and key enabling technologies. Section 3 an- ders to line breakers (control) just near the power lines. Beyond that,
alyzes potential EC application scenarios in SGs. Section 4 studies the operation of the power grid is typically configured as a top-down
how SGs support the vision of EC to become sustainable computing. hierarchical structure. The lowest hierarchical structure lies in a single
Section 5 summarizes open research questions. Finally, conclusions are substation or power plant.
given in Section 6. However, it may be a misunderstanding to contest that EC has al-
ready been completely realized in SG. These early attempts lay a per-
2. An overview of EC in SG fect foundation for implementing EC in SGs, but they do not entirely
equal to EC. Compared with these existing equipment, EC is more in-
This section engages with the definition, architecture, characteris- telligent and flexible. For the conventional equipment, the computation
tics, and key enabling technologies of EC in SG. only takes place on its established programs to carry out fixed tasks. In
contrast, for EC systems, the computation could be migrated to different
2.1. EC definition in SG places, and thus tasks become changeable and on-demand. Meanwhile,
through task migration and dispatching, EC is an excellent choice to
To date, there is no consensus about the official definition of EC. The upgrade the computing capability of SGs without overthrowing exist-
definitions of EC in different industries embody various concepts since ing ICT systems. The computational resources already put in use can be
different industrial fields may have different requirements for EC [33]. fully utilized. These legacy computational resources are perfect candi-
For mobile communications applications, EC refers to the computing date resources for EC (e.g., smartphones, laptops, base stations could be
architecture that ”provides an IT service environment and cloud com- utilized in mobile edge computing) [20].
puting capabilities at the edge of the mobile network, within the radio In the near future, with the growing need for SGs to inter-react with
access network and in close proximity to mobile subscribers.” [19] In millions of distributed energy resources and electrical users, the impor-
4
Fig. 2. Potential users of EC in SGs.
tance of integrating EC in SGs will become increasingly evident. It is measurements to monitor grid statuses. Power quality data captured by
particularly true for the power distribution network, which is in the harmonic meters are sometimes also needed. For control, digital protec-
blind spot of the protection, control, and communication in the power tion relays and intelligent electronic devices (IEDs) are microprocessor-
system today. based controllers. IEDs transmit data to the master supervisory control
and data acquisition (SCADA) system. Necessary meteorological data
2.2. EC users in SG and geographic information and integrates different management sys-
tems should also be integrated into the SCADA. Currently, the SCADA
Before laying out more details about EC, it is necessary to clarify system is advancing to the wide-area monitoring system (WAMS). The
the scope of the edge in SGs and potential users of EC in SGs. They are WAMS deploys the phasor measurement unit (PMU) or micro PMU
summarized in Fig.(2) [37]. In this paper, we divide potential users mostly in power transmission networks but rarely in the power distri-
of EC (data consumers) in SG into four categories: grid users, market bution network. PMUs use time synchronization provided by the global
operators, energy service companies and system operators [38]. positioning system (GPS) to generate phasor data with a high resolution
Grid users are owners of various electrical equipment, including en- of up to 60 samples/s. These are primary data to represent the operation
ergy generator owners, energy storage owners, and electrical end-users. and control behaviors of the grid. To further monitor the health of elec-
Environmental sensors (e.g., temperature sensors) and electric meters trical equipment assets, various monitoring measures are also required:
are their primary data sources. They can use EC to build an energy man- surveillance video cameras are deployed to monitor the status of lines
agement platform to interact with the broader power grid and optimize and substations; hydrogen and moisture sensors are deployed to mon-
energy use or generation [39]. itor the status of transformers. System operators can use EC to offload
Market operators are organizers of various types of power markets. these tasks related to control, dispatch, and monitoring to the edge.
Usually, market operators need to collect real and reactive power in-
jections, operating limits, transaction bids and offers from market par-
2.3. EC architecture in SG
ticipants into their data centers. After clearing the market, the market
operator publishes tariff information and market outcomes. Market op-
A three-tier architecture to implement EC in SGs formed by the
erators can use EC to address the meter-to-cash process and organize
cloud-edge-thing continuum is suggested in this study, and it is shown
small-scale markets in the power distribution network.
in Fig.(3). Each tier can communicate both vertically and horizontally
Service companies are providers of energy services, including the re-
within the architecture. The details of each tier and their links are given
tailer and aggregator. They mainly collect user-side profiles and design
as follows:
customized energy services. They can use EC to efficiently deploy their
energy management schemes and publish energy scheme recommenda-
tions to users. 2.3.1. Tiers
System operators are in charge of the operation and control of the • Thing tier: This layer has been widely studied in the context of the
physical power system. Along a power transmission line or in a power IoT. Thus, this tier is named after the thing tier. The thing tier covers
substation, voltage, current, frequency, and power flow meters are basic most of the electrical equipment and the communication access in
5
Fig. 3. The three-tier architecture for implementing EC in SGs.
the SG. This tier is in charge of carrying the operations of SG and re- For example, UAVs could be emergency reserves for computation
alizing operation/control orders. Communication connections need resources with high mobility when SG encounters natural disasters.
to be established to transfer data to high layers. In this tier, various Usually, the closer an EC resource is to the thing tier, the faster the
devices, especially in the power distribution network, need to be- response can be achieved. In contrast, less powerful computing ca-
come end-to-end interconnected for different purposes, such as local pability can be provided. Therefore, there is a performance tradeoff
coordination and location-based services in SGs. between computing capability and delay. Different EC resources, ei-
• Edge tier: The edge tier contains the intermediate storing, communi- ther in the same layer or in different layers, may establish sponta-
cation, and computing resources between the SG control center and neous and constant connections to provide task migration, system
the things. The “edge” is a relative concept. Take the smart meter as resilience, and data sharing.
an example. If it is purely utilized to sense and transfer data (sens- • Cloud tier: Following the convention in the IoT and communication
ing), it belongs to the thing tier; if it is utilized as the platform for industries, “cloud” is used to name this layer. Nevertheless, cloud
home energy analytic (computation), it belongs to the edge layer. centers in the context of the communication industry may not cur-
The extensive resources in the edge layer can potentially be divided rently exist in most power grids. It does not indicate that SG opera-
into sub-layers according to their locations. The sub-layers closer tors must build a cloud data center before deploying EC. Before the
to the things consist of various low-power and fine-performance re- concept of “clouds” is proposed, SG operators and service compa-
sources (a single chip). They are mainly designed to perform local nies have been already equipped with the dispatch/control centers
and instant analytics. The sublayers closer to the control center com- with powerful computing resources comparable to cloud data cen-
prise resources precisely deployed with more powerful computing ters today. These resources are only dedicated to fixed power system
performance. They can communicate via reliable communication tasks and not designed for pay-per-use on-demand jobs. These dis-
access to perform complex and large-scale analysis and monitor- patch/control centers can be treated as the core computational re-
ing [40]. These delicate resources mainly contain small computers source as the same as the cloud data center and thus can be included
and servers with multi-cores. in this tier. This paper embraces all these resources in the domain of
Some potential components of the edge layer are exhibited in Fig. 3. ”clouds” for convenience.
Existing computation resources already at the edge of the power sys- This tier consists of control, storing, and computing centers. Com-
tem, if they are capable of running data analytics, could be added to pared with the edge tier, the clouds are designated with high-
the edge resource pool. The RTU, smart relay, smart controller, and performance storage and computing elements. These powerful re-
smart meter in Fig. 3 belong to this type. Besides, computation resources are deployed to perform complex analyses with a long-term
sources near the electrical users are also candidates for data analytic time scale and a grid-wide geographical scope. This layer is in charge
in EC, such as routers, gateways, mobile phones, PCs, and base sta- of global and centralized monitoring and control. Cloud centers also
tions (MEC resources) in Fig. 3. These resources are conventionally connect to migrate data and applications.
excluded from computation resources in SG. However, they can be
fully utilized by offloading SG data analytic when they are spared
from other tasks. Last, newly deployed resources are also promis-
ing, including the microcomputer, UAV, local data center, and local 2.3.2. Links
• Thing-edge link: The thing-edge link provides connections in the con-
control center. Especially, intelligent UAVs are often equipped with
powerful computation resources now. They are studied to provide text of the IoT in SGs. This link allows equipment and users in the
computing resources in both urban cities and remote rural areas. thing tier to access to the network via heterogeneous communica-
tion access. Data are uploaded, and applications are offloaded to the
6
edge layer via this link. This link also carries the feedback informa- users’ behaviors with learning ability. They can then select the most
tion and control signals from the edge. appropriate content to store as the user-side knowledge. For instance,
• Edge-cloud link: The edge-cloud link is established to enable the in- the correlations between electrical users and the environment can be
teractions of the edge and the cloud for further analysis. The edge learned as the basis for behavior prediction and energy management.
and the cloud deliver data and offload applications to each other Cognition EC releases the burden of the cloud by avoiding high storing
through this link. The volume, complexity, and frequency of data are demand and communication overhead. In the future, SG is innovated to-
expected to decrease along the thing-edge-cloud transmission path. wards customized services. The cognition of EC can support customized
Thus the frequency and granularity of the information along this link energy service recommendations.
should be carefully designed.
• Thing-cloud link: The thing-cloud link is built either before or after 2.4.3. Flexibility
the deployment of edge computing resources. This link is necessary The flexibility of EC is reflected in both the planning phase and
even with the existence of EC. The computing performance of the the operation phase. During the planning phase, the cloud needs to be
edge resources is less powerful than that of the cloud. Thus when planned with excess capacity to remain elastic. With EC, when new com-
most cloud resources are idle, cloud computing may be faster with putation tasks are introduced to SG, computing resources could just be
relatively less computing delay. Whether the thing-cloud link is in added locally near the source of new demands. The expansion of EC
operation should be dependent on the availability of cloud resources. closely follows the expanded needs of computation in SG. Thus EC is
oriented for on-demand use with less backup capacity and aligns with
2.4. EC characteristics and strengths the ”plug-and-play” proposal in SGs [43]. EC resource placement and de-
ployment can be deliberately optimized with more flexible choices. Ap-
Before EC, centralized cloud computing is a widely-accepted solu- propriate planning for EC can simultaneously reduce upfront investment
tion to analyze SGs’ data. The cloud connects distributed SG equipment and operational expenses. During the operation phase, EC can be tai-
to the central data center to provide computation resources. Neverthe- lored as modular solutions and offers a programmable ecosystem. Thus,
less, centralized cloud computing architecture may stumble SG’s further operators and service companies can deploy their service algorithms ef-
development. First, cloud computing is constrained by the limited band- ficiently. In addition, tasks can be intelligently distributed across differ-
width available in the distribution power network. Connecting millions ent computing resources in the thing-edge-cloud resource continuum.
of devices to the centralized data center may cause communication con- Thus, SG operators and service companies are able to deploy different
gestion and packet loss and may be technologically unfeasible. Second, computing paradigms in Table 1 (the table in section 1) flexibly.
the cloud may be too occupied to satisfy the heterogeneous requirements
of millions of SG devices. The details of SG devices and end-users could
2.4.4. Reliability
hardly be noticed and analyzed, degrading data analysis performance.
EC is reliable in terms of both physical operation security and data
Last, cloud computing may hazard the privacy and data security of SG
security and privacy. Uninterrupted and safe operation is the top prior-
device owners and electrical users. Uploading data increases the possi-
ity in SG. Leave-one-out (N-1) security analysis is a common criterion
bility of leakage of private or sensitive data. Thereby, it is acknowledged
in SG. The single-point failure of EC affects a much smaller scope than
that cloud computing may not be a cure for all industrial applications.
that of cloud computing. EC is also resilient to a single-point fault by
Compared with the cloud computing, the main characteristics and
further replicating the failed tasks to other EC nodes to maintain essen-
strengths of EC are as follows:
tial functions. EC can serve as emergency machine-to-machine (M2M)
communication resources. Cascading failure due to communication con-
2.4.1. Low-latency nection breakdown can then be avoided. From the perspective of data
The latency to transmit data and execute SG data analytic is formu- security and privacy, the cloud works on a shared background. It may
lated as Eq.(1). increase the possibility of leakage of private or sensitive data due to a
∑( ) ∑
𝐷= 𝑡𝑞𝑢𝑒,𝑘 + 𝑡𝑝𝑟𝑜𝑐,𝑘 + 𝑡𝑡𝑟𝑎,𝑘 + 𝑡𝑝𝑟𝑜𝑝,𝑖 + 𝑡𝑒𝑥𝑒 (1) lack of privacy safeguard machinery. EC can selectively process data lo-
𝑘 𝑖 cally or act as the proxy for encrypting sensitive data before data leave
for the cloud. In this way, EC reduces the quantity of grid users’ data at
where 𝑘 is the subscript for communication middle-ware along the data
risk.
transmission path; 𝑖 is the subscript for communication links along the
data transmission path. 𝑡𝑞𝑢𝑒,𝑘 , 𝑡𝑝𝑟𝑜𝑐,𝑘 , 𝑡𝑡𝑟𝑎,𝑘 refer to the queuing delay,
processing delay, and transmission delay for middle-ware 𝑘. 𝑡𝑝𝑟𝑜𝑝,𝑖 is 2.5. Key enabling technologies
the propagation delay passing link 𝑖. 𝑡𝑒𝑥𝑒 is the computation execution
time. Analytics can be directly undertaken close to the data sources. Key enabling technologies of EC in SGs are summarized on the phys-
∑ ( )
Thus, EC can reduce communication hops ( 𝑘 𝑡𝑞𝑢𝑒,𝑘 + 𝑡𝑝𝑟𝑜𝑐,𝑘 + 𝑡𝑡𝑟𝑎,𝑘 ) ical level, the presentation level, the task level, and the application level
∑
and the propagation distance ( 𝑖 𝑡𝑝𝑟𝑜𝑝,𝑖 ). For conventional power grid as follows:
applications, the delay for wide-area monitoring is usually on the or-
der of 100 ms; the delay for stability control is usually considered in 2.5.1. Physical-level technology
the range of 150 to 200 ms [41]. EC can also achieve a one-hop delay Physical-level technology is mainly utilized for industrial embedded
within 1 ms for picture processing according to existing tests [42]. In chips, including heterogeneous computing, lightweight function library, in-
addition, to guarantee low latency and jitter (the instability of latency), memory computing, and ML processors.
task offloading technology is used to dispatch and schedule tasks among Heterogeneous computing and lightweight function libraries are pro-
EC resources. Thereby, tasks can dynamically migrate to the places to posed to enhance the performance of existing embedded chips. The edge
obtain low latency, especially when the nearest EC resource is busy or nodes in SGs may consist of distinct commercial and mature micropro-
some communication link congests. cessors. Heterogeneous computing is required to coordinate these het-
erogeneous computing resources and usespecialized processing capabil-
2.4.2. Cognition ities to handle particular tasks [44]. The lightweight function library
EC resources only need to be in charge of devices and users within enables more applications to be run on full-fledged operating systems
their territory. Thus it gains significant advantages in understanding di- in resource-limited environments of SGs [45]. Most of the data analytic
versified requirements and perceiving environmental information near tasks are related to machine learning (ML). Thus some pioneering work
devices and users. EC resources can learn the patterns of devices’ and to build lightweight ML libraries on general chips is ongoing [46].
7
Table 3 2.5.3. Task-level technology

The main characteristics of EC which benefit SG. Tasks are smaller executable units of specific applications. Task-level
Characteristics Detailed description Application scenarios technology is utilized to dispatch and schedule SG tasks, also defined as
task offloading.
Low-Latency Less hops, less middlewares; Time-Sensitive
Task offloading technology is mainly related to optimization. It is
Isolation from other parts. Applications
Cognition Similar environment with SG agents; Cogitive generally formulated as a multi-object constrained math optimization
Tailored with user-side knowledge. Applications problem. The form of task offloading is free in the thing-edge-cloud con-
Flexibility Plug and play/On-demand; Community tinuum: devices in the thing layer can offload tasks to the edge layer to
Distribute tasks within community. Applications
save energy and prolong battery life; the cloud layer can offload tasks
Reliability Resilient to N-1 failure; Security
Cyber security and privacy. Applications
to the edge layer to reduce latency; computational resources in the edge
layer can offload tasks to each other to enhance the overall system per-
formance. The objective for task offloading is diversified, encompassing:
minimal delay [53,54], better computing performance [55,56], higher
In-memory computing and ML processors are innovations for em- reliability [57], less energy consumption [53,55] of the devices, or any
bedded chips to be used in the future. In-memory computing enables combination of them [58]. Readers can refer to Ref.([59]), which com-
high-performance chips by storing data in RAM and processing the data prehensively reviews existing research about task offloading problems.
in parallel [47].
ML processors are chips designed for edge learning tasks. The perfor- 2.5.4. Application-level technology
mance of existing computing resources at the edge of the power system EC is utilized for fulfilling data analytic tasks in SG, many of whose
and the typical commercial edge ML processors are shown in Tab.5. algorithms depend on machine learning (ML). Application-level tech-
The computing performance is measured as floating-point operations nologies enable data analysis of EC, specifically ML on EC, or and they
per second (FLOPS), which means the number of floating-point calcula- are sometimes referred to as edge intelligence. Typical application-level
tions a computing resource could perform per second. The larger FLOPS technologies are summarized as follows:
is, the more powerful the computing performance is [48]. The result is • Divide-and-conquer learning: Learning sometimes needs to be carried
shown in Table 5.
out in a divide-and-conquer manner if datasets or computing re-
Generally, training a machine learning model to classify images re-
sources are geo-distributed. Thus, a single EC resource is responsible
quires about 1G floating-point operations in each iteration [49]. FPGA
for part of the learning tasks. The global result is then aggregated
chips already deployed in SGs are qualified candidates for EC, and they
in different manners. For instance, optimal distributed learning re-
are already put in use to perform fast inference tasks for edge learning
quires frequent data exchange among EC nodes to reach equilibrium;
in other industries [50]. Other New ML computation resources, such as
federated learning only gathers individual results for EC nodes [60].
the graphics processing unit (GPU), the application-specific integrated • Offline-aided online learning: In most cases, EC performs online learn-
circuit (ASIC) are shown in Table 5. They satisfy the memory and com-
ing to capture real-time dynamics. However, EC’s computing capa-
putation performance requirement of ML analytic, and they are easily
bility is not as powerful as that of the cloud, especially when deal-
deployed and managed at the edge of the power system.
ing with the training process. Thus, EC can be aided through offline
learning in the cloud. Transfer learning, incremental learning, and
other new learning paradigms are promising approaches to address
this issue [61]. Training the ML model in the cloud, and making
2.5.2. Presentation-level technology inferences in the edge is also a feasible solution [45].
• Reinforcement learning: Reinforcement learning is one of the most
Presentation-level technology offers presentation and translation for
the detailed implementations of ICT, in which operators and service crucial enablers for the cognition of EC. Moreover, the hybrid could-
companies do not have expertise. These techniques mainly include vir- edge analytic framework is often used to release the burden of rein-
tualization and abstract as well. forcement learning computation for EC. It utilizes both the powerful
Virtual machine (VM) technology and container technology are vital computing performance of the cloud and the fast reactions of the
facilitators for EC users in SGs to dispatch and schedule computing re- edge. Through periodical communication between the cloud and the
sources. Different tasks may be programmed in different languages and edge, the RL models could update incrementally to adapt to the dy-
executed on heterogeneous operating systems (OSs) in SGs. These issues namic environment of SG and contribute to the cognition of EC [62].
give rise to interoperability and programmability problems, hindering
SG operators from migrating tasks [4]. VM technology and container 3. EC empowers SG applications
technology isolate tasks and abstract in-depth operating system details.
EC users in SG could dispatch tasks in VMs and containers instead of This section explores potential SG applications empowered by EC
simple tasks, which reduces the difficulty of migrating tasks among dif- from four dimensions: time-sensitive applications, cognitive applica-
ferent computational resources. A detailed introduction about virtual tions, community applications, and security applications.
machine and container technology can be found in Ref.([51]).
Software-defined networking (SDN) and network function virtualiza- 3.1. Time-sensitive applications
tion (NFV) are crucial tools for communication network management.
To dispatch SG tasks, SG operators and service companies need to con- First, time-sensitive applications cover SG operations and businesses
figure communication networks. The knowledge about the communica- requiring accurate actions and timely feedback. As shown in Fig. (4)(a),
tion network is quite specialized. Thus, it is difficult for SG operators to these applications directly benefit from the low-latency characteristic
control the communication network directly. SDN and NFV abstract the of EC. These scenarios cover both the power distribution network and
physical functions of the communication network in software. A com- the power transmission network. Real-time and even super real-time
prehensive management platform for communication networks can be control and protection and online applications are typical examples.
delicately built for EC users in SGs. Therefore, they can directly redesign For instance, frequency and volt/VAR regulation needs low latency to
and configure the communication network through these platforms ac- react to the fast dynamics of the frequency and voltage; coordinated
cording to their needs. A detailed introduction to SDN and NFV tech- control for electric vehicles needs low latency to guarantee the per-
nology can be found in Ref.([52]). formance and timeliness of coordination [63]; online state estimation,
8
Table 4
The main enabling technologies which empower EC’s applications in SG.
Category Detailed description Enabling technologies
For Existing Heterogeneous Computing

Hardware-level Enable efficient usage of embedded chips Hardware Lightweight Function
deployed in SG for EC computation For New In-momery Computing
Hardware ML Processors
Computation Virtual Machine
Presentation-level Managing how SG computation tasks are Dispatch Container
distributed among different nodes. Communiction SDN
Management NFV
Managing how SG computation tasks are
Task-level Task Offloading/Migrating
distributed among different nodes
Providing SG services and data analytics, Divide-and-conquer Learning
Application-level
especially data-driven learning Offline-aided Online Learning
Reinforcement Learning
Table 5
Comparison of the performance among different computation resources. (1T=103 G=106 M=1012 ).
Computation Memory/
Category Hardware performance/FLOPS Byte
Complex Programmable Logic Device (CPLD) 1M 1k

Low-end microcontroller unit (MCU) 10M 10k
Existing computation resources in SGs High-end microcontroller unit (MCU) 500M 500k
Digital signal processor (DSP) 100G 500k
Field programmable gate array (FPGA) 1T 1M
Graphics Processing Unit (GPU) 1T 1G
Low-end application-specific integrated circuit (ASIC) 100G 1G
Off-the-shelf edge learning resources High-end application-specific integrated circuit (ASIC) 1T 10G
Micro computer 100G 100G
Multi-core server 1T 1T
Fig. 4. (a) SG operations and businesses requiring accurate actions

and timely feedback (If one task requires higher reliability and lower
latency, it is contained in the darker red region.); (b) EC can offload
tasks to different places to reduce latency
online demand response, and online asset monitoring need the fresh- tency, as shown in Fig.(4)(b). Several studies have already investigated
est data obtain the desired result. Noticeably, some of these applica- time-sensitive tasks by using existing ICT infrastructures as EC computa-
tions listed in Fig.(4)(a) may have already been designed to be exe- tion resources. [64] proposes an edge framework for wide-area backup
cuted near the sensors. Nevertheless, in EC, task offloading offers a more protection. The protection scheme uses IEDs as primary EC nodes for
guaranteed low-latency performance. Applications can be re-dispatched fast and low-latency short circuit diagnosis. [65] invented an accurate
and re-scheduled in the cloud-edge-thing continuum to assure low la- fault location model trained by the radial basis function neural network
9
and stored in the substation EC server. [66] uses FPGAs as edge com- data according to the requirements for different services. It is vexing for
putational resources to detect high-impedance faults in real-time. The the cloud to determine the degree and granularity of every cleaning re-
limited resources on FPGAs are utilized via hardware parallelism and quirement for every application. Different services may have different
pipelining. requirements. The cloud may not be capable of handling all these trivial-
Furthermore, EC brings support for measuring the effects of com- ities. If there is too much information kept, extra bandwidth and storage
munication latency quantitatively rather than qualitatively, especially will be occupied, and the data needs to be further processed. If too much
in the power distribution network, where fiber communication could raw data are filtered out, data re-request is required, and it may incur
hardly reach. Under the support of EC, the communication latency high communication overheads. Data generated by neighboring sensors
reaches the millisecond resolution. It is comparable to the typical sys- are often only locally relevant; thus, EC is suitable for data-processing
tem sine wave period of SG (1/50 s or 1/60 s). If the network latency and caching needs. EC can use its cognition to identify the importance
and jitter are low and predictable, SG applications can quantitatively of the data and determine how different data are managed differently.
measure latency as endogenous parameters instead of disturbance. For For data analytics, a part of SG data analytics is also recommended to
instance, the determined bounded delay 𝜏 can be treated as a quantita- be accomplished by cognitive EC, as shown in Fig. (5). They are featured
tive parameter in the controller design and real-time demand response by extracting knowledge from raw data or fusing different types of data.
programs. Quarantining the errors created by the communication pro- Notably, some of these listed analytic tasks can be carried out either on
cess, the accuracy and predictability of the control and dispatch a are the edge or the cloud. If they fall into the following four categories, EC
enhanced [67]. may be a more reasonable choice [72].
In addition to SG operations and services, EC is also crucial for han-
dling and optimizing SG data flow/traffic, especially with the advent of • Only the information after being processed instead of raw data is
big data. In the power distribution network, a GPS phase-locked oscilla- needed. Cognition EC is equipped with enough computational re-
tor provides an accurate reference source for PMUs. However, without sources for fulfilling tasks. Transferring them to the cloud for exe-
data traffic management, phasors causally are delayed, and the order cution is unnecessary-for example, asset monitoring and alarming in
of phasors at different measurement points are distorted; thus, SG ob- power substations.
servability is reduced. The decisions for data flow are related to: the • Information is not permitted to transfer to the faraway cloud for
composition of the data packet (e.g.,whether the data load consists of privacy concerns. Additionally, extracting and filtering useful infor-
time), the sampling rule (how often data should be sampled), the send- mation by using local access on-premises is permitted. EC resources
ing rule (how often data should be sent), the queuing strategy (how can support the deployment of these filtering algorithms-for exam-
to organize the data queue), and the routing policy (how to schedule ple, private electrical user load information analytics.
the route of the transmission) [68]. These data flow optimizations can • The dilemma of handling data traffic pulls applications. Transferring
be made by observing the quality of service (QoS) metrics, including them to the cloud incurs high communication overhead. Variations
bit rate, throughput, delay, availability, jitter, and age of information among different nodes are high and can be paralleled by cognition
(AoI). On the one hand, data flow optimizations need to be executed EC to reduce the complexity. For example, visual recognition of the
ahead of other data analytics on SG operations and services. Thus, data states of the power lines by UAVs.
flow optimization problems need to be solved as soon as possible. Be-
sidesBe, it should be executed online to capture the dynamics of data For example, [73] investigates dynamic residential energy optimiza-
flow for stochastic events. Thus, solving them in the edge layer becomes tion utilizing the memorization functions of EC. The results of short-
the foremost selection. [69] uses smart meter embedded computational term sub-problems are buffered, and long-term problem solutions can
resources for data analytics in the context of EC and exposes the impact retrieve these results instead of recomputation and thus, the execution
of sampling frequency and digitization resolution. time can be reduced. [74] proposes an edge-cloud integrated solution
for building demand response using reinforcement learning (RL). An RL
3.2. Cognitive applications agent learns an optimal control policy on the cloud, and the policy is
then distributed to EC resources for execution. [75] present an energy
SG equipment is becoming more diversified; SG services are becom- trading management system where the edge node acts as a retail energy
ing customized. Cognitive applications grasp distinct features among market server, which provides energy services to the end-users.
equipment and end-users, offering fine-grained intelligence to SGs. Among cognitive applications, this paper highlights cognitive busi-
Intelligent agents are one of the approaches for achieving cognition ness services for electrical consumers offered by service companies in the
in SGs. The EC resource is the ideal carrier for intelligent agents [28,62]. power distribution network. A broad range of services can be provided
Intelligent agents are utilized for two classes of functions in SGs, one by service companies, including personalized energy pricing, energy
representing the energy market issues, including energy balance man- scheme recommendation, energy efficiency optimization, sustainability
agement, energy pricing, and energy scheduling. The other focuses on- advisory, and energy infrastructure sharing. These can be considered as
network security, reliability, fault handling, and efficiency [70]. Never- the EaaS (energy as a service) business model in SGs, similar to infras-
theless, especially in the power distribution network, modelling millions tructure/platform/software as a service (IaaS/PaaS/SaaS) model in the
of SG agents in the cloud is complex. Modeling the action set 𝐱𝑡 , the state communication industry. EaaS models indicate that electrical users can
( )
set 𝐬𝑡 , the environment 𝛏𝑡 , and the transition dynamics 𝐬𝑡+1 = 𝐝 𝐱𝑡 , 𝐬𝑡 , 𝛏𝑡 completely outsource energy portfolio option design, and grid users do
explicitly for every agent in the cloud incurs high communication, com- not need to address detailed implementations. Thus far, only some small-
putational, and storage costs. EC gains advantages over clouds since it scale pilot tests of EaaS have been launched in the form of one-to-one
can achieve high granularity for perceiving the parameters and build- negotiation. If EaaS is rolled out to cover all the electrical consumers,
ing the models. EC can act as the agent for market organizers in the EC can be a promising enabling technology for its cognition of electrical
distributed energy market and selectively buffer the trading records or user behavior patterns. In addition, it is also programmable and can be
upload records to the cloud for further checks. Utilizing EC resources written with constraints to guarantee that the EaaS schemes are phys-
for blockchains is also a future choice [71]. ically achievable and promise reasonable profit for service companies.
Utilizing data-driven methods to analyze equipment and end-users in Initial attempts are ongoing, [76] introduced a novel platform for home
the power distribution network is another approach for achieving cogni- energy management-as-a-service platforms over EC. The platform fea-
tion. However, how and where data-driven models, including data pre- tures open-source software/hardware to enable the user to implement
processing and data analytics, are implemented is still an open question. energy management with customized control-as-services while minimiz-
For data pre-processing, EC is appealing for cognitively pre-processing ing the implementation cost.
10
Fig. 5. (a) Cognition EC and main EC applications; (b) The taxonomy of cognitive EC analytics.
3.3. Community applications tion rate of energy storage is performed at the minute level in a re-
gional cooperative manner. Economic bidding optimization and other
EC is flexible for using the computing resources in the thing-edge- grid-connected decisions may need to be made at the quarter-hour level
cloud resource continuum. Community applications represent those ap- in a full global manner. EC can adaptively optimize how computational
plications that can utilize and dispatch computational resources in the resources are organized according to requirements of applications and
continuum. characteristics of computational resources inside the community. This
Forming the community of intelligent agents for SG operations and approach differs from the conventional multi-agent system, as the frame-
services can be fulfilled with EC. The community could be formed by works of cooperation, competition, supervision coexist. For example,
the power transmission network alone, the power distribution network the hybrid cloud-edge analytic scheme, which dynamically distributes
alone, or simultaneously. Developing community agents does not equal computation tasks among EC nodes and the cloud, prevails in many
pure aggregations of computational resources. Cooperation and com- works. [45] designs a city-wide energy management system based on
pletion, negotiation and compromise coexist inside a community to EC infrastructure. Through deep reinforcement learning (DRL), each
complete different objectives with different time scales and geographic edge node becomes an edge DRL agent, and the cloud becomes the
scopes [77]. In SGs, micro-grids and virtual power plants (VPPs) are cloud DRL agent. The training and inference process is cooperatively
typical examples of forming community resources. In these systems, distributed among different agents. [78] uses a three-layer cloud-edge
the transient stability and system inertia need to be supported at the computing architecture for energy management by reconfiguring the
sub-second level in a fully distributed manner to respond to contin- power network topology considering the dynamic thermal line rating
gency rapidly. The power balance, frequency, and voltage regulation (DLR) constraint. EC resources are designed to generate Benders cuts,
services need to be handled at the second level in a regional autonomous and the cloud is designated as the coordinator. [79] proposes a dy-
manner. The optimization of the cost of generations and the utiliza- namic economic dispatch paradigm based on a local DSP chip (edge)
11
Fig. 6. (a) Crowd-sensing: (I) EC nodes publish incentives for data collections and recruit data workers. (II) End devices collect and submit data according to
incentives; (III) EC nodes distribute rewards according to individual contributions to tasks; (b) EC-based digital agent simulation: (I) Physical sensors collect and
transmit real-time data to EC nodes. (II) EC nodes receive data and build digital agents; (III) Digital agents form community just as the physical power system does,
and simulate the behaviors of the physical power system.
Fig. 7. (a) Emergency replica through EC:
when the main computing resource fails, back-
up EC nodes serve and cascading failure is
avoided; (b) AAA management taken over by
EC: EC nodes are in charge of AAA management
in place of the cloud ; (c) User data privacy me-
diator through EC: EC nodes process sensitive
data before it leaves for the cloud to protect
privacy.
Fig. 8. The approaches for EC to become sustainable

with the support of SGs.
12
and a remote cloud computing platform (cloud). The remote cloud con- For physical operation, inevitably, the stability of SGs will be affected
ducts offline calculations, while DSP chips help with real-time decision by ICT resources. The breakdown of ICT may incur cascading failures to
making. challenge the operation of the entire SG. EC can improve the resilience
In addition to the electrical agent community, the sensor commu- of SG in comparison with cloud computing in two typical cases [87]. In
nity is also crucial in SG for collecting and aggregating data. The con- one case, when the power system fails first, then the power supply losses
ventional data collection policy is marked by sampling data with fixed of the ICT resources will follow. A wider-area breakdown could develop,
intervals and compulsorily updating data. However, this fixed policy and it may raise the difficulty of recovering. Under this circumstance, SG
may encounter several problems with a growing need for more data in can use idle EC resources (such as base stations with UPS and UAVs) to
the power distribution network. First, data collection devices are owned support contingency. EC can support emergency communication (in the
by different stakeholders. They may not voluntarily submit their sensor way of D2D communication) and offload essential tasks to ensure that
data. Proper incentives are needed to exchange/trade to gather the nec- dispensable services operate continuously. In the other case, when the
essary data [80]. Second, the data collected from different devices may ICT resources fail first, a single-point failure will not be as destructive
have different values for different application scenarios. Operators and as the case with cloud computing alone. SG can utilize EC resources by
service companies should determine the importance/value of data and replicating tasks to other functioning devices as contingency plans. In
allocate more computation and communication resources to more valu- this way, possible cascading failures due to ICT resource breakdown can
able data. Crowd-sensing, widely studied to be launched with EC, is a be prevented [88]. These duplicates also enable SG immunity to most
solution to these challenges [81]. The sensors are called data workers in minor ICT disruptions. Ref.[11] highlights that EC could replicate the
crowd-sensing and are paid for their efforts to collect high-value data. power plant’s operation software and support its continuous operation.
EC in crowd-sensing is responsible for detecting newly available workers Additionally, local resources can be spared for updates.
and pushing notifications to data collection requirements and data gath- For cybersecurity, relevant challenges can also be addressed with EC.
ering. By recording the reputation (a measurement of trustworthiness Authentication, authorization, and accounting (AAA) are fundamental
for the quality of contributed data) within the data worker community, elements of cybersecurity. They control who is permitted to use net-
EC could identify the importance of sensors [82]. In addition, instead of work resources (through authentication), what they are authorized to
continuously collecting numerous data, EC can publish data collection do (through authorization), and capture the actions performed while ac-
demands that matter to SG operations and businesses and alleviate the cessing the network (through accounting). Using the cloud to monitor
increase in data traffic. for a large number of geo-distributed devices incurs massive overhead.
Ideally, the cyber communities built with EC can evolve as a digital EC nodes can delegate the task of security monitoring and AAA manage-
twin system for SGs in both the power transmission network and the ment [11,52]. Edge-based AAA management is particularly significant
power distribution network. Traditionally, to predict the behaviors of for the power distribution system, where SG equipment communicate
the power grid, operators depend on computer simulation composed of through public communication network instead of utilities’ private net-
programmed analytic solutions for SG equipment. The increasing com- work.
plexity of equipment and users degrades the accuracy of the conven- For protecting electrical users’ privacy in the power distribution net-
tional simulations. With cognition, EC can use data to “clone” the SG work, some privacy-protecting schemes also agree with the ideas of EC.
devices in its territory while retaining high-level details. Furthermore, These schemes are data perturbation (add noise to a subset of data that
EC can digitally replicate the states and actions just as the same cy- users deem sensitive without compromising the accuracy of the data);
ber representation of a single device or user. Afterward, the hardware/ data encryption (use cryptographic techniques for security among multi-
human-in-the-loop simulation can be equivalently carried with EC re- party computations, adding keys for accessing the contents of data); and
sources. For example, if operators want to test a dispatch/schedule strat- data anonymization (removing relevant information that could make
egy, EC resources can act as the corresponding grid components. If ser- others identify the identity of users). These schemes are virtually added
vice companies want to test the effect of a demand response program, EC privacy-oriented data analytic tasks. EC can act as privacy mediators to
can mimic grid users’ negotiation behaviours. The digital twin platform carry out these privacy-oriented tasks before data depart for the cloud.
should be maintained by the platform owner and test projects should Notably, the excessive work for securing privacy needs to be balanced
be paid by the platform users. Thus, a digital twin of the physical sys- with its cost (energy, computation, and communication cost)[89].
tem is established in an agent-based, parallel, and geo-distributed man-
ner. The connection between the physical model and the corresponding
virtual model is established through real-time data. Simulations in SG 3.5. Application cases
become living: digital simulation models could update just the same as
their physical counterparts change. Simulations can be run on this dig- The communication service companies are pioneers who attempt to
ital replica recorded in EC resources, and the result will be the same as implement EC. They take advantage of their hardware and platforms
the physical network. To reduce the communication overheads in the built in the cloud computing era to provide EC products and services.
digital twin system, elastic updating methods, or new means of com- Their products have already been put into the market and supplied to
munication such as 5G communication could be used. The dispatch and different industrial applications. Micro data centers, micro computers,
scheduling of computation resources also require to be optimized to im- and EC chips embedded in gateways are typical examples.
prove the performance of the twin system. This idea originates from Currently, equipment and service providers in the energy industries
the aerospace/ satellite system [83], the digital twin in the industrial have gradually attached to importance EC products’ development. Their
IoT [10], and distributed co-simulation in SG [84,85]. products focus on more specialized application scenarios suited for SGs,
not constrained to conventional data centers and servers. To summa-
rize, their EC products offer EC solutions covering edge-based hardware,
3.4. Security applications edge-managed platforms, and edge-enable services.
From the perspective of physical operations, a secure power grid • Edge-based hardware: EC controllers in the power distribution net-
should resist accidents or disasters and fully ensure power supply un- work are typical examples of edge-based hardware. For the Volt-VAR
der contingency in the power transmission and the power distribution. control solution, Varentec provides Edge of Network Grid Optimiza-
From the perspective of cyber operations, a secure SG should guarantee tion (ENGO), a dynamic VAR source that can be mounted on a pole
cybersecurity to avoid the threat of data leakage and deliberate attacks top or a meter pedestal. It is usually connected to the customer-side
jeopardizing SGs [86]. of the transformer. ENGO dynamic VAR devices autonomously sense
13
and regulate voltages by injecting dynamically reactive power at the the air and the energy in the electromagnetic wave generated through
grid edge [90]. wireless communication and power transmission. Energy harvesting can
• Edge-managed platform: Edge-managed platforms provide software recycle thermal energy and electromagnetic energy [99]. These energy
platforms for EC users in SGs. This platform allows customers to de- generations are potential new local power supplies for EC nodes.
velop SG applications without caring about the complexity of build- Trading energy is an indirect measure for achieving economic sus-
ing and maintaining the hardware. For example, APM Edge is an on- tainability by cutting energy costs to achieve economic sustainability. In
premise asset performance management platform in both the power SGs, bidirectional information and power flow are both available. Local
transmission and distribution network. APM Edge enables asset own- self-generated energy can be sold to other electrical users via distributed
ers to leverage their online and offline data to drive more intelli- energy trading or transactive energy trading [100]. The EC provider
gent, risk-based approaches to asset management. This edge-based can gain arbitrage through different levels of real-time prices. Alterna-
platform offers quick installation and seamless migration from mon- tively, it can sign contracts with sizable renewable generation plants
itoring a single transformer to having the insight of an entire fleet to absorb its energy. In addition, through demand response programs,
of various electrical assets [91]. Similar applications include SINU- EC resources can provide different types of services, such as frequency/
MERIK Edge [92] and Predix Edge [93]. voltage regulation [101].
• Edge-enable services: Edge-enable services are designed upon edge- EaaS models are also promising for a sustainable EC. The reliabil-
managed platforms. Through edge-enable services, EC providers, in- ity of EC cannot be guaranteed by relying solely on renewable energy
stead of EC users in SGs, are responsible for building, maintaining, supply. A backup secondary power supply (such as energy storage or
and updating SG applications on these platforms. For example, Grid- power grid cables) is needed. The two sets of power supply (primary
stream is an edge-based energy management service solution sup- power supply and backup power supply) add enormous pressure to the
porting advanced multi-energy metering and personal energy man- capital cost and operational cost for EC service providers. In SGs, this
agement. It supports up to 5-minute interval data collection, load problem can be addressed with the storage renting or storage sharing
management, home area networking, and outage management ap- market, which spares EC users’ efforts to build and manage storage by
plications [94]. themselves. EaaS models can further support outsourcing the manage-
ment of EC operators’ energy usage and carbon emission profiles [102].
Besides, industrial providers in both the communication and the Remarkably, EC in the communication industry and micro-grids in
power industry begin to cooperate, to complement their advantages. In SG are symmetrical. EC can be considered to push the computing func-
China, several related programmes are undergoing. In the Qi-Yu ultra tions to the edge; the micro-grid can be treated as migrating the electri-
high voltage substation in Henan province, EC is used to analyze data cal functions of the power system to the edge. EC can connect to the
collected by robots the inspection robot, the surveillance camera and cloud and solely manage local computation tasks; the microgrid can
other monitoring equipment. After being processed, part of these data operate in grid-connected and stand-alone modes. The integration of
are routed to the substation control room, and part of these data are EC and the microgrid, these two geo-distributed models reinforce each
transmitted to the provincial dispatch center. Some programmes com- other’s functions through interaction and collaboration between the sys-
bine 5G and EC to provide communication network isolation, ensuring tems. EC supports the control and communication of the microgrid. The
the security of sensitive data [95]. microgrid can also supply power to EC resources using renewable en-
ergy. If EC is utilized within a microgrid, it can be treated as a flexi-
4. SG supports sustainable EC ble electrical load [103]. The local generations in a microgrid are usu-
ally composed of renewable energy with low marginal prices. Thus, EC
At present, the power consumption of ICT resources (excluding providers do not need to to deliberately build a self-supporting renew-
smartphones) amounts to 8% of the total worldwide power consump- able generation to reduce carbon emissions and operation costs.
tion. It is projected to reach 14% by 2020 [96]. Overall, EC itself is more
sustainable than cloud computing. First, it avoids large cooling systems 5. Research challenges and future trends
in the cloud data center, making a significant share (around 30%) of
the cloud’s energy consumption. Second, the request of SG data analytic Although EC has drawn significant attention from energy industries,
may not always be active, and thereby EC resources could be turned off there are still some challenges ahead. This paper identifies relevant chal-
adaptively to save energy. Third, EC reduces the data traffic between the lenges and future trends, which are urgently needed to be researched to
edge layer and the cloud layer, and intra-data center networking, which pave the way for EC’s large scale implementations in SG.
reduces the energy consumption of network switches and routers [97].
However, the energy efficiency of EC may not be competent to the cloud, 5.1. Application-aware data management scheme
resulting in more energy consumption per computation instruction. Ac-
cording to estimates, EC consumes less 10%-30% energy than the cloud In the near future, the implementation of EC in SGs will be accompa-
combining all factors above [98]. nied by the proliferation of massive data. The SG data flow could pose
Nevertheless, leaving expanding EC resources to become denser will more and more influence on the operation of SG operations and services,
still incur unacceptable energy costs. Additionally, the government and and thus the management of SG data becomes increasingly significant.
regulators use mandatory quotas and carbon taxes to push ICT resources The present data management scheme is isolated from SG applications,
to consume renewable energy. Considering the above, although EC is and the performance of SG operations could not give feedback to the
technically promising, it may be neither environmentally nor econom- data management scheme, which is not “application-aware.”
ically sustainable. Some studies have already attached importance to Take the meter data collection process at the electrical user side as
energy-aware task offloading to save energy for EC resources. Here, we an example. During the standard operations, the user’s electric load’s
illustrate how SGs can assist in the reform of a more sustainable EC. collection interval shall not be greater than 15min. During the demand
Finding alternatives for fossil fuels as generation resources and re- response, the collection interval shall not be greater than 1min for
placing these brown energy sources can fundamentally achieve the sus- contract-based demand response, and the collection interval shall not
tainability of EC. EC infrastructures can directly utilize energy sources be greater than 30s for real-time demand response. In the case of ac-
in the environment (including wind, solar, thermal energy) to provide curate load shedding, the collection interval shall be at the millisecond
self-supporting electricity. In addition, part of the electricity is not en- level. In order to satisfy the requirements of the accurate load shed-
tirely converted into the target energy and escapes as other means of ding, the data collection is now considered to remain a similar level of
energy. These types of energy typically include the thermal energy of that. It may significantly increase the overheads of communication and
14
memories, resulting in communication congestion and memory explo- are honest. If the attackers physically compromise the majority of the EC
sion, affecting other SG applications’ real-time performance. Dynami- resources in a house, a building, or even an entire region, these mecha-
cally changing the data collection frequency rather than keeping con- nisms will no longer be adequate. The maintenance, repair, and replace-
stant frequency according to the real-time application requirements can ment of compromised EC nodes is also a considerable project.
alleviate the adverse effects. Beyond hardware security problems, most EC nodes are occupied
In addition to the frequency of data acquisition, the number of data with operational tasks. Thus these resource-constrained devices can
samples, data transmission routing can all managed in this application- hardly make room for selected security functions to protect themselves.
aware way. For example, the data routing could monitor the power sup- The public space of EC nodes poses a significant challenge to establish
ply and bypass routers of unreliable power supply to enhance the relia- a trusted authentication and access control mechanism.
bility of SGs.
5.5. Development environment and business models
5.2. Modular and dispatchable data analytic
EC technologies are still young and may need years of progressive
More and more data analytic methods are adopted to serve SGs, es- research to evolve. The development of EC requires the involvement and
pecially for electrical user-oriented applications like energy usage opti- support of a complex set of industries, communities, and standard orga-
mization. Electrical users’ behaviors and various demand-side resources nizations. There must be some industries to be the pioneer in leveraging
(like electric vehicles) could be highly nonlinear and dynamic. Thus EC. What distinguishes a killer application for EC also remains an open
some complex data-driven models are needed to be utilized, which usu- question. SG may become the main precursor for EC. However, there
ally ensemble several simple learning processes. However, most of these should be enough incentives for SG operators to adopt EC. Nevertheless,
data-driven models are organized as a chunk file, integrating the whole to date, there have been no established business models and ubiquitous
algorithm from the start to the end. That may incur a high complex- infrastructures for EC. This may prevent SG operators from moving most
ity and resource usage for algorithms, resulting in difficulty for EC to of these applications to the edge. Thus, from a non-technological per-
migrate these tasks. spective, the industry community should pay sufficient attention to eco-
Nowadays, like the development of the microservice in the data cen- nomic incentives and accounting mechanisms. The mechanism should
ters, algorithms’ modular design is prevailing. Chunk data-driven user provide EC solutions to ensure a sustainable development environment
analytic tasks could be separated into several modular subtasks. Each for EC technologies.
subtask could be easily put into a lightweight container, which releases The highly geo-distributed and heterogeneous nature of EC adds dif-
the burden of task migration and operation. Beyond that, modular data ficulty in determining unifying standards and interfaces. In addition,
analytic also does good to fast creation and building of different SG EC users already have some standards or protocols in their existing sys-
tasks. For example, suppose one electrical user analytic task requires tems. Therefore, adaption to these old standards in different industries
the support vector machine (SVM) and the decision tree. In that case, is indispensable. Some protocols are IEC 61850 protocols for IED com-
the modular subtasks for the SVM and the decision tree can be put in two munications used in power and energy society; IEEE 1451 standard for
containers and migrated to the edge gateway of that user. A complete connecting sensors or actuators to microprocessors used in instrumen-
task could thus be efficiently built. tation and measurement society; IEEE P2668 standard for facilitating
the evaluation of IoT maturity newly brought by Industrial Electronics
5.3. Performance-complexity trade-off evaluation Society. These standards and interfaces are supposed to enable different
computational resources in different layers to interact with each other
Due to EC resources’ heterogeneity, tasks should be adapted to EC smoothly. In this sense, an open-source platform formed by different
resources when migrated to the edge. Namely, the performance of data industry communities should be established to create a collaborative
analytic algorithms and the constraints for EC resources should be bal- development environment. There should be a long way to address this
anced. Up to now, most of the data analytic frameworks in SGs assume issue.
that computation resources are sufficient for services, lack of consider-
ation for real implementations for these algorithms.
At the theoretical level, the result under partial execution should 6. Conclusions
be carefully studied. A performance guarantee bound should be given
if a task is carried out with limited computational resources within a In this survey, we aim to discuss various theoretical and practical as-
given time limit. That bound lets us know at most how far the partial pects when EC encounters SGs. First, related works of EC applications in
optimization result is away from the optimal result. For example, for different industries are summarized to gain a first glimpse of EC. Next,
battery charging and discharging management, how far is the gap be- we dive into specific insights into EC in SGs, including the definition,
tween the result under the complete optimization and the result under architecture, characteristics, and key enabling technologies of EC from
optimization with shorter time horizons and shorter time limits? the prospect of the power industries. A comprehensive investigation into
At the application level, it is necessary to study the practical meth- the potential EC applications in SG is presented to understand the role
ods of model simplifying and model compression. Simplifying the model of EC. Assessment of how to make EC sustainable with SGs is also thor-
is extremely important for tasks that contain large amounts of parame- oughly examined.
ters, such as Deep Neural Networks for SG fault diagnosis. Luckily, some Considering that EC technology is still young and immature, open
studies have put related efforts. The early inference is used to reduce the questions of EC in SGs may still require further research and develop-
number of network parameters; the gradient compression and pruning ment. Improvement could be made in terms of EC system design, algo-
are used to reduce the bits of gradient data; the knowledge distillation rithms applicable for EC, resources management under the EC environ-
is used to reduce the model size. ment, and even hardware accelerations. Both opportunities and techni-
cal challenges of EC implementation in SGs may emerge in the future.
5.4. Edge security and maintenance
Declaration of Competing Interest
Although EC can enhance security compared with clouds, it still faces
several security challenges yet. The geo-distribution of EC resources in- The authors declare that they have no known competing financial
creases the risk of physical attacks. The existing security mechanisms interests or personal relationships that could have appeared to influence
typically rely on the assumption that the majority of the data sources the work reported in this paper.
15
Acknowledgements [29] Naranjo PGV, Pooranian Z, Shojafar M, Conti M, Buyya R. FOCAN: a fog-supported
smart city network architecture for management of applications in the Internet of
everything environments. J Parallel Distrib Comput 2019;132:274–83.
This work was supported by Major Smart Grid Joint Project of Na- [30] Kraemer FA, Braten AE, Tamkittikhun N, Palma D. Fog computing in healthcare
tional Natural Science Foundation of China and State Grid under Grant review and discussion. IEEE Access 2017;5:9206–22.
U2066205, and the National Natural Science Foundation China and [31] Chen M, Li W, Hao Y, Qian Y, Humar I. Edge cognitive computing
based smart healthcare system. Future Gen Comput Syst 2018;86:403–11.
Joint Programming Initiative Urban Europe Call (NSFCJPI UE) under doi:10.1016/j.future.2018.03.054.
Grant 71961137. [32] Rahmani AM, Gia TN, Negash B, Anzanpour A, Azimi I, Jiang M, et al. Exploiting
smart e-health gateways at the edge of healthcare Internet-of-things: A fog com-
References puting approach. Future Gen Comput Syst 2018;78:641–58.
[33] Premsankar G, Di Francesco M, Taleb T. Edge computing for the Inter-
[1] Fang X, Misra S, Xue G, Yang D. Smart grid the new and improved net of things: a case study. IEEE Internet Things J 2018;5(2):1275–84.
power grid: a survey. IEEE Commun Surv Tutor 2012;14(4):944–80. doi:10.1109/JIOT.2018.2805263.
doi:10.1109/SURV.2011.101911.00087. [34] Ibrahim MS, Dong W, Yang Q. Machine learning driven smart electric power
[2] Bera S, Misra S, Rodrigues JJPC. Cloud computing applications for smart grid: a systems: current trends and new perspectives. Appl Energy 2020;272:115237.
survey. IEEE Trans Parallel Distrib Syst 2015;26(5):1477–94. doi:10.1016/j.apenergy.2020.115237.
[3] Shi W, Cao J, Zhang Q, Li Y, Xu L. Edge computing: vision and challenges. IEEE [35] Ndikumana A, Tran NH, Ho TM, Han Z, Saad W, Niyato D, et al. Joint communica-
Internet Things J 2016;3(5):637–46. tion, computation, caching, and control in big data multi-access edge computing.
[4] Satyanarayanan M. The emergence of edge computing. Computer 2017;50(1):30–9. IEEE Trans Mobile Comput 2019:1. doi:10.1109/TMC.2019.2908403.
[5] Postscapes. IoT edge computing software. https://www.postscapes.com/iot-edge- [36] Liu F, Tang G, Li Y, Cai Z, Zhang X, Zhou T. A survey on edge computing systems
computing-software/; 2019. and tools. Proc IEEE 2019;107(8):1537–62. doi:10.1109/JPROC.2019.2920341.
[6] Dolui K, Datta SK. Comparison of edge computing implementations: fog computing, [37] Zaballos A, Vallejo A, Selga JM. Heterogeneous communication architecture for
cloudlet and mobile edge computing. In: Proceedings of the 2017 Global Internet the smart grid. IEEE Netw 2011;25(5):30–7. doi:10.1109/MNET.2011.6033033.
of Things Summit; 2017. p. 1–6. [38] European Commission DE. The role of DSOS in a smart grid environment. Technical
[7] Pan J, McElhannon J. Future edge cloud and edge computing for In- Report; 2014.
ternet of things applications. IEEE Internet Things J 2018;5(1):439–49. [39] Mouftah HT, Erol-Kantarci M, Rehmani MH. Wiley; 2019. p. 217–45. null
doi:10.1109/JIOT.2017.2767608. [40] Green RC, Wang L, Alam M. Applications and trends of high performance com-
[8] Greentech. What is grid edge?- grid edge solutions are at the center puting for electric power systems: focusing on smart grid. IEEE Trans Smart Grid
of the electricity market’s evolution.https://www.greentechmedia.com/articles/ 2013;4(2):922–31. doi:10.1109/TSG.2012.2225646.
read/what-is-the-grid-edge; 2017. [41] Zhou Y, Ghosh S, Ali MH, Wyatt TE. Minimization of negative effects of time de-
[9] Molzahn DK, Drfler F, Sandberg H, Low SH, Chakrabarti S, Baldick R, et al. A survey lay in smart grid system. In: Proceedings of the IEEE Southeastcon; 2013. p. 1–6.
of distributed optimization and control algorithms for electric power systems. IEEE doi:10.1109/SECON.2013.6567474.
Trans Smart Grid 2017;8(6):2941–62. doi:10.1109/TSG.2017.2720471. [42] Liu J, Shou G, Liu Y, Hu Y, Guo Z. Performance evaluation of integrated
[10] Lu Y, Huang X, Zhang K, Maharjan S, Zhang Y. Communication-efficient federated multi-access edge computing and fiber-wireless access networks. IEEE Access
learning for digital twin edge networks in industrial IoT. IEEE Trans Ind Inform 2018;6:30269–79. doi:10.1109/ACCESS.2018.2833619.
2020;Early Access:1. doi:10.1109/TII.2020.3010798. [43] Moslehi K, Kumar R. A reliability perspective of the smart grid. IEEE Trans Smart
[11] Chiang M, Zhang T. Fog and IoT: an overview of research opportunities. IEEE In- Grid 2010;1(1):57–64. doi:10.1109/TSG.2010.2046346.
ternet Things J 2016;3(6):854–64. [44] Zhang K, Mao Y, Leng S, Zhao Q, Li L, Peng X, et al. Energy-efficient offloading for
[12] Cintuglu MH, Mohammed OA, Akkaya K, Uluagac AS. A survey on smart grid mobile edge computing in 5G heterogeneous networks. IEEE Access 2016;4:5896–
cyber-physical system testbeds. IEEE Commun Surv Tutor 2017;19(1):446–64. 907. doi:10.1109/ACCESS.2016.2597169.
doi:10.1109/COMST.2016.2627399. [45] Liu Y, Yang C, Jiang L, Xie S, Zhang Y. Intelligent edge computing for IoT-based
[13] Greer C, Burns MJ, Wollman DA, Griffor ER. Cyber-physical systems and energy management in smart cities. IEEE Netw 2019;33(2):111–17.
Internet of things. Special Publication (NIST SP) - 1900-202 2019;75:82–4. [46] ARM. CMSIS NN software library. https://www.keil.com/pack/doc/CMSIS/NN/
doi:10.1016/j.future.2017.05.040. html/index.html; 2019.
[14] Bonomi F, Milito R, Zhu J, Addepalli S. Fog computing and its role in the Internet [47] Singh S. Optimize cloud computations using edge computing. In: Proceedings of
of things. In: Proceedings of the first edition of the MCC workshop on mobile cloud the 2017 international conference on big data, IoT and data science (BID); 2017.
computing, MCC ’12. New York, NY, USA: Association for Computing Machinery; p. 49–53. doi:10.1109/BID.2017.8336572.
2012. p. 13–16. ISBN 9781450315197. doi:10.1145/2342509.2342513. [48] Wikipedia. FLOPS. https://en.wikipedia.org/wiki/FLOPS; 2020.
[15] Chiang M, Ha S, I C-L, Risso F, Zhang T. Clarifying fog computing and net- [49] Github. Estimates of memory consumption and FLOP counts for various convolu-
working: 10 questions and answers. IEEE Commun Mag 2017;55(4):18–20. tional neural networks. https://github.com/albanie/convnet-burden; 2020.
doi:10.1109/MCOM.2017.7901470. [50] Biookaghazadeh S, Ren F, Zhao M. Are FPGAs suitable for edge computing? CoRR
[16] Mach P, Becvar Z. Mobile edge computing: a survey on architecture and computa- 2018;abs/1804.06404.
tion offloading. IEEE Commun Surv Tutor 2017;19(3):1628–56. [51] Morabito R, Cozzolino V, Ding AY, Beijar N, Ott J. Consolidate IoT edge
[17] Taleb T, Samdanis K, Mada B, Flinck H, Dutta S, Sabella D. On multi-access edge computing with lightweight virtualization. IEEE Netw 2018;32(1):102–11.
computing: a survey of the emerging 5G network edge cloud architecture and or- doi:10.1109/MNET.2018.1700175.
chestration. IEEE Commun Surv Tutor 2017;19(3):1657–81. [52] Gedeon J, Brandherm F, Egert R, Grube T, Muhlhauser M. What the fog? edge
[18] Sabella D, Vaillant A, Kuure P, Rauschenbach U, Giust F. Mobile-edge computing computing revisited: promises, applications and future challenges. IEEE Access
architecture: The role of MEC in the Internet of things. IEEE Consum Electron Mag 2019;7:152847–78. doi:10.1109/ACCESS.2019.2948399.
2016;5(4):84–91. doi:10.1109/MCE.2016.2590118. [53] Chen M, Hao Y. Task offloading for mobile edge computing in software
[19] Abbas N, Zhang Y, Taherkordi A, Skeie T. Mobile edge computing: a survey. IEEE defined ultra-dense network. IEEE J Sel Areas Commun 2018;36(3):587–97.
Internet Things J 2018;5(1):450–65. doi:10.1109/JSAC.2018.2815360.
[20] Hu P, Dhelim S, Ning H, Qiu T. Survey on fog computing: architecture, key [54] Guo H, Liu J, Zhang J, Sun W, Kato N. Mobile-edge computation offload-
technologies, applications and open issues. J Netw Comput Appl 2017;98:27–42. ing for ultradense IoT networks. IEEE Internet Things J 2018;5(6):4977–88.
doi:10.1016/j.jnca.2017.09.002. doi:10.1109/JIOT.2018.2838584.
[21] Liu F, Tang G, Li Y, Cai Z, Zhang X, Zhou T. A survey on edge computing systems [55] S S, G S, S B. Joint optimization of radio and computational resources for multicell
and tools. Proc IEEE 2019;107(8):1537–62. doi:10.1109/JPROC.2019.2920341. mobile-edge computing. IEEE Trans Signal Inf Process over Netw 2015;1(2):89–
[22] Omoniwa B, Hussain R, Javed MA, Bouk SH, Malik SA. Fog/edge computing-based 103. doi:10.1109/TSIPN.2015.2448520.
IoT (FECIoT): architecture, applications, and research issues. IEEE Internet Things [56] K Z, Y M, S L, Q Z, L L, X P, et al. Energy-efficient offloading for mobile
J 2019;6(3):4118–49. doi:10.1109/JIOT.2018.2875544. edge computing in 5G heterogeneous networks. IEEE Access 2016;4:5896–907.
[23] Wang X, Han Y, Leung VCM, Niyato D, Yan X, Chen X. Convergence of edge doi:10.1109/ACCESS.2016.2597169.
computing and deep learning: a comprehensive survey. IEEE Commun Surv Tu- [57] Mao Y, Zhang J, Letaief KB. Dynamic computation offloading for mobile-
tor 2020;22(2):869–904. doi:10.1109/COMST.2020.2970550. edge computing with energy harvesting devices. IEEE J Sel Areas Commun
[24] Hou X, Li Y, Chen M, Wu D, Jin D, Chen S. Vehicular fog computing: a viewpoint 2016;34(12):3590–605. doi:10.1109/JSAC.2016.2611964.
of vehicles as the infrastructures. IEEE Trans Veh Technol 2016;65(6):3860–73. [58] Funai C, Tapparello C, Heinzelman W. Computational offloading for energy con-
[25] Liu J, Wan J, Zeng B, Wang Q, Song H, Qiu M. A scalable and quick-response soft- strained devices in multi-hop cooperative networks. IEEE Trans Mobile Comput
ware defined vehicular network assisted by mobile edge computing. IEEE Commun 2020;19(1):60–73. doi:10.1109/TMC.2019.2892100.
Mag 2017;55(7):94–100. [59] Wang J, Pan J, Esposito F, Calyam P, Yang Z, Mohapatra P. Edge cloud offloading
[26] Kumar N, Zeadally S, Rodrigues JJPC. Vehicular delay-tolerant networks for algorithms. ACM Comput Surv 2019;52(1):1–23. doi:10.1145/3284387.
smart grid data management using mobile edge computing. IEEE Commun Mag [60] Wang S, Tuor T, Salonidis T, Leung KK, Makaya C, He T, et al. Adaptive feder-
2016;54(10):60–6. ated learning in resource constrained edge computing systems. IEEE J Sel Areas in
[27] Perera C, Qin Y, Estrella JC, Reiff-Marganiec S, Vasilakos AV. Fog computing for Commun 2019;37(6):1205–21. doi:10.1109/JSAC.2019.2904348.
sustainable smart cities: a survey. ACM Comput Surv 2017;50(3):32:1–32:43. [61] Hou T, Feng G, Qin S, Jiang W. Proactive content caching by exploiting transfer
[28] He Y, Yu FR, Zhao N, Leung VCM, Yin H. Software-defined networks with mo- learning for mobile edge computing. In: Proceedings of the GLOBECOM 2017 -
bile edge computing and caching for smart cities: a big data deep reinforcement 2017 IEEE Global Communications Conference; 2017. p. 1–6. doi:10.1109/GLO-
learning approach. IEEE Commun Mag 2017;55(12):31–7. COM.2017.8254636.
16
[62] Chen B, Wan J, Lan Y, Imran M, Li D, Guizani N. Improving cognitive ability [82] Zhu G, Liu D, Du Y, You C, Zhang J, Huang K. Toward an intelligent edge: wireless
of edge intelligent IIoT through machine learning. IEEE Netw 2019;33(5):61–7. communication meets machine learning. IEEE Commun Mag 2020;58(1):19–25.
doi:10.1109/MNET.001.1800505. doi:10.1109/MCOM.001.1900103.
[63] Shang Y, Liu M, Shao Z, Jian L. Internet of smart charging points with [83] Zhang X, Wang J, Jiang C, Li H, Guo Q, Wang X, et al. Hardware-in-the-loop simu-
photovoltaic integration: a high-efficiency scheme enabling optimal dispatch- lation system for space information networks. J Commun Inf Netw 2017;2(4):131–
ing between electric vehicles and power grids. Appl Energy 2020;278:115640. 41. doi:10.1007/s41650-017-0046-2.
doi:10.1016/j.apenergy.2020.115640. [84] Huang T-e, Guo Q, Sun H. A distributed computing platform supporting power
[64] Tong J, Wu H, Lin Y, He Y, Liu J. Fog-computing-based short- system security knowledge discovery based on online simulation. IEEE Trans Smart
circuit diagnosis scheme. IEEE Trans Smart Grid 2020;11(4):3359–71. Grid 2017;8(3):1513–24. doi:10.1109/TSG.2016.2571442.
doi:10.1109/TSG.2020.2964805. [85] Li X, Huang Q, Wu D. Distributed large-scale co-simulation for IoT-aided smart grid
[65] Peng N, Liang R, Wang G, Sun P, Chen C, Hou T. Edge computing control. IEEE Access 2017;5:19951–60. doi:10.1109/ACCESS.2017.2753463.
based fault location in distribution networks by using asynchronous tran- [86] Lai K, Illindala M, Subramaniam K. A tri-level optimization model to mitigate co-
sient amplitudes at limited nodes. IEEE Trans Smart Grid 2021;12(1):574–88. ordinated attacks on electric power systems in a cyber-physical environment. Appl
doi:10.1109/TSG.2020.3009005. Energy 2019;235:204–18. doi:10.1016/j.apenergy.2018.10.077.
[66] Sirojan T, Lu S, Phung BT, Zhang D, Ambikairajah E. Sustainable deep learning at [87] Jin D, Li Z, Hannon C, Chen C, Wang J, Shahidehpour M, et al. Toward a cyber re-
grid edge for real-time high impedance fault detection. IEEE Trans Sustain Comput silient and secure microgrid using software-defined networking. IEEE Trans Smart
2018:1. doi:10.1109/TSUSC.2018.2879960. Grid 2017;8(5):2494–504. doi:10.1109/TSG.2017.2703911.
[67] Hui H, Ding Y, Song Y. Adaptive time-delay control of flexible loads [88] Cheng X, Lyu F, Quan W, Zhou C, He H, Shi W, et al. Space/aerial-assisted com-
in power systems facing accidental outages. Appl Energy 2020;275:115321. puting offloading for IoT applications: a learning-based approach. IEEE J Sel Areas
doi:10.1016/j.apenergy.2020.115321. Commun 2019;37(5):1117–29. doi:10.1109/JSAC.2019.2906789.
[68] Habault G, Lefranois M, Lemercier F, Montavont N, Chatzimisios P, Papadopou- [89] Zhang J, Chen B, Zhao Y, Cheng X, Hu F. Data security and privacy-preserving in
los GZ. Monitoring traffic optimization in a smart grid. IEEE Trans Ind Inform edge computing paradigm: survey and open issues. IEEE Access 2018;6:18209–37.
2017;13(6):3246–55. doi:10.1109/TII.2017.2742584. [90] Varentec. Edge of network grid optimization (ENGO).
[69] Sirojan T, Lu S, Phung BT, Ambikairajah E. Embedded edge computing for https://varentec.com/products/engo-v10/
real-time smart meter data analytics. In: Proceedings of the 2019 international [91] ABB. APM edge. https://new.abb.com/products/transformers/service/
conference on smart energy systems and technologies (SEST); 2019. p. 1–5. advanced-services/apm-edge;
doi:10.1109/SEST.2019.8849012. [92] Simense. The power of digitalization in manufacturing: Sinumerik edge.
[70] Malik FH, Lehtonen M. A review: agents in smart grids. Electr Power Syst Res https://new.siemens.com/global/en/markets/machinebuilding/machine-tools/
2016;131:71–9. doi:10.1016/j.epsr.2015.10.004. cnc4you/fokus-digitalisierung/sinumerik-edge.html;
[71] Guo S, Hu X, Guo S, Qiu X, Qi F. Blockchain meets edge computing: a distributed [93] Digital G.E.. Predix edge - asset connectivity and local intelligence to power your
and trusted authentication system. IEEE Trans Ind Inform 2020;16(3):1972–83. IIoT solution. https://www.ge.com/digital/iiot-platform/predix-edge;
doi:10.1109/TII.2019.2938001. [94] Landis+Gyr. Gridstream RF advanced metering solution.
[72] Li E, Zhou Z, Chen X. Edge intelligence: on-demand deep learning model co-infer- https://www.landisgyr.com/product/gridstream-rf-2/;
ence with device-edge synergy. CoRR 2018;abs/1806.07840. [95] China Southern Power Grid, China State Grid, China Mobile, China Telecom, China
[73] Samie F, Bauer L, Henkel J. Cham: Springer International Publishing; 2019. Unicom, Huawei. 5G deterministic network @ power industry series white paper
p. 21–42. ISBN 978-3-030-03640-9. i: demand, technology and practice. Technical Report; 2020.
[74] Zhang X, Biagioni D, Cai M, Graf P, Rahman S. An edge-cloud integrated solution [96] Jones N. How to stop data centres from gobbling up the worlds electricity. Nature
for buildings demand response using reinforcement learning. IEEE Trans Smart 2018;561:163–6. doi:10.1038/d41586-018-06610-y.
Grid 2021;12(1):420–31. doi:10.1109/TSG.2020.3014055. [97] Jalali F, Hinton K, Ayre R, Alpcan T, Tucker RS. Fog computing may help to
[75] Yaghmaee Moghaddam MH, Leon-Garcia A. A fog-based Internet of energy ar- save energy in cloud computing. IEEE J Sel Areas Commun 2016;34(5):1728–39.
chitecture for transactive energy management systems. IEEE Internet Things J doi:10.1109/JSAC.2016.2545559.
2018;5(2):1055–69. doi:10.1109/JIOT.2018.2805899. [98] Ahvar E, Orgerie A, Lbre A. Estimating energy consumption of cloud, fog
[76] Al Faruque MA, Vatanparvar K. Energy management-as-a-service over fog comput- and edge computing infrastructures. IEEE Trans Sustain Comput 2019:1.
ing platform. IEEE Internet Things J 2016;3(2):161–9. doi:10.1109/TSUSC.2019.2905900.
[77] McArthur SDJ, Davidson EM, Catterson VM, Dimeas AL, Hatziargyriou ND, Ponci F, [99] Tan T, Yan Z, Zou H, Ma K, Liu F, Zhao L, et al. Renewable energy harvesting and
et al. Multi-agent systems for power engineering applications part i: Concepts, ap- absorbing via multi-scale metamaterial systems for Internet of things. Appl Energy
proaches, and technical challenges. IEEE Trans Power Syst 2007;22(4):1743–52. 2019;254:113717. doi:10.1016/j.apenergy.2019.113717.
doi:10.1109/TPWRS.2007.908471. [100] Lin J, Pipattanasomporn M, Rahman S. Comparative analysis of auction mecha-
[78] Dabbaghjamanesh M, Kavousi-Fard A, Dong ZY. A novel distributed cloud-fog nisms and bidding strategies for P2P solar transactive energy markets. Appl Energy
based framework for energy management of networked microgrids. IEEE Trans 2019;255:113687. doi:10.1016/j.apenergy.2019.113687.
Power Syst 2020;35(4):2847–62. doi:10.1109/TPWRS.2019.2957704. [101] Fu Y, Han X, Baker K, Zuo W. Assessments of data centers for provision of frequency
[79] Wang S, Wang X, Wu W. Cloud computing and local chip-based dynamic eco- regulation. Appl Energy 2020;277:115621. doi:10.1016/j.apenergy.2020.115621.
nomic dispatch for microgrids. IEEE Trans Smart Grid 2020;11(5):3774–84. [102] Huang P, Copertaro B, Zhang X, Shen J, Lfgren I, Rnnelid M, et al. A review
doi:10.1109/TSG.2020.2983556. of data centers as prosumers in district energy systems: renewable energy inte-
[80] Gao G, Xiao M, Wu J, Zhang S, Huang L, Xiao G. DPDT: a differentially private gration and waste heat reuse for district heating. Appl Energy 2020;258:114109.
crowd-sensed data trading mechanism. IEEE Internet Things J 2020;7(1):751–62. doi:10.1016/j.apenergy.2019.114109.
doi:10.1109/JIOT.2019.2944107. [103] Yang T, Zhao Y, Pen H, Wang Z. Data center holistic demand response algorithm
[81] Marjanovic M, Antonic A, Zarko IP. Edge computing architecture for mobile crowd- to smooth microgrid tie-line power fluctuation. Appl Energy 2018;231:277–87.
sensing. IEEE Access 2018;6:10662–74. doi:10.1109/ACCESS.2018.2799707. doi:10.1016/j.apenergy.2018.09.093.
17

1 s2.0 S2666792420300068 Main

Uploaded by

Copyright:

Available Formats

1 s2.0 S2666792420300068 Main

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

1 s2.0 S2666792420300068 Main

Uploaded by

Copyright:

Available Formats

Advances in Applied Energy 1 (2021) 100006

Contents lists available at ScienceDirect

Advances in Applied Energy

Smart grid encounters edge computing: opportunities and applications☆

Reference Contribution Inspiration to EC in SG Scope

F. Bonomi et al.[14] Pioneering work to discuss EC SG is treated as a typical application. ✓ ✓ ✓ × ✓ × ×

T Taleb et al.[17] A survey focusing on EC’s fundamental Complement materails on enabling ✓ × ✓ ✓ ✓ ✓ ✓

Advances in Applied Energy 1 (2021) 100006

Fig. 2. Potential users of EC in SGs.

Fig. 3. The three-tier architecture for implementing EC in SGs.

Table 3 2.5.3. Task-level technology

Category Detailed description Enabling technologies

For Existing Heterogeneous Computing

Complex Programmable Logic Device (CPLD) 1M 1k

Fig. 4. (a) SG operations and businesses requiring accurate actions

Fig. 8. The approaches for EC to become sustainable

You might also like