IEEE Computational Intelligence Magazine - November 2019 - 47

it can run on an arbitrary connected communiADMM is able to solve large-scale optimization
cations topology. This makes DC-ADMM more
communication-efficient and scalable. The selecproblems by decomposing them into multiple small
tion of agents' neighbors, i.e., B i, will influence
subproblems that agents can handle locally.
the ADMM convergence performance. Generally,
increasing the numbers of neighbors speeds up
convergence but causes higher overall communimeters. For large agents such as microgrids and wind farms,
cations cost. The selection of B i will be further discussed in
DCSs can be implemented as virtual machines that run on
the following sections.
edge servers. In addition, a center would be needed to manage
DCS's authentication, algorithm parameter tuning, and
B. ADMM-Based Distributed Intelligence
software update, guaranteeing efficient collaboration among
In energy management, system-wide optimization requires
massive DCSs in the distributed computing environment
global information from all elements in the system. Howevof IoE.
er, the energy devices, resources, and infrastructures in a system may be possessed by different agents. Agents are
independent of each other and unwilling to share their priIII. Examples of ADMM Applications in IoE
vate data. ADMM is a promising solution to this issue. As
This section provides some examples of ADMM applications in
introduced previously, ADMM is able to solve large-scale
practical distributed energy management in IoE. The examples
optimization problems by decomposing them into multiple
range from the transmission level, distribution level, to consumsmall subproblems that agents can handle locally. The local
er level, which are illustrated in Fig. 5.
computation allows an agent's private data not to be disclosed to others, protecting the privacy of agents. Moreover,
A. Transmission-Level Applications
since agents collectively participate in the computation in
parallel, ADMM can achieve moderately accurate conver1) Gas-Power Co-Scheduling
gence in a relatively small number of iterations. For convex
Due to the environmental benefit and relatively low price of
programs, ADMM has a sublinear convergence rate of
natural gas, Natural Gas-fired Units (NGUs) have been increasingly installed for electricity generation. A large number of
O (1/t ) [25]. Due to these advantages, we propose using
NGUs give rise to the necessity of coordinative scheduling of
ADMM to guide the design of distributed computational
both gas and power. As shown in Fig. 5(a), NGUs consume gas
intelligence in IoE infrastructures.
from the gas network and generate electricity in the power
Fig. 4 illustrates the proposed distributed computational
network; NGUs are coupling points between the two energy
intelligence architecture, in which each agent relates to a Disnetworks. The goal would be to achieve optimal scheduling for
tributed Computing System (DCS). The DCS is a software
both gas and power while guaranteeing the autonomy of the
system dedicated to local computing (i.e., variable update) in
two energy network operators. The co-scheduling problems
the ADMM framework. In Fig. 4, an EV agent is exampled,
can be solved using standard ADMM [19], [26], in which the
whose DCS has a database storing and managing the EV's pritwo energy operators act as agents and the shared data between
vate data such as battery states, power consumption profiles
them only relates to the gas-power conversion of NGUs. Each
(from sensors), locations, preference parameters (set by drivers),
operator is still independent in managing its own energy infraetc. An agent's private data would be included in its cost funcstructures. There is no need to send gas-only (power-only) data
tion f i($) and local constraint set X i . In the ADMM frameto the power (gas) operator.
work [see (3) and (7)], f i($) and X i are only used for the x i
update which is performed by agent i itself. Thus, private data
2) Optimal Power Flow
is only used for local computing in a DCS without being sent
Optimal power flow is a typical type of energy management
to any other. The computation of a DCS also requires data of
problems in power transmission systems. The considered scecoupling variables from other DCSs, and this data refers to
nario is shown in Fig. 5(b), in which the goal would be to
shared data which can be known by coupling agents. In the
coordinate a set of power generation units to match the loads
case of Fig. 3(b), the physical meaning of the shared data n i is
on buses at minimum cost. In past research, a bus or a region is
energy price. It is shown that the privacy of an agent is prousually considered as a subsystem (i.e., agent), and subsystems
tected via the DCS. The actuator (e.g., EV charger) is responsive
are coupled via transmission lines [13], [18], [27]. Besides, we
to the computational results for executing power control. Difcan consider a more general scenario where a bus is connected
ferent applications can be installed on a DCS for different
with generation units possessed by different companies. In
energy management goals.
Fig. 5(b), we use colors to distinguish the companies. In this
Some agents may only require relatively small amounts
case, each company corresponds to an agent in the ADMM
of data storage and computing resources. For these small
framework, and the coupling among agents is associated with
agents, DCSs can be installed on intelligent terminals such
transmission lines as well as power balance on buses. Standard
as EVs' on-board computer systems and buildings' smart

NOVEMBER 2019 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

IEEE Computational Intelligence Magazine - November 2019

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2019