Autonomous Energy Grids

Preprint
Benjamin Kroposki, Emiliano Dall’Anese,
Andrey Bernstein, Yingchen Zhang, and
Bri-Mathias Hodge
National Renewable Energy Laboratory
To be presented at the Hawaii International Conference
on System Sciences
Waikoloa, Hawaii
January 3–6, 2018

NREL is a national laboratory of the U.S. Department of Energy

Office of Energy Efficiency & Renewable Energy
Operated by the Alliance for Sustainable Energy, LLC
This report is available at no cost from the National Renewable Energy
Laboratory (NREL) at www.nrel.gov/publications.

Conference Paper
NREL/CP-5D00-68712
October 2017

Contract No. DE-AC36-08GO28308

 NOTICE

The submitted manuscript has been offered by an employee of the Alliance for Sustainable Energy, LLC
(Alliance), a contractor of the US Government under Contract No. DE-AC36-08GO28308. Accordingly, the US
Government and Alliance retain a nonexclusive royalty-free license to publish or reproduce the published form of
this contribution, or allow others to do so, for US Government purposes.
This report was prepared as an account of work sponsored by an agency of the United States government.
Neither the United States government nor any agency thereof, nor any of their employees, makes any warranty,
express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of
any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately
owned rights. Reference herein to any specific commercial product, process, or service by trade name,
trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation,
or favoring by the United States government or any agency thereof. The views and opinions of authors
expressed herein do not necessarily state or reflect those of the United States government or any agency thereof.

This report is available at no cost from the National Renewable Energy

Laboratory (NREL) at www.nrel.gov/publications.

Available electronically at SciTech Connect http:/www.osti.gov/scitech

Available for a processing fee to U.S. Department of Energy

and its contractors, in paper, from:

U.S. Department of Energy

Office of Scientific and Technical Information
P.O. Box 62
Oak Ridge, TN 37831-0062
OSTI http://www.osti.gov
Phone: 865.576.8401
Fax: 865.576.5728
Email: [email protected]

Available for sale to the public, in paper, from:

U.S. Department of Commerce

National Technical Information Service
5301 Shawnee Road
Alexandria, VA 22312
NTIS http://www.ntis.gov
Phone: 800.553.6847 or 703.605.6000
Fax: 703.605.6900
Email: [email protected]

Cover Photos by Dennis Schroeder: (left to right) NREL 26173, NREL 18302, NREL 19758, NREL 29642, NREL 19795.

NREL prints on paper that contains recycled content.

 Autonomous Energy Grids
Benjamin Kroposki Emiliano Dall’Anese Andrey Bernstein Yingchen Zhang Bri-Mathias Hodge
National Renewable National Renewable National Renewable National Renewable National Renewable
Energy Laboratory Energy Laboratory Energy Laboratory Energy Laboratory Energy Laboratory
Benjamin.kroposki Emiliano.DallAnese Andrey.Bernstein Yingchen.Zhang Bri.Mathias.Hodge
@nrel.gov @nrel.gov @nrel.gov @nrel.gov @nrel.gov

Abstract resources (DERs) (solar, fuel cells, microturbines,

gensets), distributed energy storage (batteries, ice
Current frameworks to monitor, control, and storage), and new loads (electric vehicles, light-
optimize large-scale energy systems are becoming emitting diode lighting) are being added to electric
increasingly inadequate because of significantly high grids and causing bidirectional power flows and
penetration levels of variable generation and voltage fluctuations that impact optimal control and
distributed energy resources being integrated into system operation. In addition, because of increased
electric power systems; the deluge of data from numbers and types of sensors, massive amounts of new
pervasive metering of energy grids; and a variety of data are being collected on energy grid conditions.
new market mechanisms, including multilevel ancillary
services. This paper outlines the concept of
autonomous energy grids (AEGs). These systems are
supported by a scalable, reconfigurable, and self-
organizing information and control infrastructure, are
extremely secure and resilient (self-healing), and can
self-optimize in real time to ensure economic and
reliable performance while systematically integrating
energy in all forms. AEGs rely on cellular building
blocks that can self-optimize when isolated from a
larger grid and participate in optimal operation when
interconnected to a larger grid. This paper describes
the key concepts and research necessary in the broad Figure 1. Energy systems integration occurs at a
domains of optimization theory, control theory, big variety of physical scales.
data analytics, and complex system theory and The use of natural gas-fired generation has also
modeling to realize the AEG vision. increased significantly, both at the bulk level and
1. Introduction locally through combined heat-and-power applications.
Finally, new ideas such as islanded microgrids [2] are
Energy systems generate, transport, convert, and being used locally to disconnect parts of an energy
consume energy. They encompass a wide range of system from a larger system for economic reasons and
domains, including electric power systems, thermal to improve customer reliability and resilience. The
systems used for heating and cooling, and fuel systems formation of microgrids can be seamless when there is
such as natural gas or hydrogen networks. In addition, no interruption to customer loads or they can isolate
a large number of interrelated domains influence the from the grid with a discontinuity of service.
operation of these energy systems, including
communications, water, and transportation networks. Current grid control systems that operate at the bulk-
Energy systems can function at a variety of physical system level typically control about 10,000 points even
scales, from a small individual consumer, to on the largest systems. As additional smart,
communities and cities, to larger regions that controllable devices are integrated into the grid, the
encompass transmission networks. Figure 1 shows number of control points could easily reach hundreds
these interrelationships at a variety of physical scales of millions, significantly increasing the complexity of
[1]. how to control and optimize the system. All these new
technologies (Figure 2) are increasing the complexity
Current power systems deliver electricity primarily of energy systems to a point at which existing
in one direction: flowing from large central plants to techniques to monitor, control, and optimize them will
customer loads; however, increasing amounts of be inadequate. For example, existing techniques might
variable generation (wind and solar), distributed energy not offer decision-making capabilities that are

1
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 consistent with the form and function of future large- capabilities along with comprehensive situational
scale systems, which will be governed by faster awareness. Currently, energy systems of significant
dynamics, include heterogeneous energy assets that are size make a large number of simplifications and rely on
controllable at different temporal resolutions, and the underlying physics of the systems to operate
accommodate a variety of deregulated energy markets properly. A further challenge is associated with the
[3]. This paper proposes the concept of autonomous range of timescales that are important to the real-time
energy grids (AEGs) and describes the research needed operation of energy grids. In power systems, this can
to fully exploit new techniques that enable their secure, range from ultrafast protection schemes, through
resilient, and economic operation. Similar to transient and dynamic stability in the seconds range, to
autonomous vehicles—which do not require a physical power flow and unit commitment, which can be hours
driver and can make decisions on how to most and days.
effectively transport a person from one place to
As more flexibility [4], [5] is added to energy
another—AEGs do not require physical operators, can
systems, the complexity of controlling and optimizing
be extremely secure and resilient (self-healing), and
these systems will increase to a point at which they
can self-optimize in real time to ensure economic and
will be too difficult to manage using conventional
reliable performance while integrating energy in all
techniques. Future energy systems will need techniques
forms. To achieve these goals, AEGs rely on scalable
that fully use big data analytics and advances in
cellular blocks that can self-optimize when isolated
optimization, control, and complex system theories. As
from a larger grid and participate in optimal operation
the grid evolves, there will also be challenges to
when interconnected to a larger grid. These scalable
operation caused by the ownership of various
cells (Figure 2) can be areas of the grid that can run
controllable assets. Because distributed generation,
independently as microgrids or be parts of the grid that
storage, and local loads might be under the ownership
are segregated from a control perspective, but they do
of customers instead of power system operators, how
not have enough local generation to carry the full load
these devices participate in real-time operations will
of the cell. The AEG concept allows for the use of
need to be further defined to ensure grid stability and
optimization and control across cells in cases when the
reliability.
cells can form independent microgrids and when they
can control assets but not intentionally island. 2. Real-Time Optimization
This section focuses on real-time optimization
methods for AEGs. The term real time refers to an
operational setting in which the power set points of the
DERs within each cell are updated on a second or
subsecond timescale to maximize the operational and
economic objectives while coping with the variability
of ambient conditions and noncontrollable energy
assets [4] and achieving intercell coordination to
ensure reliable system-wide operation. The main
challenge in this context is related to the solution of
pertinent optimization problems at a timescale that
matches the dynamics of the prevailing ambient
conditions, noncontrollable assets, and changes in the
AEGs’ configurations. In fact, solving optimization
problems to convergence (i.e., in a batch setting) every
second or every few seconds is currently impractical
because of the following challenges:
(c1) Complexity and convergence rate. For large-scale
Figure 2. Energy systems incorporate a variety of grids, computational complexity might prevent the
energy sources and can form cells.
solution of optimization problems at the required
A critical factor that makes the operation of energy timescales [3], [6]. When an optimization problem is
grids challenging is that underlying problems solved in a distributed fashion (e.g., with cell-to-cell
associated with control, optimization, and monitoring communications as well as intracellular message
tasks need to be solved in real time and in a distributed passing), multiple communication rounds are necessary
fashion. Energy is constantly being produced and used, to converge to (possibly optimal) solutions.
and this balance requires fast decision-making

2
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 (c2) Model inaccuracy. Approximate linear models or Capitalizing on this powerful mathematical model,
convex relaxation methods might be leveraged to first-order (gradient-descent based) methods can be
derive convex problems that facilitate the design of used to design online algorithms to track optimal
computationally affordable solutions; however, solution trajectories of the formulated time-varying
approximate/relaxed convex problems might involve problems; however, existing online optimization
only an approximate representation of a system’s methods implicitly operate in an open-loop
physics and constraints, and therefore the optimal configuration and might still suffer the challenges
solutions of the convex problem might not be feasible noted above in (c2)–(c3). A powerful way to address
for the original problem. (c2)–(c3) is to suitably modify the first-order methods
to accommodate appropriate measurements from the
(c3) Pervasive metering. Solving optimization
energy grids and DERs [10], [11], [14], [16]. The
problems such as relaxations/linearizations of the AC
resultant feedback-based online optimization methods
optimal power flow (OPF) requires pervasive metering
can cope with inaccuracies in the representation of the
to collect measurements of the noncontrollable loads at
AC power flow and avoid pervasive metering to gather
all locations in real time, which might be impractical.
the state of noncontrollable resources. These
Regarding (c1), note that the optimization tasks optimization algorithms also naturally afford a
related to AEGs are markedly different from traditional distributed implementation by leveraging the
settings wherein energy systems are optimized at the decomposition of the Lagrangian function associated
wholesale level using economic- and market-based with the optimization problems. Particularly, the
objectives. The modus operandi in bulk systems— decomposition of the Lagrangian function can lead to a
wherein a few large-scale generators need to be distributed message passing that entails:
dispatched and the noncontrollable net load varies
(i) Communications between a cell-level control
slowly—is not compatible with the form and function
platform and individual customers, which are
of AEGs with massive integrations of DERs or when
necessary to optimize customer-level objectives while
optimization models require accurate representations of
respecting electrical limits within a cell; and
the AC power flows as well as the controllability
region of the DERs. In these cases, optimization (ii) Message passing among cells to optimize the flow
problems are nonconvex, nondeterministic polynomial- of power based on economic and reliability targets.
time hard (NP-hard), and therefore they are infeasible
to solve at the envisioned timescale [3], [6].
Nonconvexity often implies that distributed solution
methods might not be provably convergent to Karush–
Kuhn–Tucker points. To address (c1), and to facilitate
the development of provably stable and optimal
distributed solution methods for AEGs, a first step is to
develop convex relaxations and linear approximations
of pertinent nonconvex problems [17], [18]. Section
2.1 explains how to appropriately modify existing
relaxation/approximation methods to ensure the
feasibility of the produced solutions. The next steps
propose developing real-time optimization algorithms
that resolve the challenges in (c1)–(c3); these are
described in the ensuing subsections.
2.1 Feedback-Based Online Optimization Figure 3. Communications architecture for
distributed and real-time optimization of AEGs
To capture time-varying operation and economic
Figure 3 illustrates this architecture, wherein
objectives and constraints, the time-varying
communications among cells occurs because of
optimization formalism [9], [10] is leveraged.
consistency constraints for electrical quantities that
Specifically, pertinent time-varying convex
pertain to the cell-to-cell connections—for example,
optimization problems can model well-defined
adjacent cells agree on the real and reactive power
objectives and constraints of DERs located within each
exchanges at the points of interconnection. Intracellular
cell as well as consistency constraints for electrical
communication is used to ensure that the set points of
quantities that pertain to the cell-to-cell connections.
the DERs are computed to maximize the given
operational objectives while ensuring that electrical

3
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 limits are satisfied. See, e.g., [10] for intracellular the theory for iterative methods that involve gradients
communication and [19] for an example of cell-to-cell with errors can be leveraged. In fact, it can be shown
consensus-based message passing. As shown in, e.g., that a packet loss leads to the computation of primal or
[10], optimality and stability results for these online dual gradient steps with outdated information [30]. A
algorithms can be established to track the solution of plug-and-play operation wherein cells can switch from
the formulated time-varying optimization problems. an islanded mode to a larger grid-connected mode
should be ensured; this can be modeled as a time-
The idea of accommodating measurements into
varying constraint in the underlying optimization
optimization methods goes back to, e.g., [7], [8]; and it
problem. Similarly, flexible operation, wherein a cell
was applied to the real-time optimization of power
(or a portion of an cell) switches to an autonomous
systems in [11], wherein a centralized controller was
control setting during a prolonged communications
developed based on measurement-based projected-
outage, should be enabled.
gradient methods. Online algorithms were developed in
[10], [12], [13], and [14] to find solutions to AC OPF Last, distributed optimization will cross-fertilize
problems; the authors of [10] established results to with the design of (albeit futuristic) market
track solutions of a time-varying linearized AC OPF, mechanisms to systematically account for
and the authors of [14] tracked solutions of a time- payment/rewards to exchange energy exchange and
varying relaxed AC OPF. Recently, a projected- provision ancillary services among autonomous cells.
gradient method on the power flow manifold was This is in the spirit of transactive energy frameworks
proposed in [15], [16]. Although [11], [14], [15], [16] [25], which will need to be considerably broadened to
can be implemented in a centralized controller, [10] accommodate the proposed architecture and to enable a
affords “star” communications between cell-level rigorous mathematical analysis of system stability and
controllers and DERs. optimality. For example, the stability analysis in [31]
will be extended to account for cell-to-cell interactions.
2.2 Enabling Real-Time Optimization of AEGs
To enable real-time optimization of AEGs, the
3. Control Theory to Ensure Cell Stability
approaches in [10]–[16] should be broadened to and Synchronization
accommodate a communication architecture that AEGs pose significant challenges in terms of
includes cell-to-cell and cell-to-customer message optimal operation and the analysis of their stability.
passing. This can be obtained by dualizing the This is particularly the case when distributed or
consensus constraints in the optimization problem that decentralized control algorithms are used to operate the
are used to ensure that adjacent cells agree on the system (as discussed in Section 2) because these
power flows from one cell to another. Overall, the algorithms are inherently asynchronous as a result of
resultant feedback-based online optimization methods communications delays, losses, and distributed
need to provably track the solution of the convex (asynchronous) control actions. The typical approach
optimization problems by modeling well-defined to stability analysis involves analyzing a continuous-
objectives and constraints of each cell as well as the time system of differential equations; however, for
consistency constraints for electrical quantities that networked systems with digital controllers, this
pertain to the cell-to-cell connections. It is worth standard analysis naturally disregards computational
emphasizing that the design of the distributed and communications latencies as well as asynchronous
algorithm as well as the overall communications actions. Another challenge arises because cells must
strategy will depend on the type of actors participating operate autonomously when they are isolated and
in the real-time optimization process—i.e., end cannot rely on the frequency and voltage signals from
customers, cell controllers, or aggregators that the larger grid.
participate in the overall optimization process while
retaining controllability of their own assets. To overcome these difficulties, a multi-timescale
stability analysis is used. We illustrate the idea using
The algorithm operates in an asynchronous way to the following multi-timescale control architecture:
account for different communications latencies and for
devices that can be controlled at different timescales • On a slow timescale (e.g., every 5–15 minutes),
(for example, inverter-interfaced devices are controlled optimal droop coefficients are computed by a
at fast timescales, whereas thermostatically controlled central entity in the cell (using, e.g., the
loads are controlled every few minutes). Resilience to methodology of [20]).
communications drops and asynchronous operation • On a faster timescale (e.g., every second or tens of
should be analytically established through pertinent seconds), optimal power set points are computed
input-to-state stability and tracking results. To this end, using the methods discussed in Section 2.

4
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 • On a fast timescale (e.g., every 50–100 msec), the and temporal characterizing; ii) state estimation and
power set points of the controllable assets in the forecasting; iii) autonomous decision-making.
cell are adjusted using the proportional droop
Effectively characterizing the system operation
controller.
status in both the spatial and temporal domains using
• In real time, the inverters track the given power set big data analytics [26] is the critical first step toward
point as closely as possible, and they ensure phase knowledge discovery. Big data analytics in this sense
and voltage synchronization in case of islanded includes advanced concepts such as artificial
operation. intelligence and machine learning techniques to help
The idea is that the stability of the system needs to understand the data and make critical decisions in real
be ensured by design. To this end, the following time. Although the high-volume and high-velocity
considerations need to be made: nature of big data means that many dimensions of
information can be extracted, it also brings challenge to
(i) The computation of the optimal droop coefficients processing these data because of possible overfitting.
on the slow timescale must ensure stability of the Therefore, in the spatial domain, clustering the system
resulting dynamic system on the fast timescale. and using optimal sensor selection is critical. The
Because the time step of the fast-timescale operation is ultimate goal of is to be able to fully observe the
more than 20 milliseconds, a quasi-static system under any conditions using interrelated
approximation can be used—in between the set point information to enhance the accuracy and speed of
update, it is assumed that the system is in its steady knowledge discovery.
state and phasor-based power-flow analysis is
applicable. This yields a discrete-time dynamic system It is understandable that in a networked energy
that can be analyzed using, e.g., standard bounded- system, there are close correlations among cells. Only
input/bounded-state notions [20]. a limited number of sensors within each tightly
correlated cell are needed to fully monitor the entire
(ii) To design and analyze real-time controllers, a cell. As an example, secondary voltage control (SVC)
continuous-time methodology can be used to ensure [27] is one way to group such information in each cell.
the stability and synchronization of the cell. The If a power system is represented by a linearized model
current state of the art provides several alternatives in of the power flow equations, the voltage coupling of
this context, such as [21], [22], and [23]. The control the two buses can be used to define the relative electric
framework in [23] is a promising approach to enable distance between any two buses. If the reactive power
low-inertia AEGs to operate in a plug-and-play mode variance at two buses causes coupled voltage variance
while stabilizing cells around set points produced by at those buses, the relative electric distance given this
the real-time optimization algorithms described in definition is small. Therefore, the buses that have small
Section 2. relative electric distances can be clustered in one SVC
Note that these timescales are typically analyzed cell using a method such as the κ-means clustering
separately in the literature. This results in suboptimal algorithm. Within each SVC cell, the minimum
strategies that do not use the interrelationships between number of sensors can be determined for full
the timescales; hence, a new multi-timescale control observability in state estimation and forecasting. Figure
framework is needed to design optimal and stable 4 shows an example of SVC clustering and optimal
controllers for cellular energy grids. sensor selection. The IEEE 39-bus system is clustered
into 7 areas (divided by blue lines), and each area has a
4. Big Data Analytics for Energy Systems limited numbers of sensors (red dots) to ensure full
Increasing amounts of heterogeneous sensor data observability.
and information are becoming available in energy grids In the temporal domain, although many variables
from sources such as smart meters, distributed can be useful to extract the features of any given
generation, and smart home energy management system, the time-frequency information is most critical
systems. Being able to collect, curate, and create for many energy system applications. Methods such as
actionable information with these data will be critical matching pursuit decomposition (MPD) [28] are very
to creating AEGs. AEGs will need to be able to digest effective for signal time-frequency analysis.
data, evaluate data, and make decisions faster than in Implemented with the Gaussian atom dictionary [29], it
real time in both centralized and distributed settings. can be used to characterize big data represented by the
The big data analytics that are needed for the amplitude, time shift, frequency shift, and variance of
envisioned AEG involve three critical steps: i) spatial Gaussian atoms. The basic premises of using MPD for
big data analytics in energy systems are as follows:

5
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 (i) Any signal can be represented using decomposition Relevant machine learning and signal processing
with finite iterations and a limited residual. problems for AEGs should be implemented by using
distributed algorithms. Similar to Section 2, once the
(ii) The group of parameters that are identified by
problem is modeled via a well-defined optimization
MPD for any signal can be used for correlation
problem, primal-dual-gradient-type methods can be
analysis and can therefore overcome data overfitting.
used to design distributed and online algorithms to
track optimal solution trajectories of inferential
problems and machine learning problems.
5. Complex Systems Theory and Modeling
Energy systems include all energy domains
(electricity, fuels, thermal) as well as other domains
(communications, water, transportation) that influence
how energy is generated, distributed, converted, and
used. The connections among these domains create
interdependencies that challenge overall system design,
planning, control, and optimization. Historically, little
attention has been paid to the overlap among these
areas, but AEGs will need to be able to account for
Figure 4. Secondary voltage control clustering and these interdependencies because of the increasing
optimal sensor selection using IEEE 39-bus system impacts that each energy and infrastructure domain has
on another. Similarly, most energy system modeling
Being able to estimate and forecast the system’s true
methods are not conducted on a full time spectrum,
states under different aspects of grid operation
wherein changes in transient and dynamic behavior
(including steady states, dynamic and transient
influence steady-state performance. New techniques
conditions) in an online setting (real time) are also
are needed to address the uncertain and stochastic
essential elements toward automatic decision-making.
nature of variable resources and consumer choices
To handle nonlinear model analysis, a variety of across the full time spectrum of operation, from
machine learning-based approaches are used to analyze milliseconds to years.
and forecast system information (e.g., electrical price,
To accurately model the interactions among such
load, voltage deviation, congestion). When forecasting
disparate energy systems, new modeling approaches
a relatively small number of signals, standard statistical
are needed that move beyond the single modeling
methods and human-driven variable selection are
formalism techniques that are most prevalent when
sufficient for many applications (for example,
representing single energy domains. Multi-paradigm
relatively simple forecasting models can predict
modeling allows for a domain-independent framework
upcoming electrical demand at the level of an entire
that aids in more accurate multisystem representations
region); however, if a large amount of data from
through multiple formalisms, multiple levels of
heterogeneous sources is considered, then even
abstraction, and meta-modeling [24].
determining the proper correlations and features to use
becomes a challenging problem. Also, the scalability The choice of formalism is an important issue in
and adoptability of a single location’s forecast can modeling and simulation. Often a certain formalism is
achieve good results in certain scenarios or a certain used for a particular problem because it seems the most
time period, but it is not conveniently applied or natural or because of historical precedence. For
generalized to other scenarios or different time periods. example, in power systems the unit commitment and
economic dispatch problem is naturally solved through
Therefore, new techniques in machine learning—
mathematical programming (optimization) techniques;
including deep neural networks, multikernel learning,
however, mathematical programming would not be a
Monte Carlo tree searches, dimensionality
natural fit to solve the dynamic flow of natural gas
reduction/data reconstruction, matrix completion, and
through a pipeline because this involves solving the
reinforcement learning—are needed to help estimate
partial differential equations of the Navier-Stokes
and forecast system variables that can decide true
equations. Thus, if the goal is to model the interactions
operation states.
between these two systems, a multi-formalism
Innovations in probabilistic decision networks and approach that allows each subsystem to be represented
conditional data analytic structures for energy grids are in its most natural form is needed.
the keys to enabling autonomous decision-making.

6
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 Using multiple levels of abstraction is another necessary to simulate AEGs. The design and operation
important concept in modeling and simulating complex of AEGs will require further advancements in
systems. This is often performed without prior modeling complex systems because the systems are too
planning because of data or computational time large, complex, and costly to create physical
limitations. For example, a power system planning representations. Because much of the design will occur
generation expansion model would not include details through computational simulations, further
of the electromagnetic transients in a generator because advancements in modeling and simulating these
(among other reasons) simulating these subsecond systems will be necessary to ensure that they are safe,
phenomena make the simulation of the yearly or economically efficient, reliable, and robust.
decadal time steps of the generation planning decisions
When considering coupled energy infrastructures,
computationally infeasible. Therefore, these processes
the coupling factors among energy carriers are
are currently modeled separately. However, there are
oftentimes assumed constant. This is the case for the
times, such as planning for contingency events, when
fuel-to-power and fuel-to-heat conversion efficiencies
considering both of these aspects in an integrated
for combined-heat-and-power units, for example.
fashion would lead to more optimal decisions or a
However, a number of coupling factors are, in fact,
better understanding of the trade-offs involved in
nonlinear: examples include the efficiency and power
system design and operation.
consumed by a variable-speed water pump, which are
Therefore, to answer some of the more complex nonlinear functions of the pump frequency. Further, the
questions about how AEGs would be designed and operational region of some type of cogeneration units
operated requires the ability to model different levels or absorption and compression chillers might be
of abstraction in the same model. Often, this is to be nonconvex. Multisystem modeling might introduce
able to answer different questions about the same sources of nonconvexity via bilinear or trilinear terms
system under varying scenarios, and thus dynamically appearing in equality constraints, which might render
changing the level of model abstraction is a significant the optimization tasks outlined in Section 2 hard to
modeling capability. Figure 5 provides an example of solve. It is thus apparent that trade-offs between the
how multi-timescale models of the electricity system complexity of the modeling approach and achieving
could be linked together to make the computation flexibility must be taken into account in the system
tractable. design and operational processes.
6. Integration to Solve the Challenges of
AEGs
To develop AEGs that are scalable, reconfigurable,
and self-organizing, four research areas (Figure 6) have
been discussed that, when properly integrated, will
play a key role in how future energy systems operate.
AEGs rely on scalable, self-configuring cellular
building blocks that ensure that each “cell” can self-
Figure 5. Diagram of how multiple models would optimize when isolated from a larger grid as well as
interact at different timescales and fidelities partake in the optimal operation of a larger grid when
interconnected.
The concept of meta-modeling [32] is very useful for
integrating the different levels of model abstraction and A fundamental underpinning of this system is the
modeling formalisms together when considering ability to accurately model the cellular building blocks
complex multi-energy systems. Meta-modeling is the and their interactions with the rest of the systems so
process of defining the rules and constructs necessary that control, optimization, and forecasting methods
for creating models. This allows for creating tools that might be applied in operation. Complex system theory
can automatically convert models from one formalism therefore serves as a means for modeling and
to another or automatically create a model at another simulating the different energy domains and their
level of abstraction. This allows using the same basic interactions. These models can then be used for real-
simulation data at various levels of aggregation and in time optimization and control of the systems and
multiple simulation formalisms, depending on the subsystems, using information gained from big data
current required level of modeling fidelity and the analytics to provide forecasts that serve as parameters
questions under study. Meta-modeling is critical for in the control and optimization algorithms as well the
integrating the different types of models that are algorithmic computational awareness to apply regime-
switching approaches.

7
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 8. References
[1] M. O’Malley, B. Kroposki, B. Hannegan, H.
Madsen, M. Andersson, W. D’haeseleer, M.
McGranaghan, C. Dent, G. Strbac, S. Baskaran, M.
Rinker, “Energy Systems Integration: Defining and
Describing the Value Proposition”, NREL Technical
Report, NREL/TP-5D00-66616, June 2016
[2] B. Kroposki, R. Lasseter, T. Ise, S. Morozumi, S.
Papatlianassiou, and N. Hatziargyriou, “Making
Microgrids work”, IEEE Power and Energy magazine,
Vol.6, Issue 3, May-June 2008
Figure 6. Integration of four technical areas needed [3] J. A. Taylor, S. V. Dhople, and D. S. Callaway,
to solve the challenges of AEGs “Power systems without fuel,” Renewable &
Sustainable Energy Reviews, vol. 57, pp. 1322–1336,
In addition to these four areas of basic research,
May 2016
other important aspects will need further refinement to
achieve the full vision of AEGs. New smart grid [4] E. Dall’Anese, P. Mancarella, and A. Monti,
devices such as the power electronic interfaces of DER “Unlocking Flexibility: Integrated Optimization and
inverters will need some embedded autonomous Control of Multienergy Systems”, IEEE Power and
functionality to ensure rapid response to grid anomalies Energy Magazine, Vol. 15, issue 1, Jan-Feb. 2017
without communications. This would ensure that the
[5] J. Cochran, M. Miller, O. Zinaman, M. Milligan, D.
devices can respond to voltage and frequency
Arent, B. Palmintier, M. O'Malley, S. Mueller, E.
fluctuations quickly enough to damp grid disturbances.
Lannoye, A. Tuohy, B. Kujala, M. Sommer, H.
To ensure that all these smart devices work together in
Holttinen, J. Kiviluoma, S. K. Soonee, “Flexibility in
a reliable and resilient fashion, interoperability and
21st Century Power Systems. 21st Century Power
cybersecurity must be embedded into the devices and
Partnership”, NREL/TP-6A20-61721, 2014
used across the power system. Finally, because the
power system operates using both reliability and [6] S. Paudyaly, C. A. Canizares, and K. Bhattacharya,
economics as mechanisms to ensure stable supply of “Three phase distribution OPF in smart grids:
power at equitable costs, the ownership of devices used Optimality versus computational burden,” in 2nd IEEE
in future grid operations and how compensation will PES Intl. Conf. and Exhibition on Innovative Smart
flow to devices that provide a range of grid services Grid Technologies, Manchester, UK, Dec. 2011.
will need to be further defined. Addressing all of these
[7] A. Jokic, M. Lazar, and P. Van den Bosch, “Real-
concerns will be a multiyear effort, but it could provide
time control of power systems using nodal prices,” Intl.
a power system that will operate in an autonomous
J. of Electrical Power & Energy Systems, vol. 31, no.
fashion that can self-optimize for both reliability and
9, pp. 522–530, 2009.
economics while improving resilience, security, and
the ability to integrate energy in all forms. [8] J. Wang and N. Elia, “A control perspective for
centralized and distributed convex optimization,” in
7. Acknowledgements Proc. of 50th IEEE Conf. on Decision and Control,
This work was supported by the U.S. Department of Orlando, FL, Dec. 2011.
Energy under Contract No. DE-AC36-08GO28308 [9] A. Simonetto and G. Leus, “Distributed
with Alliance for Sustainable Energy, LLC, the asynchronous time varying constrained optimization,”
Manager and Operator of the National Renewable in 48th Asilomar Conference on Signals, Systems and
Energy Laboratory. The U.S. Government retains and Computers, Nov 2014, pp. 2142– 2146.
the publisher, by accepting the article for publication,
acknowledges that the U.S. Government retains a [10] E. Dall’Anese and A. Simonetto, “Optimal power
nonexclusive, paid-up, irrevocable, worldwide license flow pursuit,” IEEE Trans. on Smart Grid, May 2016.
to publish or reproduce the published form of this [11] A. Bernstein, L. Reyes Chamorro, J.-Y. Le
work, or allow others to do so, for U.S. Government Boudec, and M. Paolone, “A composable method for
purposes. real-time control of active distribution networks with
explicit power set points. part I: Framework,” Electric
Power Systems Research, vol. 125, no. August, pp.
254–264, 2015.

8
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.
 [12] L. Gan and S. H. Low, “An online gradient Control for Islanded Three-Phase Microgrids," in IEEE
algorithm for optimal power flow on radial networks,” Journal of Photovoltaics, vol. 4, no. 1, pp. 387-395,
IEEE Journal on Selected Areas in Communication, Jan. 2014.
vol. 34, no. 3, pp. 625–638, March 2016.
[23] M. Colombino, D. Gross, and F. Dörfler. “Global
[13] E. Dall’Anese, S. V. Dhople, and G. B. Giannakis, phase and voltage synchronization for power inverters:
“Photovoltaic inverter controller seeking ac optimal a decentralized consensus-inspired approach.” In
power flow solutions,” IEEE Transactions on Power Proceedings of the 56th IEEE Conference on Decision
Systems, Volume: 31, Issue: 4, July 2016 and Control, March 2017.
[14] Y. Tang, K. Dvijotham, and S. Low, “Real-time [24] P. Mosterman and H. Vangheluwe, “Computer
optimal power flow,” IEEE Trans. on Smart Grid, Automated Multi-Paradigm Modeling: An
2017, DOI: 10.1109/TSG.2017.2704922 Introduction”, SIMULATION, Vol.80, Issue9,
September 2004, pp.433-450
[15] A. Hauswirth, S. Bolognani, G. Hug, and F.
Dorfler, “Projected gradient descent on Riemannian [25] “GridWise Transactive Energy Framework”,
manifolds with applications to online power system PNNL Technical Report no. PNNL-22946.
optimization,” in 54th Annual Allerton Conference on
[26] H. Jiang, X. Dai, D. Gao, J. Zhang, Y. Zhang, and
Communication, Control, and Computing, Sept 2016,
E. Muljadi, “Spatial-Temporal Synchrophasor Data
pp. 225–232.
Characterization and Analytics in Smart Grid Fault
[16] A. Hauswirth, A. Zanardi, S. Bolognani, G. Hug, Detection, Identification, and Impact Causal Analysis”,
and F. Dorfler, “Online optimization in closed loop on IEEE Transactions on Smart Grid, Vol. 7, Issue 5,
the power flow manifold,” in 12th IEEE PES September 2016, pp. 2525-2536.
PowerTech conference, 2017.
[27] J. L. Sancha, J. L. Fernandez, A. Cortes, and J. T.
[17] J. Lavaei and S. H. Low, “Zero duality gap in Abarca, “Secondary voltage control: Analysis,
optimal power flow problem,” IEEE Transactions on solutions and simulation results for the Spanish
Power Systems, vol. 27, no. 1, pp. 92–107, 2012. transmission system,” IEEE Transactions on Power
Systems, vol. 11, no. 2, pp. 630–638, May 1996.
[18] A. Bernstein, C. Wang, E. Dall’Anese, J.-Y. L.
Boudec, and C. Zhao, “Load-flow in multiphase [28] S. G. Mallat and Z. Zhang, “Matching pursuits
distribution networks: Existence, uniqueness, and with time-frequency dictionaries,” IEEE Trans. Signal
linear models,” arXiv preprint arXiv:1702.03310, Process., vol. 41, no. 12, pp. 3397–3415, Dec. 1993.
2017.
[29] S. S. Chen, D. L. Donoho, and M. A. Saunders,
[19] E. Dall'Anese, S. V. Dhople, B. B. Johnson, and “Atomic decomposition by basis pursuit,” SIAM
G. B. Giannakis, “Decentralized Optimal Dispatch of Journal on Scientific Computing, vol. 20, no. 1, pp.
Photovoltaic Inverters in Residential Distribution 33–61, 1998.
Systems," IEEE Transactions on Energy Conversion ,
[30] N. Gatsis and G. B. Giannakis, “Residential load
vol. 29, n. 4, pp. 957-967, Dec. 2014.
control: Distributed scheduling and convergence with
[20] K. Baker, A. Bernstein, C. Zhao, and E. lost AMI messages,” IEEE Transactions on Smart
Dall’Anese, “Network-cognizant design of Grid, vol. 3, no. 2, pp. 770–786, 2012.
decentralized Volt/VAR controllers,” in The Eighth
[31] X. Zhou, E. Dall’Anese, L. Chen, and A.
Conference on Innovative Smart Grid Technologies
Simonetto, “An Incentive-based Online Optimization
(ISGT 2017), 2017.
Framework For Distribution Grids,” IEEE transactions
[21] S. D’Arco, J.A. Suul, O. B. Fosso, “A Virtual on Automatic Control, 2017. [Online]
Synchronous Machine implementation for distributed https://arxiv.org/abs/1705.01482
control of power converters in SmartGrids”, in Electric
[32] J. de Lara and H. Vangheluwe, “AToM3: A Tool
Power Systems Research, 122, pages 180-197, 2015.
for Multi-Formalism and Meta-Modelling” Lecture
[22] B. B. Johnson, S. V. Dhople, J. L. Cale, A. O. Notes in Computer Science, Volume 2306, pp. 174-
Hamadeh and P. T. Krein, "Oscillator-Based Inverter 188, 2002.

9
This report is available at no cost from the National Renewable Energy Laboratory at www.nrel.gov/publications.