You are on page 1of 6

Smart Greedy Distributed Allocation in Microgrids

Ilai Bistritz, Andrew Ward, Zhengyuan Zhou, Nicholas Bambos


Department of Electrical Engineering, Stanford University, Stanford, CA, 94305

Abstract—We consider a microgrid that consists of N providers undergoes an even more complex conversion: wind turbines
and B consumers. Each provider has a certain supply and each need to convert their AC in its natural frequency to DC and
consumer has a certain demand. The efficiency of transmitting back to AC in the grid’s frequency. This also incurs significant
energy between providers and consumers is modeled using a
bipartite graph G. Our goal is to maximize the amount of utilized losses. By exchanging energy locally without exporting it
energy using a distributed algorithm that each provider runs to the macrogrid, these conversion losses can be saved [8].
locally. We propose a non-cooperative energy allocation game, Both the intermittent availability and production properties
and adopt the best-response dynamics for this game as our of renewable energy sources motivate an infrastructure that
distributed algorithm. We prove that the best-response dynamics supports a direct exchange of electricity between different
converge in no more than N steps to one of at most N ! pure Nash
equilibria of our game. Despite the fact that some of these Nash consumers in the same local environment [9].
equilibria are suboptimal, we are able to prove that our algorithm Nevertheless, an infrastructure of transmission lines that
achieves near-optimal performance in “almost all” games. We do connects different consumers is not enough. Given such an
so by analyzing the best-response dynamics in a random game, infrastructure, different allocations of energy from providers
where the network is generated using a random model for the to consumers will result in significantly different efficiencies.
graph G. We prove that the ratio between the utilized energy
of our algorithm and that of the optimal solution converges to The optimal allocation depends on the profile of supplies
one in probability as B increases (and N is any function of B). and demands that changes with time. This calls for an effi-
Using numerical simulations, we demonstrate that our asymptotic cient matching between providers and consumers that uses
analysis is valid even for B = 10 consumers. as much of the instantaneous available energy as possible.
Index Terms—Smart grid and cities, Game theory, Distributed This matching is not one-to-one, as each provider can provide
algorithms, Resource allocation, Random games.
electricity for multiple consumers, and each consumer can
obtain electricity from several providers.
I. I NTRODUCTION
In practice, computing a providers-to-consumers matching
The smart grid [1]–[6] is a collection of ideas to make in such a local network is a challenging task. The supplies
the electricity grid more efficient by incorporating renewable and demands change with time (typically determined every
sources, smart homes, smart meters and connecting all of 15 or 60 minutes), so there is a need to solve the problem
them using a communication network. In the smart grid, repeatedly with different parameters for each time interval.
electricity is no longer provided by a single generator via The supplies and demands are private by nature, and collecting
long, lossy transmission lines; distributed generation (DG) is them in a centralized fashion is not only demanding in terms
a key element. Thanks to the rapid advances in renewable of communication and storage, but also violates reasonable
energy technology, solar panels can be affordably installed privacy requirements and is a serious vulnerability from a
on commercial and even residential buildings. This recent cybersecurity perspective [10]. Information about the quality
development gives rise to the concept of a microgrid: several of transmission lines is also naturally obtained by local mea-
sources and consumers that are, at least partially, independent surements and is not easily available at a central computer.
of the macrogrid. Microgrids are envisioned to be one of the Thus, any practical computation of a providers-to-consumers
key components that will enable smart cities, where energy matching has to be distributed. For these reasons, the study
management is a central issue [7]. of distributed optimization algorithms for the smart grid has
Renewable energy sources are intermittent by nature. Wind become a fruitful field of research over the last decade [11].
speed and direction, shadows and passing clouds, and other In this paper, we propose a game-theoretic approach to
random phenomena all cause the available energy from a solve the fundamental providers-to-consumers energy match-
renewable source to vary rapidly both in time and space. Due ing problem in a distributed fashion. This approach yields
to the spatial diversity, in a large network of many households a very simple and efficient greedy algorithm that does not
or buildings equipped with renewable energy sources, some require any coordination between different distributed energy
will fail to fully satisfy their demand and will serve as resources (DERs) and converges in linear time (in N ). Even
consumers, while others will have an excess of energy and more surprisingly, this greedy algorithm turns out to be smart
will serve as providers. Due to the temporal diversity, the - it achieves an arbitrarily close to optimal performance for a
providers from an hour ago can now become consumers and large enough number of consumers, for “almost all” networks.
vice versa. Moreover, solar panels produce direct current (DC), We emphasize that in contrast to the majority of the game-
and converting it to alternating current (AC) in order to send theoretic literature, we do not use game theory to model
it over the macrogrid incurs significant losses. Wind energy what rational agents will do. Instead, we program agents

978-1-5386-8088-9/19/$31.00 ©2019 IEEE

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 08,2021 at 05:12:39 UTC from IEEE Xplore. Restrictions apply.
with local rules (utility functions) and analyze the global II. S YSTEM M ODEL
performance maintained at the Nash equilibria of the game.
We consider a microgrid with a set of consumers B =
By proving that this global performance is satisfactory, we
{1, . . . , B} and a set of DERs N = {1, . . . , N }. During
establish the performance guarantees of our distributed al-
each period (e.g., 15 or 60 minutes), consumer b has demand
gorithm. This approach is generally known as utility design
Db ≥ 0 and DER n has supply Qn ≥ 0. It is possible that
[12]–[17]. Specifically, we study the structure of our proposed
DERs and consumers are the same type of entity, but at this
non-cooperative game, and show that best-response dynamics
particular hour, the first group has an excess of energy while
(BRD) always converge after all DERs have acted. Using a
the second requires energy. The demands can be based on
random model for the network, we prove that for almost all
either a consumer’s current electricity usage or a prediction of
games, BRD achieve an asymptotically optimal performance
electricity usage based on the typical pattern of this consumer.
in the number of consumers B, while N may follow any
Similarly, the supplies can be based on the current stored
arbitrary sequence. This happens despite the fact that bad Nash
energy (electric vehicles and batteries [26]), or predictions of
equilibria do exist in almost all networks. Finally, although B
renewable energy production for the upcoming period.
is naturally large in practical scenarios, simulations suggest
The power-loss coefficient, or the transmission efficiency,
that our analysis is valid for as few as B = 10 consumers.
between DER n and consumer b is denoted ηn,b , and naturally
A. Previous Work 0 ≤ ηn,b ≤ ηmax for each n, b, for some ηmax ≤ 1. The
amount of energy consumer b receives from DER n is denoted
Game theory is a leading candidate to be the underlying
en,b . The amount of energy DER n transmits to consumer b
theory for smart grid distributed algorithm design [6], [18]–
is denoted an,b . Due to the transmission line loss, we have
[20]. For an excellent overview of game theory for smart
en,b = ηn,b an,b for each n, b.
grid applications, see [21]. Our work strengthens this view
Our goal is to distributedly solve the following energy
by demonstrating the usefulness of random game analysis for
matching problem for the period (e.g., hour) in consideration:
smart grid networks. Such an analysis has been successfully
applied to wireless networks in [15], [22], but has yet to be B N
max+ b=1 n=1 en,b
introduced in the smart grid literature. An alternative approach {en,b }⊂R ∪{0}
N (1)
n=1 en,b ≤ Db
is to distribute the gradient computation over the network
(see [11]). While being more straightforward, this approach s.t.∀b, n : B en,b
b=1 ηn,b ≤ Qn
requires each node to communicate large vectors (of size N B
in our case) that include data from all over the network. where the first constraint represents the finite demand of each
There are various works dealing with the economical aspects consumer and the second constraint represents the finite supply
of local energy markets [1]–[4], [6], [23]. These works are of each DER.
focused on the prices the system converges to after an auction Note that some (or even most) of the transmission lines
has taken place. Our work is substantially different since we might not exist, which is represented by setting ηn,b = 0.
are interested in an energy-efficient matching of providers However, if ηn,b > 0, it does not necessarily mean that there
to consumers, and not in a market efficient matching. There is a direct transmission line between DER n and consumer b.
are plenty of scenarios where energy-efficiency is the main Transmitting electricity between DER n and consumer b might
concern. Essentially, this is the case for any scenario where involve transmitting it between different houses that serve as
users pay a fixed rate per unit of electricity. A virtual power relays. In a relay case, the value ηn,b will typically be smaller
plant (VPP, see [24]) with many scattered renewable energy than in a direct transmission. In short, ηn,b models the energy
sources will want to optimize its energy distribution to a vari- loss of transmitting energy from n to b, possibly incorporating
ety of costumers. Another scenario is a community microgrid different causes for these losses.
[25] where private households have a long-term agreement to It is likely that the optimal solution to (1) does not fill all
exchange excesses of renewable energy between them. demands {Db }. In this case, as in practice, consumers buy the
rest of their required electricity from the traditional macrogrid.
B. Outline They can also postpone their demand to the next period, which
The paper is organized as follows: in Section II we describe is suitable for non-immediate demands (such as operating a
the system model, formulate our objective, and describe our washing machine or charging an electric vehicle).
distributed algorithm. We also state our main theorem that The problem (1) is a linear program that can, in principal,
provides theoretical performance guarantees for our algorithm. be solved if all the parameters were available on the same
In Section III we formulate the energy allocation game for server. However, in a practical microgrid, especially at the
which we develop and analyze the BRD algorithm. Section scale of those anticipated to be part of future smart cities,
IV analyzes the performance of the random NE to which our such a centralized solution is infeasible. Since problem (1) is
algorithm converges in a random game. Section V provides solved every 15 (or 60) minutes with varying supplies and
simulation results that suggest the asymptotic analysis of our demands, it is highly desirable to avoid sending all these
algorithm is valid even for B = 10. Finally, Section VI parameters from all nodes to a centralized server. Tracking the
concludes the paper. quality of the transmission lines all over the microgrid is also

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 08,2021 at 05:12:39 UTC from IEEE Xplore. Restrictions apply.
N=2 B=4
chosen by a specific DER in Algorithm 1 is never canceled in
Q1=10kWh D1=5kW
order to enable a better transaction of another DER instead.
0.9
Lemma 6 also establishes that Algorithm 1 converges after N
0.8
1
D2=7kW
steps (one for each DER). The computational complexity of
0.8 Algorithm 1 is O (B) for each DER.
Q2=15kWh
0.7
D3=10kW
Our main theorem gives theoretical performance guarantees
0.3
for Algorithm 1. Both the algorithm and the proof of the
0.95
theorem below are based on a game-theoretic analysis that is
0.9
D4=3kW given in Section III and Section IV. Informally, the theorem
states that, for a large enough number of consumers (B)
Algorithm 1 is (arbitrarily) close to optimal in almost all
Fig. 1: An example for a network with N = 2 and B = 4. networks. The simulations in Section V complete the picture
by showing that “a large enough B” is approximately B = 10.
Algorithm 1 Smart Greedy Distributed Energy Allocation Theorem 1. Let ηn,1 , ..., ηn,B be i.i.d. variables with distri-
Asynchronously, each DER runs bution Fn,η (x), independently generated for each n. Let E ∗
1) Initialize Qres = Qn . be the optimal solution to (1) for the realization of {ηn,b }. Let
2) Find b∗ that maximizes  to Dres,b > 0.
 ηn,b subject E (NB ) denote the total received energy after all NB DERs
Db∗
3) Transmit Δ = min ηn,b∗ , Qres energy to consumer have run Algorithm 1, with any arbitrary scheduling. Then,
b∗ , so now Dres,b∗ = Dres,b∗ − Δηn,b∗ . for any 0 < ρ < 1 and for any arbitrary sequence NB
 
4) Update Qres = Qres − Δ. E (NB )
5) If Qres > 0 and maxDres,b > 0, then return to Step 2. lim Pr ≥ ρ = 1. (2)
b
B→∞ E∗
End Proof: The proof uses Lemma 8 to argue that each DER
n is highly likely to contribute almost ηmax Qn to E (NB ). The
details are omitted due to page constraints.
a burden. Another barrier is computational complexity, since Theorem 1 is applicable for many plausible models for
solving linear programs with thousands of variables (or more) the efficiencies in a network. For example, a link (direct or
in reasonable time is still a challenge. The demands disclose indirect) can exist between DER n and consumer b with
information about private households, so gathering all this probability pn . The efficiency of a link might be a function
information in a single place may violate their privacy. Last but of the length of the transmission line, the random number of
not least, cyber-security is one of the most important issues hops required to transmit energy from n to b, and so on.
in a fundamental network such as the grid [10]. Storing all
the information of the microgrid at the same place offers poor III. T HE E NERGY A LLOCATION G AME
information protection. Even worse, a central computer that Our algorithm is nothing but asynchronous best-response
computes the optimal solution will need to send commands to dynamics of a certain non-cooperative game. In this section,
all the components of the microgrid to implement this solution. we define the energy allocation game (EAG) and analyze its
This is a serious security vulnerability since such a central structure. Many possible non-cooperative games could have
computer has control over all the microgrid. made sense as a model. Our motivation to choose a specific
one is not to model what rational agents will do, but to find
A. Distributed Energy Allocation a good basis for a distributed algorithm that solves (1).
We propose a (smart) greedy distributed algorithm to solve Our designed game, defined as follows, is a very natural
Problem (1), described in Algorithm 1. In Algorithm 1, DERs candidate to distributedly optimize (1).
act asynchronously over time. On its turn, a DER chooses its
Definition 2. The Energy Allocation Game (EAG) is a
transmission profile in order to maximize the total received
normal-form game with N DERs as the players and B passive
energy for its consumers. With time, the consumers start
consumers. Let a = (a1, , ..., aN ) denote the strategy profile.
to gradually fill their demands. The energy consumer b has
The strategy space of DER n is the following set of vectors
received from DER n up to time (step) t is denoted en,b (t). 

The residual demand of consumer b at time t is defined as An = an | 0 ≤ ηn,b an,b ≤ Dres,b (a) , an,b ≤ Qres,n (a)
Dres,b (t)  Db − n en,b (t).
b
In order to run Algorithm 1, each DER only needs to know (3)
its own supply, the residual demands of the consumers it is where Dres,b (a) is the residual demand of consumer b in a
connected to and the efficiency of transmitting energy to each and Qres,n (a) is the residual supply of DER n in a. The utility
of them. No knowledge of the supplies of other DERs or other of DER n in a is
efficiencies in the network is required. Furthermore, DER n B
can actually transmit energy on its turn so its action is not only 
un (a) = ηn,b an,b . (4)
a message passing. Thanks to Lemma 6, an energy transaction b=1

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 08,2021 at 05:12:39 UTC from IEEE Xplore. Restrictions apply.
One of the most common solution concepts in game theory Q1=1 D1=1
0.9
is that of a Nash equilibrium. It is of great interest in our 1 1
case, since the pure Nash equilibria are the stable points of
best-response algorithms such as Algorithm 1. 1

Q2=1 D2=1
Definition 3. A strategy profile a∗n , a∗−n ∈ A
1 × ... × AN 0.9
Nash equilibrium (NE) if un a∗n , a∗−n ≥
2 2
is called

a pure


un an , a−n for all an ∈ An and all n ∈ N . Fig. 2: A network with a suboptimal NE.
The main motivation to use the BRD as our algorithm is
the fact that our game is a potential game, defined as follows.
It is well known that the BRD converge in potential games
Definition 4. G =< N , {An }n∈N , {un }n∈N > is a general- [27]. This is also true for generalized ordinal potential games,
ized ordinal potential game if there exists a function Φ : A1 × since as long as there exists a player that can improve its utility,
... × AN → R such that for each n ∈ N and for all an , an ∈ this player can also increase the potential function. After a
An and a−n ∈ A−n , if un (an , a−n ) , −un (an , a−n ) > 0 finite amount of such improvements, we must reach a strategy
then Φ (an , a−n ) − Φ (an , a−n ) > 0. profile in which all players cannot increase their utility by
The next lemma verifies that our game is indeed a general- switching an action, which is a NE by definition. This follows
ized ordinal potential game. since a bounded potential function cannot increase forever.
Unfortunately, in generalized ordinal potential games, a NE
Lemma 5. The EAG is a generalized ordinal potential game does not necessarily maximize the potential function. Hence,
with the potential function the convenient fact that our potential function is the objective
 N
B  B 
 N of Problem (1) does not help us. The potential function
E (a) = an,b ηn,b = en,b (5) increases only until a NE is reached, which does not have
b=1 n=1 b=1 n=1 to be the optimal solution of Problem (1). In Fig. 2 we depict
Proof: The proof is immediate since each utility appears a simple example where best-response can lead to a bad NE.
in Equation (5), and is omitted due to page constraints. The optimal solution in this case is a1,1 = 1, a1,2 = 0 and
In addition to being a generalized ordinal potential game, a2,1 = 0.1, a2,2 = 0.9 which results in e1 = 0.9 + 0.1 = 1
B N
the NE of our game have the following structure. This lemma and e2 = 0.9 · 0.9 = 0.81, so b=1 n=1 en,b = 1.81.
is the key for our random game analysis of Section IV. However, if DER 2 were to perform its best-response first,
it would have transmitted a1,2 = 1 and a2,2 = 0, which is a
Lemma 6. The BRD of an EAG converge in N steps (one B  N
NE with b=1 n=1 en,b = 1, where only 55% of the energy
for each DER) to a pure NE, out of at most N ! possible pure is utilized compared to the optimal solution.
NEs.
Proof: The key observation for this result is that the IV. R ANDOM E NERGY A LLOCATION G AME
residual demands {Dres,b } never increase. The proof follows We cannot guarantee in general to which of the at most N !
by induction and is omitted due to page constraints. NE of the EAG that Algorithm (1) will converge. From the
The lemma above makes the implementation of our game traditional game-theoretic perspective, this means that the price
much easier. Without this result, the DERs would have needed of anarchy of our game might be high. It is simply possible
to distributedly detect a NE which would delay the energy that the BRD will converge to a bad pure NE.
transmission to the consumers. Since each DER is guaranteed This pessimistic conclusion is in sharp contrast to the
to act only once before the NE is reached, it can transmit the striking fact that although examples like in Fig. 2 always exist,
actual energy immediately on its turn and does not need to they are very rare. In other words, while implementing the
worry that this transmission might not be part of a pure NE. BRD of our EAG on some network in practice, one is highly
The lemma above also states that each scheduling results in unlikely to encounter a bad NE. This can be easily verified in
a pure NE. This enables running Algorithm 1 anachronistically simulations, as we present in Section V.
without coordinating between players, since each player wants This discrepancy between the traditional game-theoretic
to act only once. Any random asynchronous mechanism will answer and what actually happens calls for a more subtle
generate one of the N ! possible schedulings. The following approach, which we employ in this section. We generate a
simple lemma suggests that one of these schedulings might network at random by generating {ηn,b } for each DER n
even lead to the optimal solution of Problem (1). independently, with a distribution Fn,η (x). We show that as B
increases, the probability that any scheduling leads to a near-
Lemma 7. The strategy profile that consists of the optimal
optimal solution increases. This will analytically establish that
solution of Problem (1) is a pure NE of the EAG.
the probability that BRD in a random network will result in a
Proof: If a DER could have improved its utility, the ob- bad NE vanishes as B → ∞. The idea of analyzing random
jective of Problem (1) would have increased without violating games over networks to achieve significantly better bounds
the constraints, which contradicts the optimality. was first introduced in [15].

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 08,2021 at 05:12:39 UTC from IEEE Xplore. Restrictions apply.
The fact that the BRD of the EAG converge after each player DER transmits all of its energy to a single consumer such that
has acted only once, proven in Lemma 6, is very useful for the overall utilized energy is maximized. The weight of each
the analysis of the random EAG. Thanks to this property, our possible connection for the matching also took into account
random model is equivalent to generating {ηn,b } at random the demand and supply of that connection. While achieving
sequentially, observing the realizations for each DER only on reasonable performance, the performance of Algorithm 1 is far
its turn. If any player had acted more than once before the NE, better. The convergence of Algorithm 1 is also much faster,
its second turn would have depended on the values of {ηn,b } as it requires N iterations while

the auction algorithm has an
that were determined on that player’s first turn. Exploiting this expected running time of O N 3 iterations.
property, the following lemma bounds the expected utility of An interesting special case of networks is when each
each DER over the randomness of its {ηn,b } values. DER is connected only to nearby consumers, using relatively
N B short, dedicated, high-quality transmission lines. In Fig. 3c we
Lemma 8. Let 0 < ρ < 1 and E (t)  n=1 b=1 en,b (t).
present the histogram of the ratio between the performance of
Choose any schedule where the DER acting on turn t is
 Qn(t) Algorithm 1 and that of the optimal solution for N = B = 50.
n (t). Let D = B1 Db and define B̃ (t) = (1−ρ)D . Also The demand of all consumers was Db = 1 and the supply
 
let ηn,(B−B̃(t)) be the B̃ (t) + 1 -largest variable among of all DERs was Qn = 1. In this experiment, we generated
ηn,1 , ..., ηn,B . Define the following deterministic time at random a bipartite graph with degree d = 11 for each
 t+1 DER. Each DER was connected to d = 11 neighbors using
ηmax t =1 Qn(t ) ηn,b = 1. Algorithm 1 converged to at least ρ = 0.9 of the
τ (ρ) = min t | B > ρ or t = N . (6)
optimal solution in all networks. When applied to this kind of
b=1 Db
networks, our results provide insight about how connected a
Then for all t ≤ τ (ρ) we have for a sufficiently large B that network should be in order to efficiently utilize a target portion
 
of the available energy in the microgrid.
E {E (t) − E (t − 1)} ≥ Qn(t) E ηn,(B−B̃(t)) . (7)

Proof: The proof uses the fact that regardless of the VI. C ONCLUSIONS
current Dres,b (t) profile, on expectation each consumer still We studied the problem of distributed energy allocation
needs D − Be given E (t − 1) = e. The details are omitted due between providers and consumers that are located in the same
to page constraints. geographical region. Our goal was to utilize as much local
V. S IMULATION R ESULTS energy as possible, minimizing the electricity bought from the
macrogrid. We have proposed a smart and greedy algorithm
The importance of our analysis of the random EAG is
that consists of the best response of each player in a designed
twofold. First, it explains why in almost all networks our
energy allocation game (EAG). Although the EAG may have
algorithm achieves near-optimal performance, despite the fact
a Nash equilibrium (NE) with bad performance, we have
that bad NE do exist. Second, it provides analytical perfor-
analyzed the performance of the best-response dynamics in
mance guarantees. The analysis is asymptotic in the number
a random EAG, and shown that it is asymptotically optimal in
of consumers B, but that does not mean that we require B
almost all games with respect to the number of consumers.
to be very large. It only means that the probability for good
In a more general scenario, the supplies and demands might
performance improves with B. Additionally, N does not have
only be available in an asynchronous manner and subject to
to be smaller than B, and can be any function of B (e.g.,
prediction errors. An interesting research direction is employ-
N = B). In this section, we numerically demonstrate the
ing the learning algorithms introduced in [29]–[31] to converge
performance of Algorithm 1, and show that our analytical
to a NE of the EAG even in this case.
findings already hold for B = 10.
For each instance of supplies and demands, some of the
In Fig. 3a we present the histogram of the ratio between the
supplies might not be utilized and some of the demands are not
performance of Algorithm 1 and that of the optimal solution
fully satisfied. Since energy storage is limited, only a certain
for N = 5 and B = 10. Both the demands and supplies
amount of excess can be stored for later usage or transmission.
were random integers between 1 and 5. The efficiency ηn,b
Additionally, some of the energy demand in a given time
between DER n and consumer b is modeled as 1 − X where
slot is deferrable and can be postponed to the next time slot.
X is an exponentially distributed variable with expectation 0.2.
This gives rise to a dynamic problem where the goal is to
The results are dramatic, as ~97% of the realizations resulted
maximize the utilized energy over time. We conjecture that
in more than 95% of the optimal solution. Our results are
our instantaneous algorithm could serve as a building block to
especially surprising considering the fact the bad NE exist
solve this dynamic problem.
in almost all realizations. Nevertheless, our analysis fully
explains this seemingly surprising result.
ACKNOWLEDGMENTS
For a performance comparison, in Fig. 3b we present the
ratio of the performance of the optimal one-to-one matching This research was supported by the Koret Foundation grant
for the same scenario. Such a matching can be computed dis- for Smart Cities and Digital Living and by the Bits & Watts
tributedly using the algorithm in [28]. In this algorithm, every Initiative at Stanford University.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 08,2021 at 05:12:39 UTC from IEEE Xplore. Restrictions apply.
250 10 400

9
350

200 8
300
7

250
150 6
Frequency

Frequency
Frequency
5 200

100 4
150

3
100
50 2

50
1

0 0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Ratio to Optimal Ratio to Optimal Ratio to Optimal

(a) (b) (c)


Fig. 3: The histogram of the ratio to optimal of: our performance (in (a)), the optimal one-to-one matching’s performance (in
(b)), for N = 5 and B = 10. Subfigure (c) shows the ratio of our performance and the optimal performance, for N = B = 50,
where each DER is only connected to d = 11 consumers. All histograms are over 1000 realizations.

R EFERENCES [16] Z. Zhou, N. Bambos, and P. Glynn, “Deterministic and stochastic


wireless network games: Equilibrium, dynamics, and price of anarchy,”
[1] H. K. Nunna and S. Doolla, “Demand response in smart distribution Operations Research, vol. 66, no. 6, pp. 1498–1516, 2018.
system with multiple microgrids,” IEEE transactions on smart grid, [17] Z. Zhou, P. Mertikopoulos, A. Moustakas, N. Bambos, and P. Glynn,
vol. 3, no. 4, pp. 1641–1649, 2012. “Robust power management via learning and game design,” Operations
[2] B. Ramachandran, S. K. Srivastava, C. S. Edrington, and D. A. Cartes, Research, 2020.
“An intelligent auction scheme for smart grid market using a hybrid im- [18] C. Ibars, M. Navarro, and L. Giupponi, “Distributed demand manage-
mune algorithm,” IEEE Transactions on Industrial Electronics, vol. 58, ment in smart grid with a congestion game,” in Smart grid communica-
no. 10, pp. 4603–4612, 2011. tions (SmartGridComm), 2010 first IEEE international conference on.
[3] P. Vytelingum, S. D. Ramchurn, T. D. Voice, A. Rogers, and N. R. Jen- IEEE, 2010, pp. 495–500.
nings, “Trading agents for the smart electricity grid,” in Proceedings of [19] B. Zhang, R. Johari, and R. Rajagopal, “Competition and coalition
the 9th International Conference on Autonomous Agents and Multiagent formation of renewable power producers,” IEEE Transactions on Power
Systems: volume 1-Volume 1. International Foundation for Autonomous Systems, vol. 30, no. 3, pp. 1624–1632, 2015.
Agents and Multiagent Systems, 2010, pp. 897–904. [20] W. Saad, Z. Han, and H. V. Poor, “Coalitional game theory for cooper-
[4] A. Ward, Z. Zhou, and N. Bambos, “Bidding-based dynamic power ative micro-grid distribution networks,” in Communications Workshops
pricing scheme in smart grids,” International Conference on Computing, (ICC), 2011 IEEE International Conference on. IEEE, 2011, pp. 1–5.
Networking and Communication (ICNC), 2019. [21] W. Saad, Z. Han, H. V. Poor, and T. Basar, “Game-theoretic methods
[5] Z. Wang, A. Scaglione, and R. J. Thomas, “Generating statistically for the smart grid: An overview of microgrid systems, demand-side
correct random topologies for testing smart grid communication and management, and smart grid communications,” IEEE Signal Processing
control networks,” IEEE transactions on Smart Grid, vol. 1, no. 1, pp. Magazine, vol. 29, no. 5, pp. 86–105, 2012.
28–39, 2010. [22] I. Bistritz and A. Leshem, “Approximate best-response dynamics in
[6] W. Saad, Z. Han, H. V. Poor, and T. Başar, “A noncooperative game random interference games,” IEEE Transactions on Automatic Control,
for double auction-based energy trading between phevs and distribution vol. 63, no. 6, pp. 1547 – 1548, 2018.
grids,” in Smart Grid Communications (SmartGridComm), 2011 IEEE [23] Y. Wang, W. Saad, Z. Han, H. V. Poor, and T. Başar, “A game-theoretic
International Conference on. IEEE, 2011, pp. 267–272. approach to energy trading in the smart grid,” IEEE Transactions on
[7] C. F. Calvillo, A. Sánchez-Miralles, and J. Villar, “Energy management Smart Grid, vol. 5, no. 3, pp. 1439–1450, 2014.
and planning in smart cities,” Renewable and Sustainable Energy Re- [24] D. Pudjianto, C. Ramsay, and G. Strbac, “Virtual power plant and
views, vol. 55, pp. 273–287, 2016. system integration of distributed energy resources,” IET Renewable
[8] D. J. Hammerstrom, “Ac versus dc distribution systemsdid we get it power generation, vol. 1, no. 1, pp. 10–16, 2007.
right?” in Power Engineering Society General Meeting, 2007. IEEE. [25] L. Che, M. Shahidehpour, A. Alabdulwahab, and Y. Al-Turki, “Hi-
IEEE, 2007, pp. 1–5. erarchical coordination of a community microgrid with ac and dc
[9] F. Hvelplund, “Renewable energy and the need for local energy markets,” microgrids,” IEEE Transactions on smart grid, vol. 6, no. 6, pp. 3042–
Energy, vol. 31, no. 13, pp. 2293–2302, 2006. 3051, 2015.
[10] J. Liu, Y. Xiao, S. Li, W. Liang, and C. P. Chen, “Cyber security [26] B. Dunn, H. Kamath, and J.-M. Tarascon, “Electrical energy storage for
and privacy issues in smart grids,” IEEE Communications Surveys & the grid: a battery of choices,” Science, vol. 334, no. 6058, pp. 928–935,
Tutorials, vol. 14, no. 4, pp. 981–997, 2012. 2011.
[11] D. K. Molzahn, F. Dörfler, H. Sandberg, S. H. Low, S. Chakrabarti, [27] D. Monderer and L. S. Shapley, “Potential games,” Games and economic
R. Baldick, and J. Lavaei, “A survey of distributed optimization and behavior, vol. 14, no. 1, pp. 124–143, 1996.
control algorithms for electric power systems,” IEEE Transactions on [28] O. Naparstek and A. Leshem, “Fully distributed optimal channel as-
Smart Grid, vol. 8, no. 6, pp. 2941–2962, 2017. signment for open spectrum access,” IEEE Transactions on Signal
[12] J. R. Marden and A. Wierman, “Distributed welfare games,” Operations Processing, vol. 62, no. 2, pp. 283–294., 2014.
Research, vol. 61, no. 1, pp. 155–168, 2013. [29] Z. Zhou, P. Mertikopoulos, A. L. Moustakas, N. Bambos, and P. Glynn,
[13] ——, “Overcoming the limitations of utility design for multiagent “Mirror descent learning in continuous games,” in 2017 IEEE 56th
systems,” IEEE Transactions on Automatic Control, vol. 58, no. 6, pp. Annual Conference on Decision and Control (CDC). IEEE, 2017, pp.
1402–1415, 2013. 5776–5783.
[14] N. Li and J. R. Marden, “Designing games for distributed optimization,” [30] Z. Zhou, P. Mertikopoulos, N. Bambos, P. W. Glynn, and C. Tomlin,
IEEE Journal of Selected Topics in Signal Processing, vol. 7, no. 2, pp. “Countering feedback delays in multi-agent learning,” in Advances in
230–242, 2013. Neural Information Processing Systems, 2017, pp. 6171–6181.
[15] I. Bistritz and A. Leshem, “Game theoretic dynamic channel allocation [31] P. Mertikopoulos and Z. Zhou, “Learning in games with continuous
for frequency-selective interference channels,” IEEE Transactions on action sets and unknown payoff functions,” Mathematical Programming,
Information Theory, 2018, DOI: 10.1109/TIT.2018.2868440. vol. 173, no. 1-2, pp. 465–507, 2019.

Authorized licensed use limited to: INTERNATIONAL ISLAMIC UNIVERSITY. Downloaded on May 08,2021 at 05:12:39 UTC from IEEE Xplore. Restrictions apply.

You might also like