fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 1
AbstractDemand-side management presents significant ben- We show that this model can effectively reduce the monetary
efits in reducing the energy load in smart grids by balancing expenses of all users in the real-time market, while at the same
consumption demands or including energy generation and/or time it can provide a confident production cost estimate to
storage devices in the users side. These techniques coordinate the
energy load so that users minimize their monetary expenditure. the energy supplier. Our analysis also shows that the added
However, these methods require accurate predictions in the computational cost of these error calculations is minimal and
energy consumption profiles, which make them inflexible to real the monetary gain is significant. Moreover, the calculations are
demand variations. In this paper we propose a realistic model amenable to a distributed implementation within the commu-
that accounts for uncertainty in these variations and calculates nications network of the smart grid. This is important because
a robust price for all users in the smart grid. We analyze the
existence of solutions for this novel scenario, propose convergent distributed algorithms are beneficial in terms of communica-
distributed algorithms to find them, and perform simulations tion efficiency and computation scalability [1].
considering energy expenditure. We show that this model can To further support our results we perform simulations in
effectively reduce the monetary expenses for all users in a real- a real-time modeled scenario. We also compare our robust
time market, while at the same time it provides a reliable proposal with a naive version where users do not take into
production cost estimate to the energy supplier.
account price deviations, but where the energy supplier still
Index TermsLoad management, robust analysis, non-convex establishes a robust price. Finally, we also present results
optimization, distributed algorithms, game theory
illustrating the convergence behavior of our algorithm.
This study can be of special interest to energy supply
I. I NTRODUCTION authorities who can analyze how much energy they should
MART grids represent the concept behind the intended
S evolution of the electric grid, which through information
acquisition from the end users will allow for more efficient
account for if they plan for grid load estimation errors and
peak demand costs. Even though users may be only interested
in their monetary expenditure as energy consumers, supply au-
energy distribution, flexibility in network topology, adaptive thorities have to account for local deviations in the distribution
load management and better integration of renewable energy network and coordinate with all agents to provide unobstructed
plants with the consumption requirements of the users. service by controlling regular and peak supply stations. The
One important concept to achieve the smart grids objec- fact that an agent in the network can alter the expected energy
tives is the demand-side management, which includes the load by injecting or subtracting energy unexpectedly is a
techniques for better energy efficiency programs, distributed source of extra costs that have to be taken into account by
energy generation, energy storage and load management. The suppliers that monitor the grids energy demand. Therefore,
model presented in this work establishes a robust energy price these energy provision costs justify a modified pricing model
one day ahead, which takes into account the production costs and a worst-case analysis of the demand-side network.
of energy as well as uncertainties in the expected energy
loads. We also consider a real-time market, where users are
charged their specific demanded energy at the robust price, A. Related Work
permitting some deviation without any extra penalty. This The work presented here is focused on the study of the
model provides some flexibility to every user in their real- demand-side management framework through noncooperative
time consumption energy loads and calculates the energy game theory. We build our results from [2], where the authors
prices in a conservative manner, taking into account production consider a day-ahead scheduling algorithm and the users
costs and possible errors. In particular, we focus on a robust optimize energy generation and storage profiles in a distributed
analysis of the demand-side management algorithms, where manner. Our model extends [2] by adding error terms to
the energy strategies account for worst-case error deviations the expected consumption profiles (energy loads in the smart
in the predicted loads. In our formulation, these error terms grid) and admitting variations in the estimated values with a
vary from user to user and depend on the energy load profiles. bounded error. In addition, we analyze the effect that these
error terms produce in the expected energy demand. In order
All authors are affiliated with the Universidad Politecnica de Madrid, in
Av. Complutense 30, Madrid, 28040, Spain. to do so, we extrapolate some techniques from reference [3],
Emails: javier.zazo.ruiz@upm.es,{santiago,sergio}@gaps.ssr.upm.es where a robust algorithm for communications is developed.
This work was supported in part by the Spanish Ministry of Science Regarding a real-time energy market, several models have
and Innovation under the grant TEC2013-46011-C3-1-R (UnderWorld), the
COMONSENS Network of Excellence TEC2015-69648-REDC and by an been proposed in the literature. For instance, the authors
FPU doctoral grant to the first author. from reference [4] propose a real-time cost function with
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 2
penalization terms in which the users minimize the expected introduce the robust game and analyze its solution through an
monetary expenses. The energy production model assumes that equivalent formulation. We also present a distributed algorithm
the cost of energy is given by a quadratic function of the total that solves the game and analyze its convergence properties.
demand, and therefore, the whole energy price of the smart Finally, in Section IV we present simulations to illustrate the
grid is affected by the real-time variations. In such setting, validity of our results.
the energy price is predetermined one day ahead, but it does
not account for the unpredictable production costs of real- II. S MART G RID M ODEL
time demand. The penalty functions should account for these In this section we introduce the smart grid model, the
extra expenses, but they are also predetermined in advance. If types of users in the demand-side of the network, the energy
they are set too costly, they would affect greatly the monetary model, the pricing plan and the error terms. We consider a
expenses of users in the real-time market. On the other hand, general framework with arbitrary convex regions and convex
if the penalties are set too low, they would cause an economic objectives.
loss to the energy supplier. Our robust model, on the contrary,
sets up an energy price that will allow the energy supplier to
produce energy without incurring unplanned extra costs, and A. Demand-Side Model
will allow users flexibility in their energy demand. The smart grid topology can be divided in three parts,
Another real-time model is introduced by reference [5], namely the supply-side, a central unit or regulation authority,
which proposes an optimization problem in which users de- and the demand-side. The supply-side involves the energy
mand energy according to a personal utility function. In this producers, as well as the distribution network. The central unit
scenario prices are updated according to production costs, but coordinates the optimization process. The demand-side entails
demand is fully elastic, within some predetermined bounds, the individual end-users as consumers, possibly with means
and the problem optimizes a social equilibrium point, rather to generate and/or store energy and/or vary their consumption
than accounting for individual energy requirements. Further- profiles. The demand-side may also include users with great
more, the energy supplier may experience monetary losses impact on the smart grid, such as industries with high energy
since the pricing mechanism does not cover production costs. demands or opportunistic agents who participate as local
In contrast, our algorithm accommodates individual energy energy suppliers in the network.
requirements and ensures that production costs are covered In this paper, the demand-side management focuses on
by the pricing mechanism. calculating optimal policies for monetary savings on the con-
The authors of reference [6] propose a robust scenario sumers side, while accounting for worst-case deviations from
for a real-time market based on linear programming, which the consumption profiles due to unpredictable considerations
allocates energy load profiles within a time frame. Because in the demand-side. The optimization process involves all
of the uncertainty of prices in real-time, users decide when users together with the central unit, who exchange information
to consume energy according to some modeling parameters, for a more efficient use and distribution of energy. This
price range estimates and energy requirements. Although the communication is accomplished with smart meters, which are
algorithm is robust in the sense that it assumes a range of real- the devices in charge of the bidirectional transmission of data
time energy prices, the model considers that prices are given with the central unit and the optimization process.
by the energy supplier without taking into account production We assume that all end users indicate their intended amount
costs. Our robust model, on the other hand, establishes the of energy consumption one day ahead, so that energy prices
prices one day ahead considering possible worst-case real-time can be calculated based on demand levels during the opti-
variations, and takes into account production costs. mization process. Then, the users can readjust their generation,
A relevant survey describing some of the last results and storing and consumption profiles according to the variability of
problems of the smart grid infrastructure can be found in [7], prices. This process is repeated iteratively until convergence.
which analyzes different contributions and open problems in Algorithms that analyze this process have been studied in [2]
the topic, where demand-side-management is one example of and [10]. However, due to errors in their estimations, the
these. Another publication that models the interaction among real consumption may vary from the announced values. If
resident users and distributed energy resources is the one these variations are uncorrelated among users, sometimes the
introduced by [8], but it considers prefixed energy generation energy left unused by some consumers will be spent by others
and storage load profiles, as opposed to our framework, inducing that, on average, the supply-side will not require to
which can adapt the profiles depending on the energy prices. adjust their generation rates. However, this expected result
Regarding user behavior, the authors of reference [9] propose a may not always be unconditionally true. In addition, there is
model that takes into account the subjective behavior of users, interest from the supply-side to provide energy for unexpected
but still does not consider imprecise load predictions. events that may require higher demands. Therefore, we present
Therefore, our contribution lies in considering a day-ahead analytical results on identifying worst-case energy errors, for
energy market and minimizing the energy costs of users in a given confidence interval.
a robust situation where energy demand errors are taken into We represent the set of users who participate in this process
account. In Section II we describe the demand-side model, with D, and additionally subdivide them into the set of active
the day-ahead cost function, the real-time market and briefly users N and the set of passive users P. Active users have some
describe the day-ahead optimization process. In Section III we means of generating energy on demand (from dispatchable
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 3
X 2
sources), possess devices to store energy for future use or Ch (L(h), (h)) , Kh L(h) + n (h) (2)
adapt their consumption profiles. On the contrary, passive nD
users do not adjust their energy consumption profiles and their |H|
consumption is estimated from historical data. where the coefficients {Kh }h=1 > 0 represent the energy
The consumption values are determined one day ahead cost at each time slot determined by the supply-side. In
among all users, and are subdivided in time-slots h H = addition, the individual cost for user n D at time slot h
{1, . . . , |H|}. We define the estimated per-slot energy load is
its proportional part of the
P Ch (L(h), (h))
total cost, i.e.,
profile ln (h) that indicates the estimated energy usage of user (ln (h) + n (h))/(L(h) + nD n (h)) , so that its cumu-
n D at time-slot h. We also represent these variables as lative extension for all time-slots is given by:
row vectors ln = [ln (1), . . . , ln (|H|)] R|H| , n D. X
Additionally, let l = [l1 , . . . , l|D| ] R|H||D| denote the global fn (l, ) , Kh L(h)+1T (h) (ln (h)+n (h))+M (n )
vector with the energy loads of all users. Each individual hH
term ln includes the energy consumption of user n and the (3)
where ln is a convex function; ln ln ; all regions ln
energy contributions or added expenditures that may decrease
are convex, compact and independent among users; and 1 is
or increase the energy load due to the user owning some
a column vector of appropriate size. Additionally, we have
device to generate or store energy, respectively. Naturally,
included a penalization term M (n ) that includes management
these profiles are variables to the optimization process and will
or distribution costs, which accounts for the extra expenses
vary according to the different strategies the user can select.
that the supply authority has to account for during the pricing
It is satisfied that ln (h) > 0 if the energy flows from the grid
process. We separate these penalization terms from the energy
to the user, and ln (h) < 0 otherwise, when the user is selling
prices due to the inherent unpredictable nature of the energy
energy to the grid.
demand, which is an additional consideration with regards to
We assume that the total load in the networkaggregate per-
the models in references [2], [10].
slot energy loadat time slot h denoted by L(h) is positive:
X Since we consider an arbitrary form of ln and ln , our ro-
0 < L(h) , ln (h) (1) bust proposal can be adapted to different demand-side models.
nD For instance, in the model from [2], ln has a linear expression
and it is calculated by the central unit by aggregating every- which is a function of the generation and storing capabilities of
bodys estimated load, including passive and active users load. its users. Likewise, the model from [10] optimizes an energy
Finally, we introduce the per-slot energy load error profile user demand scheduling algorithm where ln is linear and the
n (h) as the difference from the estimated per-slot energy feasible region is convex. We can, therefore, adapt a robust
load profile ln (h) and the real value. This difference comes energy price that accounts for real-time uncertainties. And
from the assumption that demand-side users do not know their apart from the models [2] and [10], we can trivially assume a
consumption requirements with precision and they may deviate third one: one in which all users are passive, have no elastic
from their estimates. Additionally, active users may also de- demand at all, and simply do not optimize their energy load
viate from their calculated generation and storing policies for profiles. In such case, our robust model can still be applied to
unpredictable reasons such as malfunctioning, disconnections determine worst-case estimation errors.
or lack of collaboration. For these reasons, introducing an Finally, the error terms we have accounted for in (2)
error term makes the model more realistic. We represent these contribute to the users monetary expense fn (l, ), since
error terms both as row vectors n = [n (1), ..., n (H)], they affect the real-time energy price. In order to deal with
n D as the error profiles of user n and as column vectors these unknown demand terms, in Section III we propose
(h) = [1 (h), ..., |D| (h)]T , h H. Note that the index n a worst-case analysis by solving a coupled min-max game.
or h allows to easily distinguish which one we are referring This analysis will allow the supply authority to provide for
to. Additionally, let = [1 , . . . , |D| ] R|H||D| denote the excess energy demands and establish a robust energy price
global vector with the error terms of all users. that accounts for unplanned demand. In the next subsection
we present how to integrate these worst-case error terms in a
B. Day-ahead Cost Model real-time scenario to obtain a robust model.
The purpose of our day-ahead model is to establish an
estimate of the energy price that is robust against real-time C. Real-time Pricing Model
energy demand variations. From a practical perspective, it Once the robust price has been established one day ahead,
determines a worst-case energy production cost taking into if their demand remains within the worst-case estimation error
account expected user energy loads plus uncertainties, so that limits, users are charged the amount of energy they demand
the energy supplier does not suffer monetary losses due to in real time at the robust price. Outside of these limits, a
unplanned real-time demand. penalty function is introduced to cover any extra expenses.
The energy price in the network is determined by the grid Specifically, our proposed real-time cost model is
cost function Ch , which is fixed by the supply-authority and
depends on the aggregate per-slot energy load L(h) at each X
fnrt l, , lrt = Kh Lr (h) lnrt (h)+nh (l, , lrt ) +M (n )
time-slot. We use a quadratic cost function, which is widely hH
used in the literature (see e.g. [2], [10]): (4)
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 4
where l is the energy load profile obtained from the day- b) The individual cost values for the payoff functions of each
ahead optimization process, is the worst-case estimation player are equal for any NE solution of the game.
error (which is analyzed in Section III), and c) The game is monotone and NE solutions can be reached
X
through convergent algorithms, of the kind proposed in [11].
Lr (h) = lm (h) + m (h) (5)
mD Proof. See Appendix A.
is the robust total demanded energy plus worst-case estima- Convergent algorithms to reach such NE solutions are
tion errors, calculated in the day-ahead optimization process. described in [11], which depend specifically on the convex
Variable lrt refers to the actual energy load that users demand structure of the game. In particular, they require a single-loop
in real-time, and h is a penalty function that increases the best response algorithm if the game is strongly monotone, or
cost whenever a user deviates from its specified load profile a double-loop best-response algorithm if it is just monotone
range; for example the following function is a real-time cost (with an added proximity term). The specific steps, which are
model inspired by [4]: similar to those described in [2, Alg.1] for their particular
+ model, but including fixed error terms, are as follows. First,
nh (l, , lrt ) = h ln (h) n (h) lnrt (h)
+ (6) users determine their respective ln based on a day-ahead
+h lnrt (h) ln (h) n (h) optimization process (considering fixed), where every user
where h and h are scalar terms and ()+ = max(, 0). The knows their energy consumption requirements and energy
penalty values discourage the return of unused energy during costs, and calculate their best strategies depending on the
night or low demand times, and higher use of energy during energy prices. Then, all users communicate their intended
day or high demand times. energy load profile ln to the central unit and energy prices
Note that the penalty function proposed in (6) differs are recalculated. This procedure is repeated iteratively until
from [4] in that it does not charge any extra cost within the all users converge to an NE.
uncertainty margins. The reason is that our robust algorithm
already takes into account the extra costs of unplanned demand III. W ORST-C ASE A NALYSIS OF THE E NERGY L OAD
from the production model. Note also that a user can still P ROFILES
deviate further than the worst-case error estimates and, in such In the formulation of G we implicitly assumed that is
case, some penalty function should dissuade the user from given or known a priori, or otherwise is set to zero. However,
these extremes; nevertheless, these deviations are limited with these error profiles are in fact unknown random terms that
some probabilistic confidence. we added in order to correct any deviation from the real
Thus, our robust model acknowledges the cost of producing performance. In this section, these error terms will be treated
a certain amount of energy plus uncertainties, whereas current as optimization variables, and analyzed from a worst-case
real-time models (like [4]) only establish a surcharge for extra performance perspective. This allows the regulation authority
demand. Robust pricing models are of interest for energy to jointly consider both the price and management/availability
suppliers as they take into account production costs under real- costs of energy. With this idea in mind, it is realistic to consider
time price fluctuations. This is an important remark, because bounded error terms, represented by
the real-time demand affects the whole energy price, and not
only the extra demanded energy. h , {(h) R|D| | k(h)k22 (h)} h H (8)
D. Day-ahead Optimization Process where (h) are prefixed in advance and depend on the
We can now introduce the game G = hl , f i that mini- time-slot h due to more or less confidence in the regulator
mizes the users monetary expense with fixed : predictions. In practice, these values can be inferred by the
(
min fn (ln , ln , ) regulation authority, based on historic data from consumers
G : ln n D and expected variations. Quadratic constraints on the error
s.t. ln ln , terms are common in the literature, see e.g. [3], [12].
where ln = [l1 , . . . , ll1 , ll+1 , . . . , l|D| ] represents the rest of In order to propose this error model, we have assumed two
the other users policies, which influence the users decision. conditions that are satisfied generally in practice: i) indepen-
We consider the Nash Equilibrium (NE) as the solution dence of the error terms among users, and ii), no knowledge
concept of interest for this game, defined as a feasible point of the specific probabilistic error profiles of users. Property
l = [l1 , , l|D| ], such that it satisfies i) is satisfied if we assume that users alter their predicted
energy profiles independently (e.g. a user decides to charge an
fn (ln , ln
, ) fn (ln , ln , ) ln ln (7) electric vehicle for an unexpected event). The energy profiles
of different users can be correlated, but what we assume
for every player n D. In our framework, G has at least an
to be independent are the deviation errors from these users.
equilibrium point, which can be reached through a convergent
Property ii) implies that if we do not have any knowledge of
algorithm as established by the following theorem:
the error distribution of a specific user, then it is sensible to
Theorem 1. The following holds for the game G with given substitute such profile with the error distribution of an average
and fixed error values : user profile. By the central limit theorem such average user
a) It has a nonempty, compact and convex solution set. profile will approximate a Gaussian distribution. Then, the
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 5
error contribution of all users can be modeled as a Gaussian in variable ln . Furthermore, if the point that maximizes the
distribution of zero mean, and variance the sum of all user objective is unique, then the objective is differentiable in
variances. Specifically, equation (8) represents an upper bound ln with gradient ln fn (ln , ln , n , n ), and where n =
to the overall network error variance. arg maxn fn (l, n , n ) denotes the maximal point. This
In the following, we analyze a scenario that considers that property is guaranteed by Danskins Theorem [14, Sec. B.5].
every player has to account individually for their worst-case With this analysis in mind, the required assumptions from [11,
error terms and is charged accordingly. The model assumes Th.4.1.a] are satisfied and an NE exists. Furthermore, the
that each player is responsible for some maximum error Jacobian analysis performed for Theorem 1 (see Appendix A)
profiles in joint consideration with the other players. remains the same in Gm , with the exception of substituting
Before the analysis, we bring up a useful result that guar- the fixed n with the optimal solution of the max problem
antees global optimality for a maximization problem with n . Likewise, we can claim that the game is monotone and,
quadratic convex objectives and constraints. therefore, that the solution set is also convex [11, Th.4.1.b].
By Danskins theorem, if the maximum is not a unique point
Theorem 2. (from [13, Th.3]). Given the problem
then the gradient would become a subdifferential, and slight
1 T variations would have to be considered that include the use of
max x Ax + bT x + c
xRn 2 (9) multifunctions [15].
1 T
s.t. x Qx + dx + e 0 Game Gm falls into a rather novel category of min-max
2
nn problems recently analyzed in [16]. Such reference proposes
where A 6= 0, Q R are positive semidefinite matrices, an alternative formulation for min-max games, where extra
b, d Rn and c, e R, and Slaters condition is satisfied: players are added to solve the maximization parts to decouple
then x is a global maximizer iff there exists 0 satisfying the min-max form. Moreover, reference [16] proves equiv-
Ax + b = (Qx + d) (10) alence of NE for both formulations, and proposes solutions
A + Q is positive semidefinite. (11) under monotonicity and other requirements. However, our
problem Gm falls into a more general form with individual max
Note that (10) is a KKT necessary condition, plus a second objectives (rather than a common one), where the monotonic-
requirement (11) that will guarantee global maximum. ity property does not hold. Therefore, the approach from [16]
A. Worst-Case Min-max Game Formulation becomes unsuitable for distributed algorithms. Furthermore,
Gm incorporates global constraints which are not considered
In this section we consider a worst-case situation where
in the previous reference. For this reason, we have to develop
the error terms are coupled among all of the users. We
a novel approach for this particular problem.
analyze which are the users best response strategies and
In order to analyze and solve the coupled maximization
how to calculate the worst-case error terms. In particular, this
problems for all players, we first formulate an equivalent game
transforms G into a game where all users have to solve a min-
Gm (), which is more tractable, together with a nonlinear
max cost function, with coupled interactions and constraints.
complementarity problem (NCP) indicated by (12). The NCP
The new game formulation Gm , where each user n has to
is required to establish the equivalence together with other
determine variables ln and n , is given by
assumptions on the dual variables h . We have also particu-
min max fn (ln , ln , n , , n ) larized the cost term to M (n ) = m kn k22 with m 0,
ln n
Gm : n D which is motivated to penalize the users local deviations.
s.t. (h) h , h H Furthermore, this particular choice will allow us to obtain
ln ln
tractable results for the game and the NCP:
where region ln is convex, compact and nonempty. The min- (
max of game Gm for every user n D represents a game min max pn (ln , ln , n , n , )
ln n
Gm () : n D
with a convex objective function in variables ln and convex s.t. ln ln
regions. Note that for the specific case that M (n ) is quadratic
and convex in variable n , then the maximization problem falls 0 < Kh + m h k(h)k22 (h) 0, h (12)
into the category to readily use Theorem 2, since all equations
where = [1 , . . . , |H| ]T , a b indicates aT b = 0, and
are quadratic. More generally, we can provide the following
theorem for any continuous function M (n ):
X
pn (ln , ln , n , n , ) = (m h )n (h)2
Theorem 3. The following holds for game Gm with min-max X X
hH
objectives: + Kh L(h) + k (h) (ln (h) + n (h)) (13)
a) It has a nonempty, compact and convex solution set. hH kD
b) It is monotone in variables ln , and NE can be reached
Next lemma shows the equivalence between Gm and Gm ().
through convergent algorithms of the kind proposed in [11].
Lemma 1. Games Gm and Gm () together with the NCP
Proof. The max operation inside the objective of each user in
specified by equation (12) present the same NE solutions.
game Gm is performed over a convex and compact set and,
therefore, it is well defined (it always has a solution). The Proof. The KKT system of equations of Gm and those of game
maximum function in Gm preserves convexity of the objective Gm () jointly with the NCP are the same, and therefore, any
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 6
NE candidate solution can be derived from either formulation. and identify terms A = Kh + m and Q = 1. We conclude
A maximum point in both problems exists since all regions are that (11) is satisfied if h Kh + m . Finally, we obtain the
compact. Inequality h Kh + m guarantees that such point first order necessary conditions from (14) for all n D:
is maximum because Theorem 2 is satisfied for every user in X
Gm and, likewise, is maximum in Gm () because it makes the Kh ln (h)+L(h)+ k (h) +2 Kh +m h n (h) = 0,
objective concave in n . Therefore, the NE coincide. k6=n
Note that, in this equivalent formulation of Gm () we and form a system of equations together with (12). By sepa-
have focused on the special case where all h are equal rating variables and solving, we get
for all players for every h H. The equilibrium points
(h)(ln (h)+L(h)+ k (h))
P
of this formulation are normally referred to as variational n (h) = P (l (h)+L(h)+P k6=n (h))2 , n D (15)
n n k6=n k
or normalized solutions [17], and the practical implication P P
is that they establish a common price on a resource for all n (ln (h)+L(h)+ k (h))2
h = Kh + m + Kh k6=n
(16)
agents. These solutions also retain some stability properties 2 (h)
compared to other equilibrium points where the dual variables where we have chosen positive sign to satisfy h Kh + m .
are different among users, see [17] for further details. Since the solution to the fixed point equations is derived
One benefit from dealing with the new formulation Gm () is from conditions (10)(11), they are, therefore, global max-
that the strong max-min property is directly fulfilled if the ob- imum of the individual objective functions. As no player
jective is convex-concave, as opposed to Gm where it remains can further maximize their objectives given the other players
convex-convex (see [18, Ex.3.14b] for further details). This strategies, they form an NE. If the game is strongly monotone
implies that the min-max can be changed into a max-min and (Th. 4c is satisfied), then asynchronous fixed-point iterations
the solution is not altered, i.e., equivalence of both problems over these equations will converge to the global maximum.
holds. We see this is true as long as h Kh + m because If the game is not strongly monotone, then the set of op-
the objective is then concave in n . This allows us to analyze timal equations given by (15)(16) have to be solved in an
the monotonicity of the new game setting Gm () in these alternative fashion. We propose a fixed-point iterative mapping
variables. Let us present a lemma that guarantees the existence that results from calculating equation (15) repeatedly, plus a
of common h s satisfying the previous requirements: projection operation. The mapping in vectorial form becomes:
Lemma 2. There always exists h satisfying (12) (i.e., making
p ah + A(h)
Gm () concave in n ) that is dual optimal in Gm . A lower Th ((h), ah ) = Xh (h) (17)
kah + A(h)k
bound to this value is: h > m + Kh for all h H.
where ah = L(h)+l(h), A = 11T I with size |D||D| and
Proof. See Appendix B.
Xh () represents the euclidean projection onto region Xh .
Now, we can establish the following result. Such region Xh is defined as
n o
Theorem 4. The following holds: |D|
p
Xh , (h) R+ 1T (h)+ |D|1 1
1T ah (h)|D| 0
a) Game Gm () has a nonempty and compact solution set. (18)
b) A variational solution of game Gm always exists. |D|
where R+ refers to the nonnegative quadrant, so that Xh
c) If additionally h > 12 Kh (|D| + 1) + m , then the game corresponds to the upper halfspace region limited by the
Gm () is strongly monotone in variables n for all n D, hyperplane in the definition.
it has a unique NE solution, and it can be reached by a Next we give the following properties of the mapping Th :
distributed asynchronous best response algorithm, of the kind
proposed by [11, Alg.4.1]. Theorem 5. Given the mapping Th ((h), ah ) defined in (17)
where ah is fixed, then it follows:
Proof. See Appendix B a) Th : Xh Xh is a self-map for every h H.
In order to solve Gm () and the NCP described by (12), b) Th is a contraction mapping.
we need to study the inner game formed from the coupled c) The update rule (h)k+1 = Th ((h)k , ah ) converges to a
maximization problems in . In order to do that, we assume unique point (h) Xh that solves the fixed point equation:
that l is fixed. Specifically, we analyze the following game:
n (h) = Th ( (h), ah ).
max pn (l, n , n , ) n D (14)
n R|H| d) The point (h) is an NE of game Gm (), with satisfy-
where (14) is solved jointly with the NCP (12). First, note ing (12) and fixed strategies l(h).
that all users objectives can be separated in |H| independent Proof. See Appendix C
problems. This is possible because the objective function pn
is expressed as a sum over h H and (12) is expressed Because of Theorem 5 and the equivalence principle of
separately for every h H. Lemma 1, the fixed-point solution of the mapping also maxi-
Each individual problem has a quadratic form and, hence, mizes the objectives of game Gm . Thus, the mapping provides
Theorem 2 can be applied to each of them. We can transform an iterative method to find a solution of the game, regardless
the model given by (12) and (14) into a form resembling (9), of its monotonicity properties.
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 7
Region Xh in (18) becomes a technical adjustment that Algorithm 1 Distributed robust algorithm for game Gm
allows the operator to be a contraction self-mapping. It has an (S.1) Given Kh for all h H. Set iteration index i = 0.
involved interpretation, but in simple terms it limits the search 0
Set > 0. Initialize (h) = (h)
1.
space of the fixed-point to the upper region of the hyperplane. |D|
In particular, in such upper halfspace the mapping is always a (S.2) If a termination criterion is satisfied, STOP.
contraction, while in the lower halfspace such property is not (S.3) Solve for every n N
always satisfied. Therefore, a fixed point in Xh always exists,
lni+1 = arg min f (ln , ln
i
, i ) + (ln lni )2
and a recursive algorithm will converge to such point. ln ln 2
Finally, we can explicitly give an expression of the projec- (S.4) Iterate from k = 0 until convergence
tion operation into the halfspace Xh , which is given by
k+1 (h) = T ( k (h), ai+1
h ), h H (19)
if x Xh
x
where ai+1 = nD lni+1 (h) + li+1 (h).
p T 1 1 P
1
Xh (x) , x + (h) A ah + x k1k k1k h
Set the fixed-point solution: i+1 k+1 .
if x
/ Xh , (S.5) If an NE has been reached, update centroids lni = lni .
which only involves linear operations. (S.6) Set i i + 1. Go to (S.2).
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 8
3,000 80 3,000
Algorithm 1 Algorithm 1 15
2,500 60 2,500
2,000 2,000 10
1,500 40
1,500
1,000 20 1,000 5
500
0 0 500
500 1,000 1,500 2,000 0 0
500 1,000 1,500 2,000
Number of users
Number of users
Fig. 1: Real time averaged cost functions over 100 simulations
comparing robust function (4) and (20). The left axis shows Fig. 2: Energy cost of all players vs. number of users. The
the monetary cost of all users and the right axis shows (in %) left axis shows the monetary cost and the right axis shows the
the mean relative gain when using the robust algorithm. mean relative gain (in %) when using the robust algorithm.
The number of users in this plot varies from 100 to 2000 users,
users. The (h) terms have been fixed to be the 10% of the
and the total number of time slots is |H| = 24. The curve we
total energy demand per time slot h H.
refer to as naive users is the solution to the game without
A. Real-time Cost Comparison accounting for any error terms (i.e., = 0), but with added
In this subsection, we present a comparison of our robust worst-case calculations in the price determined by the energy
model with the non-robust model from [4], when users de- supplier. On the other hand, the robust curve represents the
viate from their day-ahead optimized energy load profiles. total energy cost when using Algorithm 1, which does take
Specifically, we present the average cost function over 100 into account the worst-case error terms. The average relative
independent simulations, where the real time error deviations gain over all users when using the robust algorithm is around
are given by white Gaussian processes with zero mean and the 7%, which remains almost constant with the number of
variance (h)/|D| for each user, and where the number of users; these savings values are represented in the right hand
users |D| ranges from 100 to 2000. The cost function we use axis of the figure for different number of users. This result
to evaluate the performance in real-time of Algorithm 1 is illustrates that users should adjust their energy load profiles
described in (4). Likewise, the cost function for user n we after taking the worst-case error terms into account. For these
compare with, denoted as non-robust, has the following form: simulations we fixed the term m = 0.001.
X + In Figure 3a, we plot the convergence rate of Algorithm 1,
fnnrt l, , lrt = Kh Lwc (h) lnrt (h) + h ln (h) lnrt (h) which depicts the averaged convergence rate particularized for
hH (20) 1000 users. The convergence rate is very similar to the results
+ presented in earlier approaches (see, e.g., as [2, Alg.1]). Our
+h lnrt (h) ln (h) ,
algorithm has the same algorithmic steps except for the inner
which has no error terms in contrast to (4). Note also that we loop involved in step 4, which we didnt observe to impact the
are representing the average cost of all users in the network. Fi- computation time. In Figure 3b, we plot the convergence speed
nally, values h and h are set to h = 0.2 for h [0, . . . , 8], of the contraction mapping from step (S.4) in Algorithm 1. The
h =p0.8 for h [9, . . . , 24] and h = ( h ), where x-axis shows the number of iterations and the y-axis shows the
= |D|/(h). The specific simulation parameters h and magnitude hH k k+1 (h) k (h)k. The stopping criteria
P
h are obtained from [4], and parameter is introduced to was set to 108 . Recall that convergence is guaranteed since
incorporate the uncertainty of the real-time consumption in the computation in (S.4) is a contraction mapping.
the penalty terms. Note that the value of is the inverse of The proposed algorithm was implemented in Matlab
the deviation of every users real-time error distribution. R2015a using an Intel i5-2500 CPU @ 3.30GHz. The users
The average results of the cost models are presented in optimization problems were run in sequential order (although
Figure 1. The term M (n ) is not present in (20), so we set in practice they can run in parallel). The local convex optimiza-
M (n ) = 0 in (4) for a fair comparison. Commenting on these tion problems have unique optimal points, so any optimization
results, we observe that the cost for the users is significantly solver will produce similar results as ours. In particular, we
lower than with a penalty based model as (20). Therefore, this used the fmincon function from Matlab. The whole simulation
model effectively reduces the monetary expenses of all users took about 61.38 seconds for 1000 users.
in the real-time market, while at the same time provides a As a conclusion, these simulations show consistent results
reliable production cost estimate to the energy supplier. in which there is an actual benefit of around the 7% when
B. Comparison of Algorithm 1 with naive users using a robust algorithm vs. a naive one (i.e., that does not
take into account the error terms for the optimization process),
In order to further support our results, we include the while the extra computation required is minimal.
performance of our robust algorithm and compare it to a
scenario where users do not consider the worst-case error V. C ONCLUSION
terms in their energy load profiles. In Figure 2, we plot the We presented a demand-side management model that takes
monetary cost of all users in different sizes of the smart grid. into account deviations in the estimated consumption profiles,
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 9
Convergence criteria
100 satisfied by definition of G , and that the objective functions
of all players are differentiable and convex in variables ln ,
n D. Indeed, the convexity is guaranteed since the
101
objective functions are quadratic with positive terms in the
squared elements so that the Hessian results into a diagonal
102 matrix of the form diag({Kh }h ) that is positive definite.
Hence, the claim is satisfied applying [11, Th.4.1].
103 The justification of b) is readily available by observing that
0 2 4 6 8 10
Iteration number the objective is convex in ln (h) and that other players actions
only affect the individual objective through the aggregate
(a) Convergence rate of Algorithm 1 with convergence criteria kli L(h). Therefore, the maximum value is unique for each player
li1 k2 /kli k 102 . given any strategy profile ln
since the maximum does not
105
Convergence criteria
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 10
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TSG.2016.2559583, IEEE
Transactions on Smart Grid
IEEE TRANSACTIONS ON SMART GRID 11
1949-3053 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.