Optimal Sharing and Fair Cost Allocation of Community ... - arXiv

1

Optimal Sharing and Fair Cost Allocation ofCommunity Energy Storage

Yu Yang, Student Member, IEEE, Guoqiang Hu, Senior Member, IEEE, and Costas J. Spanos, Fellow, IEEE

Abstract—This paper studies an energy storage (ES) sharingmodel which is cooperatively invested by multiple buildings forharnessing on-site renewable utilization and grid price arbitrage.To maximize the economic benefits, we jointly consider the ESsizing, operation, and cost allocation via a coalition game formu-lation. Particularly, we study a fair ex-post cost allocation basedon nucleolus which addresses fairness by minimizing the minimaldissatisfaction of all the players. To overcome the exponentialcomputation burden caused by the implicit characteristic func-tion, we employ a constraint generation technique to graduallyapproach the unique nucleolus by leveraging the sparse problemstructure. We demonstrate both the fairness and computationalefficiency of the method through case studies, which are notprovided by the existing Shapley approach or proportional method.Particularly, only a small fraction of characteristic function (lessthan 1% for 20 buildings) is required to achieve the cost allocationversus the exponential information required by Shapley approach.Though there exists a minor increase of computation over theproportional method, the proposed method can ensure fairnesswhile the latter fails in some cases. Further, we demonstrateboth the building-wise and community-wise economic benefitsare enhanced with the ES sharing model over the individual ES(IES) model. Accordingly, the overall value1 of ES is considerablyimproved (about 1.83 times).

Index Terms—energy storage sharing, coalition game, costallocation, nucleolus, fairness.

I. INTRODUCTION

Energy storage (ES) is a key technology for the world’stransition to a sustainable, flexible and reliable energy system[1]. Based on the market applications, ES are commonly dif-ferentiated as grid-level and customer-level ES 2. While grid-level ES are connected to the distribution or transmission net-works by system operators for ensuring energy supply quality,customer-level ES are usually deployed in consumer premisefor harvesting economic benefits [2]. Particularly, consumers

This work was supported by the Republic of Singapore’s National ResearchFoundation through a grant to the Berkeley Education Alliance for Researchin Singapore (BEARS) for the Singapore-Berkeley Building Efficiency andSustainability in the Tropics (SinBerBEST) Program. BEARS has beenestablished by the University of California, Berkeley as a center for intellectualexcellence in research and education in Singapore.

Yu Yang is with SinBerBEST, Berkeley Education Alliance for Researchin Singapore, Singapore 138602 e-mail: ([email protected]).

Guoqiang Hu is with the School of Electrical and Electronic En-gineering, Nanyang Technological University, Singapore, 639798 e-mail:([email protected]).

Costas J. Spanos is with the Department of Electrical Engineering andComputer Sciences, University of California, Berkeley, CA, 94720 USAemail: ([email protected]).

1The proportion of total electricity bill reduction relative to the ES capitalcost.

2grid-level ES also refers to in-front of the meter (FTM) ES and customer-level ES refers to behind-the-meter ES.

paired with ES can cut their electricity bills by synchronizinglocal volatile renewable supply with non-shiftable demand andresponding to grid price variations [3]. Over the years, grid-level ES projects are being deployed globally [2], whereasthe customer-level ES deployment is still largely impeded bythe high ES capital cost and long payback period albeit themandatory goals set by the governments [4, 5]. How to breakor lower the barriers for customer-level ES deployment hasbecome a critical issue facing the energy system transitions.

In recent years, sharing economy has manifested in trans-portation and housing systems [6]. Naturally, such sharing-oriented solution is penetrating energy systems for bringingin new technologies, such as ES [6, 7]. Particularly, multipleusers can cooperatively invest and share a central ES [8–10].The added benefits of sharing ES over installing individualES (IES) are comprehensive. First, the utilization of ES canbe increased by exploiting the complementary features of userproduction and consumption. For example, the users generallyhave different renewable generation and load patterns, render-ing them to use the ES in a time division manner. Second,duplicate installation cost can be avoided. The capital cost ofES deployment is composed of battery cost and installationcost, the latter of which accounts for 20%-50% 3. Third,additional benefit from the economy of scale can be embraced.Similar to other goods or commodities, customers are likelyto install bulks of ES at the wholesale price. Last but not theleast, locating a central ES in a community-shared space canrelieve the space concern for the customers.

The augmented benefits of cooperative ES sharing areclear, however how to reap them relies on a comprehensivebusiness model which is a nontrivial combinatorial problemthat integrates

• ES sizing. As ES is capital-intensive, determining theappropriate energy and power capacity is a prerequisiteto maximize the economic benefits of the ES sharing.

• ES operation. The operation of the shared ES shouldcoordinate the charging and discharging requirements ofall users.

• Cost allocation. Since the users are self-interested andindependent stakeholders, it is essential to study the faircost allocation among the users.

In the literature, a number of works have studied thebusiness models for cooperative ES sharing (see [7, 11, 12]and the references therein). However, most of them addressed

3A Tesla Powerwall costs 7,600$ before installation. However, accountingfor the installation cost, a rough estimate of the Tesla Powerwall cost $9,600-$15,600 for a full system installation.

arX

iv:2

010.

1545

5v5

[cs

.GT

] 5

Apr

202

1

2

the ES sizing (see [13, 14], for examples) and ES operation(see [15, 16], for examples) separately. Nevertheless, they areinterdependent and require to be coordinated so as to justifythe high ES capital cost and maximize the economic benefits.Moreover, most existing works addressed cost allocation usingrule-based cost allocation mechanisms which do not provideany notion of fairness 4.

A. Main contributionsThis paper studies a cooperative ES sharing model among

multiple buildings, each of which seeks economic benefitsfrom local renewable integration and grid price arbitrage. Ourmain contributions are as follows.• We formulate the optimal ES sharing integrating optimal

sizing, operation and cost allocation as a coalition game.• We address the fair ex-post cost allocation for ES sharing

based on nucleolus. In particular, we employ a constraintgeneration technique [17] to overcome the exponentialcomputation burden caused by the implicit characteristicfunction.

• We demonstrate the superiority of the cost allocation overthe existing Shapley approach and proportional methodby providing both fairness and computational efficiencythrough case studies. Particularly, only a small fraction ofthe characteristic function (less than 1% for 20 buildings)is required to achieve the cost allocation.

• We show the enhanced economic benefits of the ESsharing model over the IES model both at the building-wise and community-wise. Specifically, the ES sharingmodel yields higher cost reduction to each committedbuilding as well as the whole community. Accordingly,the overall value of ES is considerably improved (about1.83 times).

The remainder of this paper is organized. In Section II,we review the related works. In Section III, we presentthe coalition game formulation. In Section IV, we study thefair cost allocation. In Section V, we study the fairness andcomputational efficiency of the cost allocation as well asthe economic benefits of the ES sharing model through casestudies. In Section VI, we briefly conclude this paper.

II. RELATED WORKS

In the literature, a number of works have studied ES sharingmodels from the perspective of sizing and operation.

For ES sizing, [13] proposed an analytical approach byusing a Markovian fluid model to capture the stochastic userdemands where the ES was used as backup to ensure userdemands under grid capacity limits. [14] studied the optimalsizing of shared ES with the objective to facilitate photovoltaic(PV) utilization so as to minimize consumer cost.

The operation of cooperative ES sharing generally addressesthe charging and discharging coordination among differentusers. Typically, [9] proposed a credit-based distributed al-gorithm to manage a central ES shared by a group of cost-aware households. The discharging rate of the shared ES is

4In this paper, we refer to fairness as satisfaction and we interchangeablyuse the word “fairness” and “satisfaction”.

dynamically allocated among the households by their creditswhich characterize their accumulated stored energy in theshared ES. [18] proposed an ES sharing model for minimizingthe reputation-weighted energy cost of consumers, which arecharacterized by the proportions of their renewable injectioninto the shared ES over the historical time periods. Similarly,[15, 16] studied the control of a shared ES for optimizing thetotal weighted cost of a group of homes which reflects theiragreements on cost-saving priorities. [19] studied a sharedES model working as an energy provider to serve consumerswith elastic demand. The operation of ES is managed by anaggregator obligated to minimize the total electricity bill forconsumers. Particularly, a marginal service price model wasdeduced to charge the consumers. [11] studied the optimaloperation of a solar-plus-storage system shared across multipleconsumers for social welfare or profit maximization. Gener-ally, these works studied the ad hoc operation of a sharedES for some specified objectives. They hardly addressed theoptimal ES sizing and capital cost allocation among theparticipants which are important for practice.

Besides, the sizing and operation of cooperative ES sharingare mostly addressed separately in the literature, however theyshould be jointly considered so as to justify the high ES capitalcost and maximize the economic benefits. Moreover, it isessential to study the fair cost allocation (i.e., ES capital cost)as the users are self-interested and independent stakeholders.This paper works towards such objectives. Particularly, westudy an ES sharing model that is cooperatively invested andshared by multiple users to harness the economic benefits ofgrid price arbitrage as well as local renewable integration. Oneclose work is [12] which studied the similar sharing paradigmof ES, and an analytical fair cost allocation formula basedon core was identified. However, that work only consideredgrid price arbitrage for ES which is generally not enoughto justify the high ES capital cost, and the results (i.e., ESoperation, sizing, and cost allocation) are restricted to thetwo-period market setting and cannot be extended to thecase with time-dependent renewable generation. Exceptionally,[20, 21] studied an ES sharing model that accounts for bothprice arbitrage and local renewable integration. Whereas thoseworks focused on a different sharing paradigm in which athird-party leads the ES sharing among its “consumers” andpursues profit maximization. Another related work that sharesthe similar structure of ours by including the ES sizing,operation and cost allocation is [22]. However, that workstudied the ES sharing across multiple electricity retailers.Besides, the cost allocation was rule-based and did not addressthe fairness due to the computational challenges.

III. COALITION GAME FORMULATION OF ES SHARING

A. The Configuration of ES Sharing

Fig. 1 shows the configuration of the ES sharing modelstudied in this paper. We consider a grid-connected commu-nity composed of multiple buildings with on-site renewablegeneration (i.e., wind and solar power). Wherein the buildingscooperatively invest and share a community ES (CES) toharness renewable utilization and grid price arbitrage. The

3

Fig. 1: The configuration of a community energy storage (CES).

buildings can charge the CES with local renewable generationor the procured electricity from the grid. Conversely, they candischarge it to supply their demand when required. Besides, weallow the buildings to sell energy (i.e., renewable generation ordischarged energy from the ES) back to the grid. As the ES isshared by multiple buildings, we assume a central coordinatorobligated to coordinate their charging and discharging behav-iors of the buildings. Supportively, a net metering is attachedto each building for monitoring the energy flow. Particularly,different buildings can charge and discharge simultaneously asthe central coordinator only cares about the net power flow.In other word, if one building charges and another buildingdischarges, there would exist some cancellations. However,each individual building can not charge or discharge at thesame time due to the physical limits.

We study the optimal ES sharing model that encompassesthe optimal sizing, operation, and ex-post cost allocation.Considering the computation burden of long-term planning(i.e., sizing), we project the problem on a daily basis and studythe problem in a discretized-time framework.

B. Main Assumptions

We clarify our main assumptions as below.

(A1) The purchase price of the grid is much higher than theselling price.

(A2) The buildings do not share (or trade) energy with eachother through the shared ES.

(A3) We only consider non-elastic demand of the buildings anddemand response is not discussed.

Remark 1. (A1) is a mild assumption as renewable genera-tion is generally motivated to be consumed locally instead offed into the grid as negative load. Meanwhile, this setting canprevent opportunistic folks from disrupting electricity marketwith ES: stocking energy with lower price and selling backat a higher price. (A2) is imposed to focus on the economicbenefits of ES sharing. However, the model can be extendedto incorporate energy sharing. This supposes to enhance theeconomic benefits of ES sharing further.

C. Coalition Game Formation

We formulate the problem as a coalition game [23]. Coali-tion game is a branch of game theory that studies the coopera-tive behaviors of a group of rational agent, which accords withour settings. Following the standard terminologies, we labelthe building players by N := 1, 2, · · · , N. We study theproblem over the discretized time slots T := 0, 1, · · · , T−1.

(i) Coalitions: For a community composed of N buildings,we refer to any subset of buildings S ⊆ N cooperativelysharing an ES as an ES coalition or sub-coalition S.

(ii) Coalition value and characteristic function: Coalitionvalue ν(S) quantifies the economic benefit of an ES coalitionS. We indicate a N -player coalition game by (N , ν), wherefunction ν : 2N → R represents the characteristic function ofcoalition game that assigns value ν(S) to each sub-coalitionS ⊆ N . Particularly, the number of sub-coalitions S ⊆ Ngrows exponentially with the scale, i.e., O(2N ).

In this paper, we characterize the value of an ES coalitionν(S) by the total cost: the total electricity bills plus theES capital cost. When we use a collection of representativescenarios to capture the patterns of renewable generation andbuilding demand, the coalition value ν(S) can be characterizedby a two-stage stochastic optimization problem that couplesthe optimal ES sizing and operation:

(P) : ν(S) = minES ,PS≥0

c(xS) +∑ω∈Ω

ρωg(xS , ζω) (PU )

g(xS , ζω) = minyωi ,i∈S

g(xS ,yω, ζω) (PL)

subject to: yωi ∈ Yω

i , i ∈ S. (1a)∑i∈S

eb,ωi,t ≤ ES , t ∈ T . (1b)∑i∈S

pch,ωi,t ≤ PS , t ∈ T . (1c)∑

i∈S

pdis,ωi,t ≤ PS , t ∈ T . (1d)

where i and t are building and time indices. xS = (ES , PS)denotes the ES capacity: energy capacity ES (in kW h) andpower capacity PS (in kW). ω, ζω , Ω and pω representscenario indices, scenario realizations, scenario collection andscenario probabilities. yω

i denotes the operating strategy forbuilding i under scenarios ω, which includes the charg-ing/discharging of the ES and energy trading with the grid.Accordingly, Yω

i indicates the set of admissible strategies.• The first-stage objective (PU ) captures the total cost

which is equal to the ES capital cost c(xS) plus theweighted operation cost

∑ω∈Ω pωg(xS , ζω). For the ES

capital cost, we capitalize on an amortized price model[21, 24]: c(xS) = kpPS + keES . kp (s$/kW h) and ke(s$/kW) are the amortized ES capacity price which areobtained according to the projected ES price 100¤/kWhand 300¤/kW by 2025 [25].

• The second-stage objective (PL) characterizes the optimaloperation cost g(xS ,y

ω, ζω) for each scenario ω, whichis subject to the ES capacity xS , the scenario realizationζω , and the building operating strategies yω = [yωi ],∀i ∈S. The operation cost consists of electricity purchase costand the revenue of selling energy to the grid. We use

4

cg+t , cg-

t and cg,max to indicate the purchase, selling anddemand charge price. pg+,ω

i,t and pg-,ωi,t denote the procured

and sold energy of building i over period t. pg,max,ωi

characterizes the peak demand over the billing cycle. Wehave the operation cost:

g(x,yω, ζω) =∑i∈S

∑t∈T

(cg+t p

g+,ωi,t − c

g-t p

g-,ωi,t

)+ cg,maxpg,max,ω

i

(2)

• We define the operation strategy of building i as yωi =

[pchi,t, p

disi,t, e

b,ωi,t , p

g+,ωi,t , pg-,ω

i,t , pg,max,ωi ] which includes the

charging and discharging energy: pchi,t, p

disi,t, the stored

energy: eb,ωi,t , the procured and sold energy from/to the

grid: pg+,ωi,t , pg-,ω

i,t and the peak demand over the billingcycle: pg,max,ω

i . Correspondingly, the set of admissibleoperation strategies Yω

i is constituted by

0 ≤ pchi,t ≤ pch,max, (3a)

0 ≤ pdisi,t ≤ pdis,max, (3b)

eb,ωi,t+1 = eb,ω

i,t + pch,ωi,t ηch − pdis

i,t/ηdis, (3c)

eb,ωi,t ≥ 0, (3d)

pg+,ωi,t − p

g-,ωi,t = pch,ω

i,t − pdis,ωi,t + pd,ω

i,t − pr,ωi,t , (3e)

pg+,ωi,t , pg-,ω

i,t ≤ pg,max,ωi , (3f)

pch,ωi,t pdis,ω

i,t = 0, (3g)

pg+,ωi,t pg-,ω

i,t = 0, ∀t ∈ T . (3h)

where constraints (3a)-(3b) model the charging anddischarging rate limits. Constraint (3c) tracks the storedenergy for each building subject to the charging anddischarging efficiency ηch, ηdis. In this formulation, we donot consider energy sharing, thus each building can notover deplete its stored energy as imposed by constraint(3d). Constraint (3e) models the instantaneous balanceof supply and demand of each building, with pd,ω

i,t andpr,ωi,t denoting the non-elastic demand and local renewable

generation. Constraint (3f) captures peak demand over thebilling cycle. Particularly, the complementary constraints(3g)-(3h) enforce the physical limits of non-simultaneouscharging (purchasing) and discharging (selling).

Note that problem (P) is non-linear and non-convex dueto the presence of complementary constraints (3g)-(3h), mak-ing it computationally intractable with off-the-shelf solvers.However, considering the ES efficiency, i.e., ηch, ηdis < 1,and the grid price setting, i.e., cg+

t > cg-t , the complementary

constraints can be relaxed without affecting the optimal solu-tion. We refer the readers to an illustrative proof in AppendixA. With constraints (3g)-(3h) relaxed, there only exist linearconstraints. Besides, we note that the two-stage problem hasa min-min structure, making it possible to be converted to thesingle-stage convex problem (4) that can be tacked by someexisting commercial solvers (e.g., CPLEX).

(P′) :ν(S) = min c(xS) +

∑ω∈Ω

ρωg(x,yω, ζω)

subject to: (1b)− (1d). (3a)− (3g), ∀i ∈ S.var: ES , PS ,yω

i ,∀i ∈ S.

(4)

D. ES Coalition Game Properties

In this part, we study the characteristics of the ES coalitiongame. We have the following main result.

Theorem 1. The ES coalition game (N , ν) is subadditive, i.e.,ν(S1 ∪ S2) ≤ ν(S1) + ν(S2), ∀S1,S2 ⊆ N ,S1 ∩ S2 = ∅.

Remark 2. We defer the proof to Appendix B. Theorem 1implies that it won’t be worse off for two disjoint groups ofbuildings to merge and share a single ES. Fundamentally, theoverall economic benefits can be enhanced through merging.Therefore the buildings within a community are inclined toform a grand ES coalition N to maximize economic benefits.

IV. COST ALLOCATION BASED ON NUCLEOLUS

In Section II-D, we have proved the enhanced overll ben-efits of the ES sharing model over IES model. However,the building-wise gains relies on the ex-post cost allocation,which is supposed to be fair to ensure stable cooperation: allparticipants are satisfied and have no motivations to deviate ordisrupt the cooperation.

There exist multiple solution concepts regarding fair costallocation of coalition game [26]. One prominent one is core[23, 27]. Normally, computing core corresponds to a NP-complete linear programming (LP) that depends on the entirecharacteristic function. For example, for a N -player coalitiongame, it generally requires to solve a LP problem with (2N−1)constraints associating with O(2N−1) coalition value. This isa non-trivial task due to the exponential computation burden.Besides, the existence and uniqueness of core is anothergeneral concern for practice [23].

Another primary solution concept is Shapley value with themain idea of distributing the co-created value by the players’marginal contributions. One main advantage of Shapley overcore is the existence and uniqueness. However, it also sufferscomputation intensity from the entire characteristic function.

Nucleolus is another essential solution concept which pur-sues fairness by minimizing the dissatisfaction of all players[28]. Reasonably, a cost allocation can be viewed as fair if allplayers are satisfied (i.e., the maximum dissatisfaction of allplayers is non-positive). Besides, the nonempty and uniquenessproperty of nucleolus is preferable, motivating us to study thecost allocation of ES sharing based on nucleolus. Nevertheless,the computation challenge is yet to be addressed as computingnucleolus is as or even more difficult than the other solutionconcepts.

A. Definitions

We first introduce the main definitions.

Definition 1. (Cost allocation) We use vector x ∈ RN todenote a cost allocation of coalition game (N , ν), where theentry xi denotes the allocation to player i.

Definition 2. (Imputation) An imputation x ∈ RN is a costallocation for a grand coalition (N , ν) which is both efficientand individually rational. We have the set of imputations:

I =x ∈ RN : x(N ) = ν(N ) and ν(i) ≥ xi,∀i ∈ N

5

where we define x(S) =∑

i∈S xi,∀S ⊆ N .

Definition 3. (Core) Core refers to the imputation that nosubsets of players has incentives to deviate from the grandcoalition (N , ν). The set of core is defined as

C =x ∈ I : x(S) ≤ ν(S),∀S ⊆ N

Definition 4. (Lexographical order) Assume two N -dimensional vectors y, z ∈ RN with elements arranged innon-increasing order, i.e., yi ≥ yj and zi ≥ zj if i < j. Weclaim vector y is lexographically smaller than vector z, i.e.,y ≺lexz, if ∃k<N that yi = zi,∀i < k and yk<zk (if k = 1,we have y1 < z1).

Definition 5. (Excess of coalition) For a given cost allocationx ∈ R|S|, the excess of coalition S is defined as

e(x,S) = x(S)− ν(S)

For a coalition game characterized by cost minimization,the excess of coalition e(x,S) can be interpreted as thedissatisfaction of coalition S with the cost allocation x. Sincethere exists a group of players, the dissatisfaction of playersregarding a specified cost allocation x ∈ RN correspond toa excess of coalition vector. To address this issue, nucleolusis defined by the lexographical order of excess of coalitionvectors. More specifically, nucleolus is the imputation withthe lexicographically minimal excess of coalition vector. Theformal definition is given below.

Definition 6. (Nucleolus) For a coalition game (N , ν), letO(x) ∈ R2N−1 be the excess of the coalition vector for costallocation x ∈ RN (imputation) arranged in a non-increasingorder, i.e., Oi(x) ≥ Oj(x),∀i < j, then a cost allocationx ∈ RN is the nucleolus if we have

O(x) ≺lex O(x′),∀x

′∈ I \ x.

B. An algorithm to find the nucleolus

From the definition, we note that nucleolus is alwaysnonempty and unique. Particularly, nucleolus will locate inthe core if the latter is nonempty [17]. However, computingnucleolus is nontrivial as it corresponds to searching for thelexicographically minimal excess of coalition vector. Morespecifically, we are required to solve a sequence of lexico-graphically minimization problems, i.e., first identifying theset of cost allocations X1 that minimize O1(x), and thenminimize O2(x) over X1, where Oi(x) denotes the i-th entryof the excess of coalition vector O(x) (see Definition 6).These episodes are carried forward until the unique nucleolusis approached. Particularly, we note that the entire charac-teristic function is required in each episode to compute theexcess of coalition vector. This is computationally intensivefor ES coalition game as the characteristic function is implicitand characterized by stochastic optimization problems. Forexample, for an ES coalition with 20 buildings, we are requiredto solve the stochastic optimization problem (4) 220−1 (morethan 107) times to identify the characteristic function.

To overcome the computational challenges, this sectiondevelops an algorithm to search for nucleolus of the ES

coalition game by employing a constraint generation technique[17]. The main idea is spurred by the uniqueness of nucleolusand the underlying sparse structure of problem. Specificallyfor a N -player coalition game, the cost allocation correspondsto solving a sequence of LPs (i.e., lexicographically mini-mization) with N decision variables subject to 2N − 1 linearconstraints. This implies at most N of 2N − 1 constraintsare binding at the optima. Intuitively, if the N bindingconstraints are known a priori, only N coalition value areactually required. Indeed, the essential idea of the proposedcost allocation is to identify such critical constraints usingconstraint generation technique.

To be noted, the algorithm to search for nucleolus is consti-tuted by multiple episodes. We first starts with computing theleast core [26]. After that a sequence of lexiographically min-imization problems are solved. In each episode, we capitalizeon a constraint generation technique to identify the dissatisfiedcoalitions with the current cost allocation. The algorithm willterminate until we encounter the unique nucleolus. In thesubsequent, we introduce the details.

1) Computing least core: The least core of coalition game(N , ν) is the solutions of the LP problem [26]:

z∗ = min z

subject to: x(N ) = ν(N )

x(S)− ν(S) ≤ z, ∀S ⊆ N \∅,N.(5)

Problem (5) is equivalent to z∗ = minx∈I

maxS⊆N\∅,N

e(x,S). We

note that least core is the set of imputations with the minimummaximum excess of coalition, i.e., minx∈I O1(x). Therefore,the least core includes nucleolus based on the definition.

An example: assume a 2-player coalition game with the ex-cess of coalition vectors O(x) = [5, 3, 3], O(y) = [5, 4, 2], andO(z) = [6, 4, 1] for the cost allocation x,y, z. We can figureout the least core x,y and the unique nucleolus x becausewe have O1(x) = O1(y) ≤ O1(z) and O2(x) < O2(y).

Clearly, least core is well-defined and can be used to narrowthe search scope of nucleolus. Nevertheless, computing leastcore by solving problem (5) is computationally intensive asit requires the entire characteristic function: ν : 2N → Rwhich corresponds to the value of all sub-coalitions S ⊆ N .To overcome the computational challenge, we capitalize ona constraint generation technique to gradually approach theleast core instead of solving problem (5) all at once. Themain idea contains three steps: i) solve the relaxed problem(6) corresponding to a subset of coalitions F1 (e.g., start withF1 = 1, 2, · · · , N); ii) identify the most “violated”sub-coalition (i.e., maximum excess of coalition) with theobtained cost allocation; iii) add the identified coalition toF1. This process is repeated until no “violated” sub-coalitionswith the obtained cost allocation exists. This indicates theleast core defined in problem (5) is approached. We defer theimplementation of constraint generation technique later.

z∗,1 ==

min zsubject to:x(N ) = ν(N )x(S)− ν(S) ≤ z, ∀S ∈ F1\∅,N.

(6)

6

2) Lexicographically optimization: Intuitively, if the leastcore is unique, the nucleolus is found. However, that is notthe usual case and we usually have to carry on to identify theunique nucleolus by solving a sequence of lexicographicallyoptimization problems. For example, minimize O2(x) over theleast core x ∈ x|O1(x) = z1,∗ and so forth. Consideringthe general case, we introduce the problem of minimizingOk(x) over x|Oj(x) = zj,∗,∀j = 1, 2, · · · , k − 1 atepisode k. Akin to computing least core, we capitalize on theconstraint generation technique to overcome the computationburden by executing the three steps. Slightly different, we havethe relaxed problem (7) with blocks of binding constraintsindicated by Γj . The interpretation is that at each episode k, wesolve the lexicographically optimization minxOk(x) withinthe scope of x ∈ Rn|Oj(x) = zj,∗,∀j = 1, 2, · · · , k − 1.

zk,∗=

min zsubject to: x(S)− ν(S) ≤ z,

∀S ∈ Fk \ ∪j≤k−1Γj .zj,∗ = x(S)− ν(S),

∀S ∈ Γj , j = 1, 2, · · · , k − 1.x(N ) = ν(N ).

(7)

3) Constraint generation: This part introduces the imple-mentation of the constraint generation to identify the most“violated” sub-coalition for 1) and 2).

Suppose we have a cost allocation xk,∗ (the solution ofproblem (6) or (7) with a specific subset of coalitions Fk).The constraint generation technique requires to identify themost “dissatisfied” or “violated” sub-coalition (the subset ofplayers) not included in Fk. This requires to identify the sub-coalition with the maximum excess of coalition regarding thecost allocation xk,∗ within the remaining coalitions N \ Fk.To address such issue, we define some binary variables si ∈0, 1,∀i ∈ N to indicate whether player i is in the identifiedsub-coalition or not. Thus we can interchangeably indicatea coalition by Sj ⊆ N or a binary vector sj = sj1, s

j2,

· · · , sjN, where we have sji = 1 if player i is in coalitionSj , otherwise sji = 0. Thus, we can formulate the problem as

c∗ = maxS

∑j∈S

xk,∗j − zk,∗ − ν(S)

subject to:1 ≤∑i∈N

si ≤ N − 1.si ∈ 0, 1, ∀i ∈ N . (8a)∑i|sji=0

si +∑

i|sji=1

(1− si) ≥ 1, ∀j|Sj ∈ Fk. (8b)

where constraint (8a) is imposed to exclude the emptycoalition ∅ and grand coalition N . Constraint (8b) enforcesthe exclusion of coalitions Fk.

We note that problem (8) requires the explicit characteristicfunction ν(S). However, for the ES coalition game, thecharacteristic function ν(S) is characterized by the stochasticoptimization problem (4). To address such issue, we blend theproblem as

c∗ = max∑j∈S

xk,∗j − zk,∗ −(c(xS) +

∑ω∈Ω

ρωg(xS , ζω))

subject to: (1b)− (1d).(3a)− (3e), (3g)− (3h), ∀i ∈ N .pg+,ωi,t − p

g-,ωi,t ≥ sip

b,ωi,t + pd,ω

i,t − pr,ωi,t ,

pg+,ωi,t , pg-,ω

i,t ≤ siPg,max

1 ≤∑i∈N

si ≤ N − 1. si ∈ 0, 1, ∀i ∈ N .∑i|sji=0

si +∑

i|sji=1

(1− si) ≥ 1, ∀j|Sj ∈ Fk.

var: E,S , PS , s,yωi , ∀i ∈ N .

(9)

where P g,max indicates the maximum trading energy withthe grid of each building over single period. The objectiveof problem (9) characterizes the excess of coalition for anES coalition. Particularly, we use the combinatorial constraintspg+,ωi,t − p

g-,ωi,t ≥ sip

b,ωi,t + pd,ω

i,t − pr,ωi,t to uniformly capture the

load balance for the buildings in or out of the sub-coalition.Specifically, for the buildings in the identified sub-coalition,we have si = 1 and the procured electricity from the grid is atleast to satisfy the building demand, otherwise we have si = 0and the load balance constraints are relaxed.

Problem (9) is a mixed-integer linear programming (MILP)with O(N) binary variables, which can be handled by someexisting solvers like CPLEX for moderate scales. However,if the scale is very large and solving problem (9) directlybecomes computationally intensive, we would need to findsome other ways to handle the problem. As aforementioned,with the (most) “dissatisfied” sub-coalition S∗ regarding thecurrent cost allocation proposal xk,∗ obtained, our next stepis to added it to Fk (i.e., Fk := Fk ∪ S∗) and adjust thecost allocation accordingly.

The procedures of identifying the most “violated” sub-coalition and adjusting the cost allocation are alternated untilno “violated” sub-coalition is found, i.e., the optimal value ofproblem (8) is non-positive (c∗ ≤ 0). This implies the optimalsolution of problem (6) or (7) is approached, i.e., the leastcore or the cost allocation for minOk(x) has been identified.

We display the main procedures to search for the nucleolusof the ES coalition game in Algorithm 1. Particularly, weclarify three main points regarding the algorithm. First, thealgorithm includes two-loops: outer-loop and inner-loop. Theouter-loop associates with the lexicographically optimizationindicated by the episode k. Whereas the inner-loop iterativelysolve the lexicographically optimization by employing con-straints generation technique. In inner-loop, we alternativelyidentify the most “violated” coalition and update the costallocation. The inner-loop will terminate until no “violated”sub-coalition is found (i.e., c∗ ≤ 0 for problem (9)), whichindicates the lexicographically optimization has been solved.Second, there are two crucial steps when switching from theinner-loop to the outer-loop: i) at the end of each inner-loop, the active or binding constraints are required to beidentified (line 12). This can be achieved by checking theinequality constraints of problem (6) or (7). Specifically, withthe obtained solution xk,∗, zk,∗, we identify the active orbinding constraints Γk for next computing epoch by checking

7

the equality zk,∗ = ν(S)−x(S), ∀S ∈ Fk \∪i≤k−1Γi; ii) thesubset of coalitions is copied for the next computing epochk + 1 (line 13). Third, the overall algorithm will terminateuntil the solution of lexicographically optimization is unique(line 4-7).

Algorithm 1: Search for nucleolus of ES coalitiongame based on constraints generation

Initialize: N := 1, 2, · · ·N: building participants.Output : x ∈ RN : cost allocation for the buildings.

1 Initialize: k → 1, F1 = 1, · · · , N, STOP: =false;

2 while !STOP do3 Solve prolem (7) (or (6) if k = 1) and obtain the

solution xk,∗ and zk,∗;4 if the solution is unique then5 STOP: = true;6 break;7 end8 Solve problem (9) to identify the most

“dissatisfied” sub-coalition S∗ and thecorresponding excess of coalition c∗,k with thecurrent cost allocation xk,∗;

9 if ck,∗ > 0 then10 Add the identified sub-coalition:

Fk := Fk ∪ S∗;11 else12 Identify the active and binding constraints Γk;13 Fk+1 := Fk;14 k := k + 1;15 end16 end

V. CASE STUDY

This section reports the numeric results. We first study thefairness and computational efficiency of the cost allocationbased on nucleolus. We then investigate the enhanced eco-nomic benefits of the ES sharing model over the IES model.

A. Simulation setup

We set up the case studies based on the real building demandprofiles [29] and renewable generation profiles (i.e., wind andsolar power) [30] for one year (i.e., 365 scenarios). To accountfor the complementary feature of energy use in buildings, mul-tiple types of buildings (e.g., office, hotel, school, hospital andrestaurant) are considered. We display the typical demand andrenewable generation profile in Fig. ??. Considering the largenumber of scenarios lead to high computation cost, we chooseS = 10 representative scenarios to capture the patterns ofrenewable generations and building demands, respectively. TheES charging and discharging efficiency is set as ηch, ηdis = 0.9.For the amortized ES capital price, we assume an annualinterest rate r = 0.06 and ES lifetime L = 10 years. Westudy the problem on a daily circle with the time equallydiscretized into T = 24 time slots, corresponding to a decision

interval of ∆ = 1h. We refer to the time-of-use electricityprice of Singapore: cg+

t = 0.1271s$kW h (off-peak 23:00-7:00)and cg+

t = 0.2085s$kW h(peak 8:00-22:00) and demand chargecg,max = 0.1335s$kW (we set selling price as cg-

t = 0). Themaximum trading power with the grid is set as P g,max = 103

kW for each building.B. Fairness and computational efficiency

This part evaluates the fairness and computational effi-ciency of the cost allocation for ES sharing (Algorithm 1).We consider five ES coalition of different scales: N :=3, 5, 8, 10, 20. We compare the cost allocation based onnucleolus with proportional method [22] and Shapley ap-proach [23]. The proportional method is empirical and easyto compute whereas Shapley approach is more sophisticatedbut computationally intensive.• Proportional method: This method distributes the ES

capital cost among the buildings based on their propor-tions of operation cost (electricity bill) reduction relativeto no ES case (benchmark). Specifically, we denote theoperation cost of the buildings with no ES as xOpex ∈ RN

(this is readily available as the buildings have to purchasethe deficiency instantaneously in such case). Besides,we can obtain the operation cost xOpex ∈ RN and theES capital cost yCap ∈ R with the ES sharing modelby solving problem (4). As the operation cost of eachbuilding is readily available through the net metering, theproportional method only distributes the ES capital costof ES sharing model according to

xCapi = (xOpex

i − xOpexi )

yCap

1T (xOpex − xOpex), ∀i ∈ N .

where 1 ∈ RN denotes a N -dimensional unit vector.Thus, the total cost for each building amounts to xOpex

i +

xCapi (∀i ∈ N ).

• Shapley approach: The underlying idea of Shapley ap-proach is to distribute payoff among the players based ontheir marginal contributions to ES coalition. Specifically,for a coalition game (N , ν), the payoff assigned to eachplayer is calculated as [23]

xi(N ) =∑

S⊆N\i

|S|!(|N | − |S| − 1)!

|N |! [ν(S ∪ i)− ν(S)]

where S denotes all subsets of players of N that excludeplayer i. Clearly, the allocated payoff for player i is builtby its marginal contribution to each coalition S: ν(S ∪i)− ν(S).

As aforementioned, we evaluate a cost allocation to be fairif all the players are satisfied. In this paper, we assume allthe building players are profit-oriented and have no otherpreferences, therefore it is reasonable to quantify their dis-satisfaction (satisfaction) by their allocated cost. To accountfor the group of players, we focus on the minimum excessof coalition which can be used as an indicator of minimumdissatisfaction (DSAT) over all the buildings regarding the costallocation. Specifically, we have

DSAT = minS⊆N ,S6=∅,N

e(x,S) (10)

8

Clearly, all players are satisfied (i.e., fair) if the minimumexcess of coalition is non-positive (DSAT ≤ 0). Moreover, weprefer a cost allocation with a more negative DSAT.

For each scale, we apply the three cost allocation methodsto achieve the ex-post cost allocation across the buildingparticipants. For notation, the buildings are labeled by B1-B20 with the allocated cost displayed in TABLE I (the resultsfor N = 20 are omitted due to space limits). First of all, wenote that the total cost for each scale are the same regardlessof the ex-post cost allocation mechanism used. This is causedby the same optimization problem (4) we rely on to computeboth the Shalepy and propertional allocations. However, thereexist some differentials regarding the allocated cost to eachbuilding under the different cost allocation mechanisms, whichlead to the different DSAT as indicated in the last row of eachtable. Particularly, we find that for the scales N := 3, 5, allthe three cost allocation mechanisms can ensure fairness asindicated by the negative DSAT, whereas for the larger scalesN := 8, 10, the fairness is only ensured by the proposedmethod and Shapley approach not the proportional method.This demonstrate that the proposed method and Shapley ap-proach can ensure fairness whereas proportional method mayfail in some cases.

For the computational efficiency, we quantify the computa-tion cost of different methods by the fraction of characteristicfunction required (i.e., the number of coalition value computedby solving problem (4)). The computation cost for differentscales (N := 3, 5, 8, 10, 20) are presented in TABLE II.Particularly, for the nucleolus, we start with the singleton andgrand coalitions (i.e., F1), and we record the number of con-straint generations (i.e., K) performed in the execution. Firstof all, we observe that Shapley approach shows the highestcomputation cost with the entire characteristic function (2N−1coalition value) required. On the contrary, proportional methodis most efficient and only requires N + 1 coalition valueto achieve the cost allocation (N corresponds to computingthe cost for each building with no ES and 1 correspondsto computing the total cost of grand coalition). Notably, weobserve the computation cost with the nucleolus is slightlyhigher than the proportional method but significantly lowerthan the Shapley approach. It’s noteworthy that for N = 10,only 2.54% (26/1023) of the characteristic function is required,and when the scale is increased to N = 20, the computationburden is reduced to less than 1% (88/107). This demonstratesthe superior computational efficiency of nucleolus over theShapley approach. Therefore, we conclude that the nucleolusoutperforms Shapley approach and proportional method byproviding both computation efficiency and fairness.

C. Economic benefits of ES sharingIn this part, we study the enhanced economic benefits of

the ES sharing model (referred to CES model) over the IESmodel. For the IES model, each building invests private ESseparately where the optimal ES sizing and operation areobtained by solving problem (4) with N = 1. Particularly,for the CES model, we consider two settings: without energysharing (CES) and with energy sharing (CES + Share). For

TABLE I: Cost allocations of different methodsScale Build. Proportional Shapley Nucleolus(N) ×102(s$) ×102 (s$) ×102 (s$)

3

B1 2.63 2.64 2.69B2 4.46 4.49 4.44B3 2.40 2.37 2.38

DSAT -5.58(Y) -6.37(Y) -5.80(Y)

Scale Build. Proportional Shapley Nucleolus(N) ×102(s$) ×102(s$) ×102 (s$)

5

B1 2.57 2.57 2.59B2 4.43 4.43 4.48B3 2.30 2.27 2.26B4 6.66 6.72 6.75B5 4.72 4.68 4.60

DSAT -2.00(Y) -6.73(Y) -10.18(Y)

Scale Build. Proportional Shapley Nucleolus(N) ×102(s$) ×102 (s$) ×102 (s$)

8

B1 2.61 2.59 2.64B2 4.44 4.40 4.43B3 2.32 2.28 2.27B4 6.59 6.62 6.60B5 4.66 4.64 4.61B6 5.82 5.79 5.77B7 7.54 7.56 7.54B8 6.69 6.79 6.82

DSAT 8.34(N) -0.64(Y) -4.98(Y)

Scale Build. Proportional Shapley Nucleolus(N) ×102(s$) ×102 (s$) ×102 (s$)

10

B1 2.62 2.58 2.61B2 4.44 4.44 4.43B3 2.22 2.30 2.22B4 6.68 6.59 6.64B5 4.60 4.68 4.59B6 5.74 5.79 5.75B7 7.51 7.48 7.50B8 6.77 6.71 6.82B9 6.44 6.45 6.46B10 7.97 7.96 7.97

DSAT 10.49(N) -1.48(Y) -4.85(Y)Note: Y: satisfied N: not satisfied

TABLE II: Computation of different methodsScale Proportional Shapley Nucleolus(N) Computation Computation Computation

3 4 7 8 (4)5 6 31 13 (7)8 9 255 24 (15)10 11 1023 37 (26)20 21 > 107 88 (67)N N + 1 2N − 1 N + 1 + K(K)

the CES + Share model, we can follow the previous notationsand formulations but replace (3d) with

∑Ni=1 e

b,ωi,t ≥ 0,∀t ∈ T

which indicates the stored energy injected by the differ-ent buildings are commonly owned. Using the scale withN = 5 and N = 10 buildings as examples, we studyboth the building-wise (B1-B10) and community-wise (Com.)economic benefits with the two ES models. The building-wiseeconomic benefits with the CES and CES + Share model arecalculated based on the ex-post cost allocation of nucleolus.The community-wise economic benefits represent the overalleconomic benefits for all the buildings. Using the cost with no

9

ES as baseline, the building-wise (B1-B10) and community-wise (Com.) economic benefits can be quantified by the costreduction as shown in Fig. 2. We see the CES model yieldshigher percentage of cost reduction to each committed buildingand the whole community over the IES model. Taking the casewith N = 5 as an example ((Fig. 2(a))), the cost of B3 (i.e.,electricity bill plus ES capital cost) is cut off by 16.7% withthe CES model versus 4.9% with IES model, and the overallcost is reduced by 9.0% versus 2.5%. This implies the CESmodel can enhance both the building-wise and community-wise economic benefits over the IES model. Besides, we notethat the CES + Share model can enhance the economic benefitssignificantly further. For example, the cost reduction for B3 isup to 59.5% with the CES + Share model. We see the similarresults with the scale N = 10. Further, by comparing theresults with N = 5 and N = 10, we find that B1-B5 (appearin both scales) all gain higher cost reduction with the largercoalition (i.e.,N = 10) (the marginal decrease of B1 is causedby computing accuracy). This demonstrates that by forminga large ES sharing coalition, the economic benefits of thebuilding participants can be further enhance, however this mayrequire a more powerful central coordinator for coordination.

B1 B2 B3 B4 B5 Com.

(a)

0

10

20

30

40

50

60

Re

du

ctio

n (

%)

2.5 0.94.9

1.93.7 2.5

9.5 5.5

16.7

5.2

13.0 9.0

39.2

17.1

59.5

15.0

28.0 26.8

IESCESCES + Share

B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 Com.(b)

0

10

20

30

40

50

60

70

80

Re

du

ctio

n (

%)

2.5 1.04.8

1.9 3.6

9.1 6.6

17.8

6.7

13.1

46.2

22.6

71.8

13.6

32.3

2.1

7.6

19.3

IESCESCES + Share

Fig. 2: The building-wise (B1-B10) and community-wise (Com.) costreduction with the IES and CES model: (a) N = 5. (b) N = 10 (noES as baseline).

Further, we study the average value of ES (VoS) with thedifferent ES models (i.e., IES, CES, CES + Share). The VoS isdefined as the proportion of operation cost reduction over theES capital cost, representing the average return on investment(ROI):

VoS = [xOpex − xOpex]/xCap

where xOpex and xCap denote the optimal operation cost andcapital cost yield by an ES model. xOpex represents theoperation cost with no ES. Intuitively, there exists potential toinvest ES if and only if VoS > 1. Moreover, we would preferan ES model with a higher VoS. In this part, we study the

B1 B2 B3 B4 B5 Com.

(a)

0

1

2

3

4

5

6

VoS

1.4

2.6

6.1IESCESCES + Share

B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 Com.

(b)

0

1

2

3

4

5

6

7

8

VoS

1.4

2.5

7.2IESCESCES + Share

Fig. 3: The building-wise (B1-B10) and community-wise (Com.) VoSwith the IES and CES model: (a) N = 5. (b) N = 10.

VoS for both individual buildings and the whole community.Particularly, the ES capital cost allocated to each buildingwith the CES model can be obtained by the total allocatedcost minus the electricity bill (obtained by solving problem(4)). For the CES + Share model, we only study the VoS forthe whole community due to the lack of ES capital cost forindividual buildings. As with CES + Share model, the costof each building consists of the electricity bill, energy tradingcost, and ES capital cost. However, the latter two parts cannot be distinguished from the cost allocation. Similarly, weuse the case with N = 5 and N = 10 as examples and wepresent the results in Fig. 3. First, we observe that the VoSfor each building is apparently increased with the CES modelover the IES model. Overall, the VoS for the whole communityis about 1.8 times with the CES model than the IES model.Besides, by comparing the results with N = 5 and N = 10,we find that a large ES sharing coalition also favors the VoSfor individual buildings (B1-B5). Notably, we see the CES+ Share model can significantly increase the overall VoS forthe community than the CES model. This is reasonable thatallowing the buildings to share their surplus local renewablegeneration can reduce the over grid purchase for supplying thebuilding demand.

VI. CONCLUSION

This paper studied a cooperative energy storage (ES) busi-ness model based on the sharing mechanism. To maximizethe economic benefits of ES, we studied the problem byintegrating the optimal planning (i.e., ES sizing), operation,and fair ex-post cost allocation via a coalition game formation,thus yielding higher economic benefits to each building andthe whole community over the individual energy storage(IES) model. Particularly, the fair ex-post cost allocationwas achieved based on nucleolus which ensures fairness byminimizing the dissatisfaction of all players. To handle theexponential computation burden, we applied the constraint

10

generation technique to gradually approach the unique nu-cleolus considering the sparse problem structure, which wasdemonstrated with both computation efficiency and fairness.Further, through the case studies, we found that by enablingenergy sharing through the shared ES, the economic benefitsof ES can be further enhanced for the buildings with surpluslocal renewable generation. As the commercial deploymentof ES is currently impeded by the high capital cost, thiswork can work as an example how business model designscan benefit the practice of ES technologies. Currently, wedo not consider the degradation of ES capacity caused bythe charging and discharging circles due to the complexityof quantification, however it seems an interesting work toincorporate the recently developed convex rainflow cycle-based model [31] to address that issue in the future.

APPENDIX APROOF OF PROPOSITION 1

Proof. i) We first prove constraints pg+,ωi,t pg-,ω

i,t = 0 is re-dundant. This is intuitive as pg+,ω

i,t , pg-,ωi,t > 0 won’t happen

simultaneously when the purchase price is higher than theselling price and the objective is to minimize the total cost.

ii) We prove constraints pch,ωi,t pdis,ω

i,t = 0 can be discardedwithout affecting the optimal solution by investigating twopossible cases.

Case I: we assume an optimal operation with pch,ωi,t >

pdis,ωi,t > 0. Since ηch, ηdis < 1, we can construct another feasi-

ble ES operation strategy: pch,ωi,t = pch,ω

i,t − pdis,ωi,t /(ηchηdis) and

pdis,ωi,t = 0. It’s easy to verify that pb,ω

i,t = pch,ωi,t ηch− pdis,ω

i,t /ηdis

and eb,ωi,t+1 = eb,ω

t+1 (change the ES operation at time t butdo not affect the energy state). Consequently, we would havepωi,t = pch,ω

i,t − pdis,ωi,t = pch,ω

i,t − pdis,ωi,t /(ηchηdis) < pch,ω

i,t − pdis,ωi,t

and pg+,ωi,t − p

g-,ωi,t = pωi,t + pd,ω

i,t − pr,ωi,t < pωi,t + pd,ω

i,t − pr,ωi,t =

pg+,ωi,t − pg-,ω

i,t . Therefore, for the constructed operation, wehave pg+,ω

i,t < pg+,ωi,t or pg-,ω

i,t > pg-,ωi,t . This implies the

constructed solution will yield lower cost, which contradictsthe assumption.

Case II: Assume we have an optimal operation with pdis,ωi,t >

pch,ωi,t > 0. Similarly, we can construct another feasible ES

operation strategy: pch,ωi,t = 0 and pdis,ω

i,t = pdis,ωi,t −p

ch,ωi,t ηchηdis.

It’s easy to verify that pb,ωi,t = pb,ω

i,t and eb,ωi,t+1 = eb,ω

i,t (changethe ES operation at time t but do not affect the subsequentenergy state). For the constructed ES operation, we would havepωi,t = pch,ω

i,t − pdis,ωi,t = pch,ω

i,t ηchηdis − pdis,ωi,t < pωi,t. Similarly,

this implies the constructed solution will yield lower cost andthus contradictions.

APPENDIX BPROOF OF THEOREM 1

Proof. We prove the results by definition. Suppose there aretwo disjoint coalitions: S1,S2 ⊆ S and S1 ∩ S2 = ∅. Weassume x∗S1 , [yω(x∗S1)]ω∈Ω and x∗S2 , [yω(x∗S2)]ω∈Ω are theoptimal solution of the coalition S1 and S2, respectively, Sinceboth the objective function and the constraints of problem (4)are linear, it’s easy to verify that (x∗S1 + x∗S2 , [yω(x∗S1) +

yω(x∗S2)]ω∈Ω) are viable solution for the coalition S1 ∪ S2.Therefore, we have

ν(S1 ∪ S2) ≤ ν(S1) + ν(S2) (11)

REFERENCES

[1] F. J. De Sisternes, J. D. Jenkins, and A. Botterud, “Thevalue of energy storage in decarbonizing the electricitysector,” Applied Energy, vol. 175, pp. 368–379, 2016.

[2] “Utility-scale batteries innovation landscape brief.”https://www.irena.org/-/media/Files/IRENA/Agency/Publication/2019/Sep/IRENA Utility-scale-batteries2019.pdf. Accessed: 2020-12-24.

[3] Y. Yu, Z. Cai, and Y. Huang, “Energy Storage Arbitragein Grid-Connected Micro-Grids Under Real-Time MarketPrice Uncertainty: A Double-Q Learning Approach,”IEEE Access, vol. 8, pp. 54456–54464, 2020.

[4] D. S. Mallapragada, N. A. Sepulveda, and J. D. Jenkins,“Long-run system value of battery energy storage infuture grids with increasing wind and solar generation,”Applied Energy, vol. 275, p. 115390, 2020.

[5] W. Hu, Z. Chen, and B. Bak-Jensen, “Optimal operationstrategy of battery energy storage system to real-timeelectricity price in Denmark,” in IEEE PES GeneralMeeting, pp. 1–7, IEEE, 2010.

[6] H. Heinrichs, “Sharing economy: a potential new path-way to sustainability,” GAIA-Ecological Perspectives forScience and Society, vol. 22, no. 4, pp. 228–231, 2013.

[7] P. Lombardi and F. Schwabe, “Sharing economy as anew business model for energy storage systems,” AppliedEnergy, vol. 188, pp. 485–496, 2017.

[8] B. P. Koirala, R. A. Hakvoort, E. C. van Oost, and H. J.van der Windt, “Community energy storage: Governanceand business models,” Consumer, Prosumer, Prosumager:How Service Innovations Will Disrupt the Utility Busi-ness Model, pp. 209–234, 2019.

[9] H. Zhu and K. Ouahada, “Credit-Based Distributed Real-Time Energy Storage Sharing Management,” IEEE Ac-cess, vol. 7, pp. 185821–185838, 2019.

[10] B. P. Koirala, E. van Oost, and H. van der Windt,“Community energy storage: A responsible innovationtowards a sustainable energy system?,” Applied Energy,vol. 231, pp. 570–585, 2018.

[11] A. Fleischhacker, H. Auer, G. Lettner, and A. Botterud,“Sharing solar PV and energy storage in apartment build-ings: resource allocation and pricing,” IEEE Transactionson Smart Grid, vol. 10, no. 4, pp. 3963–3973, 2018.

[12] P. Chakraborty, E. Baeyens, K. Poolla, P. P. Khargonekar,and P. Varaiya, “Sharing storage in a smart grid: Acoalitional game approach,” IEEE Transactions on SmartGrid, vol. 10, no. 4, pp. 4379–4390, 2018.

[13] I. S. Bayram, M. Abdallah, A. Tajer, and K. A. Qaraqe,“A stochastic sizing approach for sharing-based energystorage applications,” IEEE Transactions on Smart Grid,vol. 8, no. 3, pp. 1075–1084, 2015.

11

[14] I. Kim and D. Kim, “Optimal capacity of shared energystorage and photovoltaic system for cooperative resi-dential customers,” in 2017 International Conference onInformation and Communications (ICIC), pp. 293–297,IEEE, 2017.

[15] J. Yao and P. Venkitasubramaniam, “Optimal end userenergy storage sharing in demand response,” in 2015IEEE International Conference on Smart Grid Commu-nications (SmartGridComm), pp. 175–180, IEEE, 2015.

[16] J. Yao and P. Venkitasubramaniam, “Privacy awarestochastic games for distributed end-user energy storagesharing,” IEEE Transactions on Signal and InformationProcessing over Networks, vol. 4, no. 1, pp. 82–95, 2017.

[17] A. Hallefjord, R. Helming, and K. Jørnsten, “Computingthe nucleolus when the characteristic function is givenimplicitly: A constraint generation approach,” Interna-tional Journal of Game Theory, vol. 24, no. 4, pp. 357–372, 1995.

[18] T. AlSkaif, A. C. Luna, M. G. Zapata, J. M. Guerrero,and B. Bellalta, “Reputation-based joint scheduling ofhouseholds appliances and storage in a microgrid with ashared battery,” Energy and Buildings, vol. 138, pp. 228–239, 2017.

[19] W. Zhang, W. Wei, L. Chen, B. Zheng, and S. Mei,“Service pricing and load dispatch of residential sharedenergy storage unit,” Energy, p. 117543, 2020.

[20] D. Zhao, H. Wang, J. Huang, and X. Lin, “Pricing-based energy storage sharing and capacity allocation,”in IEEE International Conference on Communications(ICC), pp. 1–6, IEEE, 2017.

[21] D. Zhao, H. Wang, J. Huang, and X. Lin, “VirtualEnergy Storage Sharing and Capacity Allocation,” IEEETransactions on Smart Grid, 2019.

[22] J. Liu, X. Chen, Y. Xiang, D. Huo, and J. Liu, “Optimalplanning and investment benefit analysis of shared energystorage for electricity retailers,” International Journal ofElectrical Power & Energy Systems, vol. 126, p. 106561.

[23] W. Saad, Z. Han, M. Debbah, A. Hjorungnes, andT. Basar, “Coalitional game theory for communicationnetworks,” IEEE Signal Processing Magazine, vol. 26,no. 5, pp. 77–97, 2009.

[24] P. Harsha and M. Dahleh, “Optimal management andsizing of energy storage under dynamic pricing for theefficient integration of renewable energy,” IEEE Trans-actions on Power Systems, vol. 30, no. 3, pp. 1164–1181,2014.

[25] H. Pandzic, “Optimal battery energy storage investmentin buildings,” Energy and Buildings, vol. 175, pp. 189–198, 2018.

[26] A. S. Schulz and N. A. Uhan, “Approximating theleast core value and least core of cooperative gameswith supermodular costs,” Discrete Optimization, vol. 10,no. 2, pp. 163–180, 2013.

[27] D. B. Gillies, “Solutions to general non-zero-sumgames,” Contributions to the Theory of Games, vol. 4,pp. 47–85, 1959.

[28] D. Schmeidler, “The nucleolus of a characteristic func-tion game,” SIAM Journal on applied mathematics,

vol. 17, no. 6, pp. 1163–1170, 1969.[29] “Openei.” https://openei.org/datasets/files/961/pub/. Ac-

cessed: 2020-10-29.[30] “Measurement and instrumentation data

center (midc), nrel transforming energy.”https://midcdmz.nrel.gov/apps/daily.pl?site=NWTC&start=20010824&yr=2020&mo=1&dy=28. Accessed:2020-10-29.

[31] Y. Shi, B. Xu, Y. Tan, D. Kirschen, and B. Zhang,“Optimal battery control under cycle aging mechanismsin pay for performance settings,” IEEE Transactions onAutomatic Control, vol. 64, no. 6, pp. 2324–2339, 2018.