This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE SYSTEMS JOURNAL 1

Distributed Algorithm for Nonsmooth Resource

Allocation Problems With Nonlinear Constraints
and Its Application to Smart Grids
Zhenhua Deng , Member, IEEE, Tao Chen, and Yan Zhao

Abstract—In this article, we study nondifferentiable resource Related Works: Recently, a number of distributed algorithms
allocation problems (RAPs). In our problem, the decisions of agents have been put forward for RAPs. For example, Lakshmanan
are subject to coupling inequality constraints, local inequality and De Farias [5] and Xiao and Boyd [6] designed dis-
constraints, and local convex set constraints. In contrast to ex-
isting RAPs, the cost functions are nonsmooth and the inequality tributed gradient-based algorithms for RAPs with differentiable
constraints are nonsmooth and nonlinear in our problem. Based cost functions, while Deng et al. [7] developed distributed
on differential inclusions and projection methods, we exploit a subgradient-based algorithms for RAPs with nondifferentiable
fully distributed subgradient-based resource allocation algorithm cost functions. Li et al. [8] and Wang et al. [9] presented dis-
to optimally allocate the network resources. With the help of the tributed resource allocation algorithms via interior-point meth-
set-valued LaSalle invariance principle, we prove the global con-
vergence of the algorithm to the optimal resource allocation of our ods and penalty function methods for RAPs with box constraints,
problem. Finally, our method is applied to the economic dispatch respectively. Further, Liang et al. [10] exploited distributed
problems of smart grids. With our method, the generations of projection-based algorithms for RAPs with convex set con-
generating units converge to the optimal power generation. straints. Moreover, Ghadimi et al. [11] and Anderson et al. [12]
Index Terms—Distributed algorithms, multiagent systems, non- designed distributed algorithms by multistep methods and ap-
linear constraints, nonsmooth analysis, resource allocation. proximate Newton methods, respectively, under which the linear
convergence is achieved. Zargham et al. [13] and Wei et al. [14]
exploited distributed algorithms with superlinear convergence
I. INTRODUCTION by virtue of accelerated dual descent methods and Newton
HERE are various distributed resource allocation prob- methods, respectively. Furthermore, Deng [15] and Deng and
T lems (RAPs) in numerous fields. For example, in power
systems, generating units should minimize the generation costs
Chen [16] developed distributed algorithms for RAPs with
second-order and high-order multiagent systems, respectively.
under the load-demand constraints (see [1], [2]). In telecom- For most of the distributed resource allocation algorithms,
munication systems, the uplink rate should be maximized by such as [5], [6], [8], [10], [11], [12], [13], [14], [15], and [16],
jointly optimizing energy and time under the energy storage con- cost functions are required to be differentiable. However, cost
straints (see [3]). In radar networks, the tracking accuracy should functions are probably nondifferentiable in practical engineer-
be maximized under the given illuminate resource constraints ing. For instance, in smart grids, the fuel payoff functions of gen-
(see [4]). On account of the universal existence of distributed erators with valve-point loading are nondifferentiable (see [17],
RAPs, it is necessary to investigate the problems. [18]). The cost functions are possibly nondifferentiable in band-
In distributed RAPs, there is a group of agents, and each width allocation and multicommodity flow problems as well
agent has a cost function, which usually is called the local cost (see [19], [20]). Therefore, it is of great value to study nonsmooth
function and is unknown to other agents. All agents are coupled RAPs. Besides, subgradient-based methods are often used to
by network resource constraints, and each agent may have its solve optimization problems with nonsmooth cost functions
own local constraints. The aim of these agents is to minimize (see [21]). In addition, continuous-time subgradient-based algo-
the global cost function of networks under those constraints, rithms are always in the form of differential inclusions, because
which is the summation of local cost functions of all agents. In the subgradient of a nonsmooth function is generally a set-valued
order to allocate network resources optimally, all agents need to map. For example, subgradient-based algorithms were exploited
cooperate with each other. in [21], [22], and [23] by differential inclusions for nonsmooth
distributed optimization and game problems.
The decisions of agents may be subject to various constraints
Manuscript received 21 April 2022; revised 7 September 2022; accepted 7 in RAPs, due to resource limitations, production capacities,
December 2022. (Corresponding author: Zhenhua Deng.) and/or security concerns. Noticeably, many distributed resource
The authors are with the School of Automation, Central South University, allocation algorithms are designed for RAPs with linear equality
Changsha 410075, China (e-mail: [email protected]; [email protected];
[email protected]). or inequality constraints, such as [5], [6], [7], [9], [11], [12],
Digital Object Identifier 10.1109/JSYST.2022.3228242 [13], [14], [15], and [24]. However, in practical engineering,

1937-9234 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE SYSTEMS JOURNAL

constraints may be nonlinear. Moreover, linear constraints are Rn | y − x < δ}. col(c1 , . . ., cq ) = [cT1 , . . ., cTq ]T . 1q and 0q
special cases of nonlinear constraints. Furthermore, in a number denote the column vectors of q ones and zeros, respectively.
of fields, some constraints must be satisfied all the time. For diag{c1 , . . ., cq } is the diagonal matrix, in which the elements
example, in smart grids, the output powers of generators must of the main diagonal are c1 , . . ., cq . Iq denotes a q × q identity
satisfy power generation capacities all the time (see [25], [26]). matrix.
These observations motivate us to study nonsmooth RAPs with
nonlinear inequality constraints and convex set constraints. II. PRELIMINARIES
This article aims to investigate constrained nonsmooth RAPs
Some preliminaries about convex analysis, nonsmooth anal-
and develop fully distributed resource allocation algorithms. The
ysis, and graph theory are given in this section.
main contributions are summarized as follows.
1) We study nonsmooth RAPs, which involve nonlinear in-
A. Convex Analysis
equality constraints and convex set constraints. The for-
mulation extends existing RAPs, including [5], [6], [7], In this section, we present some conceptions about convex
[8], [9], [10], [11], [12], [13], [14], and [15], with the con- analysis (see [27]).
sideration of nonsmooth cost functions, nonsmooth local The set Ω is convex if ∀x1 , x2 ∈ Ω and any α ∈ R with 0 ≤
nonlinear inequality constraints, and nonsmooth coupling α ≤ 1,
nonlinear inequality constraints, as well as heterogeneous
αx1 + (1 − α)x2 ∈ Ω.
local convex set constraints. Because of the nonsmooth-
ness of cost functions, the nonlinearity and nonsmoothness The normal cone of a boundary point x1 of a set Ω is defined
of inequality constraints, and the heterogeneousness of as NΩ (x1 ) = {x2 : x2 , x0 − x1  ≤ 0, ∀x0 ∈ Ω, x2 ∈ Rq }.
convex set constraints, it is hard to design distributed The convex hull of a set Ω is defined as
resource allocation algorithms for the problem and analyze
co{Ω} = {α1 x1 + · · · + αn xn | xi ∈ Ω, αi ≥ 0
the convergence.
2) We design a distributed resource allocation algorithm for i ∈ {1, . . . , n}, α1 + · · · + αn = 1}.
our problem. Compared with many distributed algorithms,
A function f : Rq → R is strictly convex if (x1 −
such as [6], [8], [9], [10], [12], and [14], our algorithm
x2 )T (∇f (x1 ) − ∇f (x2 )) > 0, where x1 = x2 .
not only can ensure the decisions satisfy the local convex
The projection of x1 on Ω is defined as ProjΩ (x1 ) =
set constraints during the running of the algorithm, but
arg minx2 ∈Ω x1 − x2 , where x1 ∈ Rq and Ω ⊂ Rq is a closed
also does not have requirements for the initial decisions.
convex set. The following lemmas about the projection hold.
Moreover, our algorithm is fully distributed, which implies
Lemma 1 (see [28, Ch. 1, Th. 2.3]): Denote Ω ⊂ Rq as a
that it is not necessary for our algorithm to use auxiliary al-
closed convex set. Then, we have the following:
gorithms to calculate algorithm parameters, in contrast to
i) ProjΩ (x1 ) − ProjΩ (x2 )2 ≤ ProjΩ (x1 ) −
numerous algorithms like [5], [7], [8], [9], [13], [15], and
ProjΩ (x2 ), x1 − x2 ∀x1 , x2 ∈ Rq ;
[16]. Besides, we analyze its asymptotical convergence
ii) x1 − ProjΩ (x1 ), ProjΩ (x1 ) − x2  ≥ 0, ∀x1 ∈ Rq ,
by set-valued LaSalle invariance principle. By our algo-
∀x2 ∈ Ω.
rithm, all agents converge to the exact optimal allocation,
Lemma 2 (see [29, Ch. 4, Th. 1.7]): Denote Ω ⊂ Rq as
rather than a neighborhood of optimal allocation achieved
a closed convex set. Let ψ(x1 , x2 ) = 12 (x1 − ProjΩ (x2 )2 −
in [10].
x1 − ProjΩ (x1 )2 ). Then, we have the following:
3) The result is applied to the economic dispatch problems of
i) ψ(x1 , x2 ) is continuously differentiable and convex with
smart grids. By taking advantage of our method, the gener-
respect to x1 on Rq , and ∇x1 ψ(x1 , x2 ) = ProjΩ (x1 ) −
ations of generating units match their generation capacities
ProjΩ (x2 );
all the time, and converge to the optimal allocation, even
ii) ψ(x1 , x2 ) ≥ 12 ProjΩ (x1 ) − ProjΩ (x2 )2 .
load demands or network configurations are changeable.
The rest of this article is organized as follows. Section II
provides preliminaries regarding convex analysis, nonsmooth B. Nonsmooth Analysis
analysis, and graph theory. Section III describes the constrained In this section, we give some concepts about nonsmooth
nonsmooth RAPs. Section IV proposes a distributed algorithm analysis (see [30]).
and proves its stability. Section V applies our method to eco- A set-valued map F : Rq ⇒ Rq is a map, which associates
nomic dispatch problems. In the end, Section VI concludes this any y ∈ Rq with a subset F(y) of Rq . A set-valued map F is lo-
article. cally bounded at y ∈ Rq , if ∃σ, o > 0 so that ∀r ∈ F(x) and x ∈
Notation: R denotes the set of real numbers. R+ is the set B(y, o), r ≤ σ holds. A set-valued map F is upper semicon-
of nonnegative real numbers. Z+ denotes the set of nonnega- tinuous at y ∈ Rq , if there exists a neighborhood U of y so that
tive integers. Rq is the q-dimensional Euclidean space. x, y F(U ) ⊂ E holds for any open set E containing F(y). A function
denotes the inner product of vectors x and y. ⊗ is the Kronecker f : Rq → R is said to be locally Lipschitz at y ∈ Rq , if ∃ θy ,
product. × is the Cartesian product. x ◦ y denotes the Hadamard τ > 0 so that f (z) − f (x) ≤ θy z − x∀x, z ∈ B(y, τ ). Let
product for x, y ∈ Rq , i.e., x ◦ y = [x1 y1 , . . . , xq yq ]. x de- Ωnd and O be the nondifferentiable point set of f and a set
notes the standard Euclidean norm of vector x. B(x, δ) = {y ∈ of measure zero, respectively. Then, the generalized gradient

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: DISTRIBUTED ALGORITHM FOR NONSMOOTH RESOURCE ALLOCATION PROBLEMS 3

of f can be defined as follows: ∂f (y) := co{limk→∞ ∇f (yk ) : III. PROBLEM FORMULATION

yk → y, yk ∈ / O ∪ Ωnd }. If f is locally Lipschitz continuous, We study a RAP with N agents, whose communication
then ∂f is an upper semicontinuous map. Moreover, ∂f is network is described by an undirected graph G. Agent i ∈
locally bounded, nonempty, compact and convex. Besides, the {1, 2, . . ., N } has a nonsmooth cost function fi (xi ) : Ci → R,
subgradient of f at y can be denoted as ∂f (y) = {θ ∈ Rq : where xi is the decision of agent i and Ci ⊂ Rn is the local
f (x) − f (y) ≥ θ, x − y∀x ∈ Rq }, if f is convex. closed convex set constraint of agent i. Moreover, xi is con-
A differential inclusion on Rq is strained by the following inequality:
g i (xi ) ≤ 0mi (2)
ẏ(t) ∈ F(y(t)) (1)
where g i (xi ) = with mi ∈ Z+ , and
col(gi1 (xi ), . . . , gimi (xi ))
gil (xi ) : Rn → R is a nonlinear and nonsmooth function with
where F : Rq ⇒ Rq is a set-valued map. A solution of (1), l ∈ {1, . . ., mi }. Furthermore, the decisions of all agents are
denoted as y(t) : [t1 , t2 ] → Rq with [t1 , t2 ] ⊂ [0, +∞), is abso- subject to the following inequality constraint:
lutely continuous, if y(t) satisfies (1) for almost all t ∈ [t1 , t2 ].

N
The existence theorem about the solution of (1) is presented as hi (xi ) ≤ 0q (3)
follows. i=1
Lemma 3 (see [31, Lemma 2.1]): Suppose that F is upper
where hi (xi ) = col(h1i (xi ), . . . , hqi (xi )) with q ∈ Z+ , and
semicontinuous, locally bounded, and has nonempty, compact,
hki (xi ) : Rn → R is a nonlinear and nonsmooth function with
and convex values. Then, for arbitrary initial value, there exists
k ∈ {1, . . ., q}.
a solution of (1).
The purpose of all agents is to minimize the cost function
If at least one solution y(t) of (1) is contained in S with
of networks f (x) under the convex set constraint C, the local
t ∈ [0, +∞), then S is called a weakly positively invariant set
inequality constraint
(2) and the coupling inequality constraint
of (1). Analogously, if any solution y(t) of (1) is contained in S
(3), wheref (x) = N i=1 fi (xi ) with x = col(x1 , . . . , xN ) and
with t ∈ [0, +∞), then S is called a strongly positively invariant
C = C1 × · · · × CN . Particularly, all agents face the following
set of (1). Besides, the set of equilibrium points (EPs) of (1) is
RAP:
denoted as {y ∈ Rq : 0q ∈ F(y)}.
Suppose that V (y) : Rq → R is a differentiable function, and 
N

then, we denote LF V (y) := {(∇V (y))T v : v ∈ F(y)} as the min f (x), f (x) = fi (xi )
x∈RN n
i=1
set-valued Lie derivative of V with respect to F at y.
We give the set-valued LaSalle invariance principle for (1) as 
N

follows. subject to hi (xi ) ≤ 0q

Lemma 4 (see [22, Lemma 2.1]): If max LF V (y) ≤ 0 or i=1

LF V (y) = ∅ ∀y ∈ S, and the evolutions of (1) are bounded, g i (xi ) ≤ 0mi , xi ∈ Ci , i ∈ {1, . . ., N }. (4)
where V (y) : Rq → R is a differentiable function, S ⊂ Rq is
a strongly positively invariant set of (1), then the solutions The aim of this article is to design a distributed resource allo-
of (1) converge to the largest weakly positively invariant set cation algorithm for agents so that their decisions can approach
Φ ⊂ {S ∩ {y ∈ Rq | 0 ∈ LF V (y)}}. Besides, if Φ is finite, to the optimal resource allocation of the RAP (4).
each solution converges to an element of Φ. In our problem, there exist three types of constraints, i.e., the
set Ci , the inequality (2) and the inequality (3), which are all
convex. In order to distinguish the set constraint Ci from the
inequality constraints, the set constraint Ci is called as convex
C. Graph Theory set constraints. Besides, here the convex set constraint Ci should
In this section, we introduce the notation about graph theory be satisfied all the time, and the inequality constraints are not
(see [32]). required to be satisfied all the time.
An undirected graph with n vertexes, denoted as G = We give some assumptions about cost functions and con-
{N , E, A}, is made up of a vertex set N = {n1 , n2 , . . . , nn }, straints as follows.
an edge set E ⊆ N × N , and an adjacent matrix A = [akl ]n×n . Assumption 1: The undirected graph G is connected.
(nk , nl ) ∈ E is an edge of G if nk and nl can communicate Assumption 2: fi (xi ) is strictly convex and locally Lipschitz
with each other. akl is the weighting of the edge (nk , nl ), where continuous, and hi (xi ), g i (xi ) are convex and locally Lipschitz
akl > 0 if (nk , nl ) ∈ E and akl = 0, otherwise. Besides,
 akk = continuous ∀i ∈ {1, 2, . . ., N }.
0, ∀k ∈ N . The degree of nk is defined as dk = nl=1 akl . Assumption N3: There exists relative interior point xi ∈ Ci
The Laplacian matrix of G is defined as L = D − A, where such that i=1 hi (xi ) < 0q and g i (xi ) < 0mi hold ∀i ∈
D = diag{d1 , d2 , . . . , dn }, which implies L is symmetric and {1, 2, . . ., N }.
positive semidefinite, and L1N = 0N . When a path exists be- For the reason of Assumption 3, the optimal allocation of the
tween any two vertices, then the undirected graph G is said to RAP (4) exists.
be connected. Denote ϑ1 , . . ., ϑn as the eigenvalues of L with On the basis of [33, Th. 3.34], we have the following lemma
ϑk ≤ ϑl for k ≤ l. Moreover, ϑ2 > 0, if G is connected. about the optimality condition of the RAP (4).

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE SYSTEMS JOURNAL

Lemma 5: Under Assumptions 2 and 3, x∗i ∈ Ci (i ∈ Step-by-step procedure of our algorithm

{1, . . . , N }) is the optimal resource allocation of the RAP (4) if
and only if there exist wpk∗ ∈ R+ , l∗ Initialization: Set xi (0) ∈ Rn , i (0) ∈ Rmi , i (0) ∈ Rq ,
i ∈ R+ with k ∈ {1, . . ., q}
and l ∈ {1, . . ., mi } such that vi (0) ∈ Rq , yi (0) ∈ Rn and stepsize.
for k = 1, . . ., do:
0n ∈ ∂fi (x∗i ) + ∂hi (x∗i )T wp∗ + ∂g i (x∗i )T ∗
i + NCi (x∗i ) (5a) Step 1: Agent i receives its neighbors’ ˜ j (k − 1) and
vj (k − 1).

N 
N
hi (x∗i ) ≤ 0q , wpk∗ hki (x∗i ) = 0 (5b) Step 2: Agent i calculates:
i=1 i=1 dyi ∈ xi (k − 1) − ∂fi (xi (k − 1)) − yi (k − 1)
g i (x∗i ) ≤ 0mi , l∗ l ∗
i gi (xi ) =0 (5c)
− ∂hi (xi (k − 1))T ˜ i (k − 1)
mi ∗
where =wp∗ col(wp1∗ , . . . , wpq∗ )
and = col( ∗
i
1∗
i ,..., i ).
Remark 1: Our problem has the following characteristics. − ∂g i (xi (k − 1))T ˜i (k − 1).
1) The cost functions and inequality constraints in our prob- d i = − i (k − 1) + ( i (k − 1) + gi (xi (k − 1)))+ .
lem are nonsmooth rather than differentiable required
by huge amounts of results like [5], [6], [8], [10], [11], d i = − i (k − 1) + ˜ i (k − 1) + hi (xi (k − 1))
[12], [13], [14], [15], and [16]. The nonsmoothness of 
N
cost functions and inequality constraints result that their − aij ( ˜ i (k − 1) − ˜ j (k − 1))
subgradients are nonglobally Lipschitz continuous, which j=1
further pose obstacles to the algorithm design and analysis
of our problem. 
N
+ aij (vi (k − 1) − vj (k − 1)).
2) The local constraints in our problem contain both the local
j=1
nonlinear and nonsmooth inequalities and the heteroge-
neous convex sets, which are more general than the case 
N

with only local convex sets, local linear inequalities as dvi = − aij ( ˜ i (k − 1) − ˜ j (k − 1)).
well as the case without any local constraints considered j=1

by masses of works, including [5], [6], [7], [9], [10], [11], Step 3: Agent i updates yi (k), i (k), i (k) and vi (k):
[12], [13], [14], [15], and [16]. yi (k) = yi (k − 1) + dyi × stepsize.
3) The coupling constraints are also nonlinear and nons- i (k) = i (k − 1) + d i × stepsize.
mooth inequalities in our problem, which extend the cou- i (k) = i (k − 1) + d i × stepsize.
pling linear equality or inequality constraints investigated vi (k) = vi (k − 1) + dvi × stepsize.
in [5], [6], [7], [9], [11], [12], [13], [14], [15], and [24]. Step 4: Agent i obtains xi (k):
xi (k) = ProjCi (yi (k)).
IV. MAIN RESULTS Step 5: Agent i sends ˜ i (k) and vi (k) to its neighbors.
In Section IV-A, we design a distributed resource allocation end for
algorithm for the RAP (4). Then, in Section IV-B, the algorithm
is analyzed.

A. Distributed Resource Allocation Algorithm Design Remark 2: The algorithm (6) is made up of four parts:
i) ∂fi (xi ) for seeking the optimal solution of RAP (4); ii)
The distributed algorithm is designed as follows: ProjCi (·) and yi for ensuring xi (t) ∈ Ci ∀t > 0, even though
ẏi ∈ xi − ∂fi (xi ) − ∂hi (xi )T ˜ i − ∂g i (xi )T ˜i − yi (6a) xi (0) may be out of Ci ; iii) ˜ i and vi for the coupling in-
equality constraint (3); iv) ˜i for the local inequality con-
˙i = − i +( i + gi (xi ))+ (6b) straint (2). The schematic of algorithm (6) is described in

N Fig. 1.
˙i= − i + ˜ i + hi (xi ) − aij ( ˜ i − ˜ j ) Remark 3: Our algorithm has the following features.
j=1 1) Compared with the algorithms in [6], [8], [9], [10], [12],

N and [14], the algorithm (6) does not have requirements for
+ aij (vi − vj ) (6c) the initial conditions, and can ensure that the decisions
j=1 always belong to the set C during the running of the

N algorithm.
v̇i = − aij ( ˜ i − ˜ j ) (6d) 2) The algorithm (6) is fully distributed, because it avoids
j=1 introducing auxiliary algorithms to calculate some global
parameters required by many algorithms, such as [5], [7],
xi = ProjCi (yi ) (6e)
[8], [9], [13], [15], and [16].
where ˜i = ( i + gi (xi ))+ , ˜i = ( i)
+
, and (x)+ = Remark 4: The step-by-step procedure of (6) is given at the
max{x, 0}. top of this page.

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: DISTRIBUTED ALGORITHM FOR NONSMOOTH RESOURCE ALLOCATION PROBLEMS 5

Because G is connected, i.e., L1N = 0N and 1TN L = 0TN , it

results from (8d) that
˜ i∗ = ˜ j∗ ∀i, j ∈ {1, . . ., N }. (9)

Left multiplying 1TN to (8c), we have


N 
N 
N
∗
i − ˜ i∗ = hi (x∗i ). (10)
i=1 i=1 i=1

Note that ˜ i = ( i)
+
and ˜ i∗ − ∗
i ≥ 0q . Hence, we have
˜ i∗ ≥ 0q ∀i ∈ {1, . . ., N } (11)

N
hi (x∗i ) ≤ 0q . (12)
i=1

Moreover, if ˜ i∗ = 0q , then
Fig. 1. Schematic of algorithm (6).

N
˜ ik∗ hki (x∗i ) = 0 ∀k ∈ {1, . . ., q} (13)
B. Convergence Analysis
i=1
To analyze the algorithm (6) conveniently, we rewrite it into 
and if ˜ i∗ > 0q , we have ik∗ = ˜ ik∗ and N ∗
i=1 hi (xi ) = 0q ,
the following compact form: N k ∗
i.e., i=1 hi (xi ) = 0 ∀k ∈ {1, . . ., q}, which implies that
ẏ ∈ x − ∂f (x) − ∂h(x)T 
˜ − ∂g(x)T ˜ − y (7a) (13) is also satisfied.
˙ = −  + ˜ (7b) By (8b) and the definition of , we have ∗ = (∗ +
g(x∗ ))+ . Obviously, ∗ ≥ 0N mi and g(x∗ ) ≤ 0N mi .

˙ = −+
˜ + h(x) + (L ⊗ Iq )v − (L ⊗ Iq )
˜ (7c) i=1 i=1
If ∗ > 0N mi , then g(x)∗ = 0N mi , which means that
i=1 i=1
v̇ = − (L ⊗ Iq )
˜ (7d) ∗ ◦ g(x∗ ) = 0N mi . If ∗ = 0N mi , we also have ∗ ◦
i=1 i=1

x = ProjC (y) (7e) g(x∗ ) = 0N mi . Therefore, we have

i=1

where y = col(y1 , . . . , yN ), x = col(x1 , . . . , xN ),  = col gil (x∗i ) ≤ 0, l ∈ {1, . . ., mi } (14a)

( 1 , . . . , N ), i = col( 1i , . . . , m i ), g(x) = col(g 1 (x1 ), . . . ,
i
l ∗
g N (xN )), g i (xi ) = col(gi (xi ), . . . , gimi (xi )),  = col( 1 ,
l i gi (xi ) = 0, l ∈ {1, . . ., mi }.
˜l∗ (14b)
q
i = col( i , . . . , i ), h(x) = col(h1 (x1 ), . . . ,
1
. . . , N ), Equations (8a) and (8e) imply that x∗ = ProjC (x∗ −
hN (xN )), hi (xi ) = col(hi (xi ), . . . , hqi (xi )), v = col(v1 ,
1
˜ ∗ − ∂g(x∗ )T ˜∗ ), which indicates that
∂f (x∗ ) − ∂h(x∗ )T 
. . . , vN ), vi = col(vi1 , . . . , viq ), and C = C1 × · · · × CN .
In the light of Assumption 2, there exists a solution to (7), on ˜ ∗ + ∂g(x∗ )T ˜∗ + NC (x∗ ).
0N n ∈ ∂f (x∗ ) + ∂h(x∗ )T 
the basis of Lemma 3. (15)
We draw the following conclusion about the relation between
By Lemma 5, (8), (9), (10), (11), (12), (13), (14), and (15)
the EP of algorithm (7) and the optimal solution of RAP (4).
indicate that x∗ is the optimal resource allocation of the RAP
Theorem 1: Under Assumptions 1–3, if (x∗ , y ∗ , ∗ ,  ∗ , v ∗ )
(4).
is the EP of (7), then x∗ is the optimal resource allocation
(ii) When x∗ is the optimal resource allocation of RAP
of the RAP (4). Conversely, if x∗ is the optimal resource al-
(4), then there exist wp∗ = col(wp1∗ , . . . , wpq∗ ) ∈ Rq+ and ∗i =
location of the RAP (4), then there exist (y ∗ , ∗ ,  ∗ , v ∗ ) ∈ mi ∗
N n× N
 col( 1∗i ,..., i ) ∈ Rm i
+ so that
R i=1 mi ×N q×N q such that (x∗ , y ∗ , ∗ ,  ∗ , v ∗ ) is an EP

of (7). 0n ∈ ∂fi (x∗i ) + ∂hi (x∗i )T wp∗ + ∂gi (xi )T ∗

i + NCi (x∗i )
Proof: (i) if (x∗ , y ∗ , ∗ ,  ∗ , v ∗ ) is the EP of (7), then
(16a)
˜ ∗ − ∂g(x∗ )T ˜∗ − y ∗
0N n ∈ x∗ − ∂f (x∗ ) − ∂h(x∗ )T 
(8a) 
N
hi (x∗i ) ≤ 0q (16b)
∗ ∗
0N mi = −  + ˜ (8b) i=1
i=1

0N q = −  ∗ + 
˜ ∗ + h(x∗ ) + (L ⊗ Iq )v ∗ 
N
wpk∗ hki (x∗i ) = 0, k ∈ {1, . . ., q} (16c)
∗
− (L ⊗ Iq )
˜ (8c) i=1

˜∗
0N q = − (L ⊗ Iq ) (8d) g i (x∗i ) ≤ 0mi (16d)
x∗ = ProjC (y ∗ ). (8e) l∗ l ∗
i gi (xi ) = 0, l ∈ {1, . . ., mi }. (16e)

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE SYSTEMS JOURNAL

 ∗
Besides, from (16b) and (16c), if N i=1 hi (xi ) < 0N q , then
It follows from Lemma 1 (i) that
∗
 ∗ ∗
1 N
we choose  i = N i=1 hi (xi ) with  ˜ i = max{ ∗i , 0q }, (x − x∗ )T (y − y ∗ ) ≥ x − x∗ 2 . (19)
N  
such that i=1  i −
∗ N
i=1  ˜ ∗i = i=1 hi (x∗i ) holds. If
N
N ∗ ∗ Lemma 1 (ii) implies that ( + g(x) − ) ˜ T (˜ − ∗ ) ≥ 0,
i=1 hi (xi ) = 0N q , then w ≥ 0N q , and we choose ∗ =
∗ ∗ ∗
N ∗
which can be written as ( − ) ˜ T (˜ − ∗ ) ≥ −g(x)T (˜ −
1N ⊗ wp with  ˜ = max{ , 0N q } such that i=1  i − ∗ ). Subsequently, together with the convexity of g(x), we have
N 
i=1 ˜∗= N ∗
i=1 hi (xi ) also holds. Consequently, we
iN ∗ − (x − x∗ )T (γ T 
˜ − γ ∗T ∗ ) − ( − ∗ )T ( − 
˜)
have ∗
i=1 ( i −  ˜ i − hi (x∗i )) = (− ∗ +  ˜ ∗ + h(x∗ ))T
(1N ⊗ ξ) = 0N q with ξ ∈ R . Based on the theorem
q
≤ − (x − x∗ )T (γ T 
˜ − γ ∗T ∗ )
of linear algebra (see [34]), RN q can be orthogonally
decomposed by ker(L ⊗ Iq ) and range(L ⊗ Iq ). Obviously,  − ∗ ) −  − 
+ g(x)T (˜ ˜ 2
1N ⊗ ξ ∈ ker(L ⊗ Iq ). Therefore, (− ∗ +  ˜ ∗ + h(x∗ )) ∈
∗ ˜ T (g(x∗ ) − g(x) − γ T (x∗ − x)) − ∗T (g(x)
= −
range(L ⊗ Iq ). Hence, there exists v ∈ RN q such that
(L ⊗ Iq )v ∗ = − ∗ +  ˜ ∗ + h(x∗ ) holds, i.e., (8c) holds. − g(x∗ ) − γ ∗T (x − x∗ )) + 
˜ T g(x∗ ) −  − 
˜ 2
According to (16d) and (16e), if g i (x∗i ) < 0mi , we have ∗i =
0mi . Further, ( ∗i + g i (x∗i ))+ = 0mi , i.e., ( ∗i + g i (x∗i ))+ − ≤ −  − 
˜ 2 . (20)
∗ ∗ ∗ ∗ + ∗
i = 0mi . If g i (xi ) = 0mi , we have ( i + g i (xi )) = i , Assumption 2 means that
∗ ∗ + ∗
which also indicates that ( i + g i (xi )) − i = 0mi . Accord-
ingly, by letting ∗ = col( ∗1 , . . . , ∗N ), (8b) is satisfied. (x − x∗ )T (β ∗ − β) ≤ 0. (21)
Take y ∗ = col(y1∗ , . . . , yN ∗
) with yi∗ ∈ x∗i − ∂fi (x∗i ) − ˜ ∗ = 0N q , we have
Because L
∗ T ∗ T
∂hi (xi ) ˜ i − ∂gi (xi ) ˜i . Based on (16a), (8a), and (8e)
hold. ( ˜ ∗ )T L(v − v ∗ ) − (v − v ∗ )T L
˜ − ˜ = 0. (22)
Therefore, (x∗ , y ∗ , ∗ ,  ∗ , v ∗ ) is the EP of (7).  By Lemma 2 (ii), we have
According to Theorem 1, if (7) is stable, then all agents
converge to the optimal resource allocation of the RAP (4).  ˜ ∗ 2 ≤ (
˜ − ˜ ∗ )T ( −  ∗ ).
˜ − (23)
Next, we analyze the stability of (7), which yields the following ˜ ∗ ≥ 0N q . Hence,
Note that hi (xi ) is a convex function and 
theorem. ∗
if 
˜ − ˜ ≥ 0N q , we have
Theorem 2: Under Assumptions 1–3, the algorithm (6) con-
verges to the optimal resource allocation of the RAP (4). ( ˜ ∗ )T (h(x) − h(x∗ ) − (x − x∗ )T η)
˜ −
Proof: Without loss of generality, let n = 1 for simplicity.
˜ ∗ )T (η − η ∗ )T (x − x∗ ) ≤ 0
−( (24)
Let V (y, , , v) = 12 (y−ProjC (y ∗ )2 −y−ProjC (y)
 ) + 12 ( − 
2 ˜ ∗ 2 −  −  ˜ 2 ) + 12 v − v ∗ 2 + 12  − and if  ˜ ∗ < 0N q , by the convexity of h(x), we have
˜ −
∗ 2
  . By Lemma 2 (i), we have
( ˜ ∗ )T (h(x) − h(x∗ ) − (x − x∗ )T η)
˜ −
1 1
V (y, , , v) ≥ x − x∗ 2 +  − ∗ 2 ˜ ∗ )T (η − η ∗ )T (x − x∗ )
− (
2 2
1 1 ˜ T (h(x) − h(x∗ ) − (x − x∗ )T η)
= ()
+  ˜ − ˜ ∗ 2 + v − v ∗ 2 . (17)
2 2 ˜ ∗ )T (h(x) − h(x∗ ) − (x − x∗ )T η ∗ ) ≤ 0.
− ( (25)
Denote L(6) V (y, , , v) as the set-valued Lie deriva- G is connected, and then, we have
tion of V (y, , , v) with respect to (7). For any ζ ∈
L(6) V (y, , , v), there exist β ∈ ∂f (x), γ ∈ ∂g(x) and η ∈ ( ˜ ∗ )T L(
˜ − ˜ ∗ ) ≥ 0.
˜ − (26)
∂h(x) so that With (18), (19), (20), (21), (22), (23), (25), (24), and (26), we
∗
ζ = − (x − x )(y − y ) + x − x  ∗ ∗ 2 have
˜ 2 − (x − x∗ )T (β − β ∗ ) ≤ 0.
ζ ≤ − −  (27)
− (x − x∗ )T (γ T 
˜ − γ ∗T ∗ ) − ( − ∗ )T ( − 
˜)
By reason of the arbitrariness of ζ, (27) implies that
+ (x − x∗ )T (β ∗ − β) − ( ˜ ∗ )T ( −  ∗ )
˜ −
maxL(6) V (y, , , v) ≤ 0. (28)
+  ˜ ∗ 2 + (
˜ − ˜ ∗ )T
˜ −
Equations (17) and (28) imply that the boundedness of
× (h(x) − h(x∗ ) − η T (x − x∗ )) x(t), (t),
˜ (t) and v(t), which mean that ˜ is bounded.
Further, by the locally Lipschitz property of f (x), h(x),
˜ ∗ )T (η − η ∗ )T (x − x∗ ) − (−
− ( ˜ ˜ ∗ )T L( ˜ ∗)
˜ −
and g(x), ∂f (x), ∂h(x), and ∂g(x) are bounded. Then,
+ (v − v ∗ )T L( ˜ ∗ ) − (v − v ∗ )T L
˜ − ˜ (18) we have x(t) − ∂f (x(t)) − ∂h(x(t))T 
˜ − ∂g(x(t))T 
˜ ≤
ψ∀t > 0, where ψ ∈ R is a positive constant. Hence, (7a)
where L = L ⊗ Iq , β ∗ ∈ ∂f (x∗ ), γ ∗ ∈ ∂g(x∗ ), and η ∗ ∈ can be viewed as a stable first-order system about y(t)
∂h(x∗ ). with bounded inputs, which indicates that y(t) is bounded.

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: DISTRIBUTED ALGORITHM FOR NONSMOOTH RESOURCE ALLOCATION PROBLEMS 7

In the same way, we can get that (t) is also bounded TABLE I
PARAMETERS OF FIVE GENERATING UNITS
by (7c). Consequently, based on Lemma 4, the algo-
rithm (7) converges to the largest weakly positively
N
invari-
ant set Φ ⊂ Ξ, where Ξ = {(x, ) ∈ R × R Nn i=1 mi
:0=
max L(6) V (y, , , v)}. Moreover, (x − x∗ )T (β − β ∗ ) >
0∀x = x∗ . Thus, limt→∞ x(t) = x∗ . 
Remark 5: In contrast to the algorithms in [10] converging to
the suboptimal resource allocation, our algorithm can converge
to the exact optimal resource allocation of the RAP (4).
smart grid RAPs with coupling inequality constraints.
V. APPLICATION Moreover, many existing algorithms, including [2], [36],
[37], [35], [38], [40], [41], [47], [42], [43], [44], [45],
In this section, our method is applied to the economic dispatch and [46], can only solve the smart grid RAPs with linear
problems of smart grids. constraints, while our algorithm can handle RAPs with
In smart grids, each generating unit has its own generation nonlinear constraints. It is well known that linear equality
cost function and generation capacity, and the generations of constraints are special cases of nonlinear inequality con-
all generating units should satisfy load demands (see [8], [9], straints. Therefore, our algorithm can solve more smart
[15]). Explicitly, the economic dispatch problem of smart grids grid RAPs.
is written as follows (see [7], [35]): 2) Many algorithms like [2], [36], [37], [35], [38], [40], [42],
[43], [44], [46], and [47] are only suitable for the smart

N
min f (PG ), f (PG ) = fi (PGi ) RAPs with differentiable cost functions, while the cost
PG ∈RN functions in smart grids may be nonsmooth (see [17],
i=1
[18]). Those algorithms may not solve the nonsmooth

N 
N
smart grid RAPs. By comparison, our algorithm can solve
subject to PGi ≥ Pdi , PGi ∈ [PGi
min max
PGi ] (29)
i=1 i=1
smart grid RAPs with nonsmooth cost functions, which
extends the applicability of our algorithm.
where fi (PGi ) is the generation cost of generating unit i, in 3) Most of existing algorithms for smart grid RAPs require
M $; PGi is the generation of generating unit i, in M W ; PG = that the initial decisions must satisfy the constraints, and
min
col(PG1 , . . . , PGN ); Pdi is the local load demand, in M W ; PGi the constraints cannot be destructed during the running of
max
and PGi are the lower and upper bounds of PGi , respectively. the algorithms, such as [41], [42], [43], [44], [45], and [46].
For the problem (29), the algorithm (6) becomes However, in smart grids, charging units and/or generation
⎧ systems are often intermittently added to or deleted from
⎪
⎪ ẏi ∈ PGi − ∂fi (PGi ) + ˜ i − yi
⎪
⎪ N the grid, which often causes the violation of constraints.
⎪
⎨ ˙ i = − i + ˜ i + Pdi − PGi − j=1 aij ( ˜ i − ˜ j ) Thus, the initialization has to be reperformed whenever
+ N aij (vi − vj ) charging units and/or generation systems plug in or leave
⎪
⎪ j=1
⎪
⎪
N
v̇i = − j=1 aij ( ˜ i − ˜ j ) off, which considerably degrades the applicability of al-
⎪
⎩
PGi = ProjCi (yi ) gorithms (see [48] and references therein). Remarkably,
(30) our algorithm has no requirements for initial decisions.
where Ci = [PGi PGi ] and ˜ i = ( i )+ .
min max 4) Many existing algorithms for smart grid RAPs have spe-
Consequently, we have the following corollary about the cial requirements for algorithm parameters, such as [2],
economic dispatch problem (29) with the algorithm (30). [37], [38], [41], [43], [44], [45], [46], and [48], while our
Corollary 1: Under Assumptions 1–3, the algorithm (30) algorithm does not have requirements for parameters.
converges to the optimal economic dispatch of problem (29). Next, the effectiveness of the proposed methods is validated
Proof: The problem (29) and the algorithm (30) are special by numerical examples. Besides, the performance evaluations
cases of the problem (4) and the algorithm (6), respectively. against plug-and-play and load changes are given.
Therefore, similarly to the proof of Theorem 2, the result is
yielded.  A. Effectiveness Validation of the Proposed Methods
Remark 6: Compared with existing algorithms for smart grid
RAPs, our algorithm has the following merits. Case 1: Consider the economic dispatch problem in the IEEE
1) Most of existing algorithms, such as [2], [36], [37], [38], 14-bus system with five generating units (see [49]). The commu-
[39], [40], [41], [42], [43], [44], [45], and [46], can nication network of generating units is an undirected ring graph.
only deal with smart grid RAPs with coupling equality The generation cost functions of generating units are (see [7],
constraints. However, there are coupling inequality con-
2
[50]): fi (PGi ) = αi + βi PGi + γi PGi , where αi , βi , γi , and ci
straints in smart grids, since, sometimes, the generated are constants. For simplicity, we use an equivalent function by
α̃i )2
energies should be greater than load demands (see [35]). changing a constant term: fi (PGi ) = (PGi2+β̃i
. Moreover, the
Our algorithm can not only deal with smart grid RAPs parameters of the five generating units are showed in Table I,
with coupling equality constraints, but also cope with which are borrowed from [49]. The local load demands are

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE SYSTEMS JOURNAL

Fig. 4. Evolutions of PGi with the algorithm in [35].

Fig. 2. Evolutions of PGi with the algorithm (30).

Fig. 5. Evolution of total generation with the algorithm in [35].

Fig. 3. Evolution of total generation with the algorithm (30).

B. Performance Evaluation Against Plug-and-Play

col(Pd1 , . . . , Pd5 ) = col(76, 53, 41, 74, 46). The initial genera-
tions of these five generating units are PG1 (0) = 0, PG2 (0) = 0, Case 2: The distributed algorithms with the plug-and-play
PG3 (0) = 0, PG4 (0) = 0, and PG5 (0) = 0, respectively. feature can improve the scalability and robustness of smart grids
Figs. 2 and 3 present the simulation results of algorithm (30). (see [51]). This case performs the plug-and-play test of the al-
In Fig. 2, the dot–dashed lines are the optimal generation and gorithm (30). The generation cost functions, initial generations,
the solid lines are the evolutions of generations. Fig. 3 reflects and parameters of generating units are the same as in Case 1.
the total generation of the five generating units. At 200 s, generating unit 2 is physically plugged out from the
Fig. 2 shows that the generations of five generating units grid, and the communication edges associated with it are also
meet the local constraints all the time, and converge to the removed. At 400 s, generating unit 2 is plugged in with its
∗ ∗ ∗
optimal generation (PG1 = 64.5031, PG2 = 69.3555, PG3 = communication edges.
∗ ∗
45.1815, PG4 = 52.7013, and PG5 = 58.2574). Fig. 3 indi- Figs. 6 and 7 show the simulation results. We can see from
cates that the coupling constraint holds at the optimal generation. Figs. 6 and 7 that generation cost minimization and load de-
For comparison, we give the results of the algorithm in [35], mand are satisfied even when plug-and-play is considered. To
which is also designed for smart grids RAPs with inequality be more specific, when generating unit 2 is plugged out from
constraints. The simulation results of the algorithm in [35] are the grid at 200 s, the remaining generating units increase their
shown in Figs. 4 and 5. It is clear from Figs. 4 and 5 that generations to the new optimal point, and the generations in
the algorithms in [35] cannot ensure that the generations of [400, 600] s are recovered to the optimal point as that in
generating units always meet the generation capacities during [0, 200] s when generating unit 2 is plugged back at 400 s.
the running of the algorithm (i.e., sometimes, the generations This case illustrates that the algorithm (30) has the ability of
min
are less than the minimum generation capacity PGi = 10). plug-and-play.

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

DENG et al.: DISTRIBUTED ALGORITHM FOR NONSMOOTH RESOURCE ALLOCATION PROBLEMS 9

Fig. 6. Evolutions of PGi with the algorithm (30). Fig. 8. Evolutions of PGi with the algorithm (30).

Fig. 7. Evolution of total generation with the algorithm (30). Fig. 9. Evolution of total generation with the algorithm (30).

C. Performance Evaluation Against Load Changes constraints are nonlinear and nonsmooth, and the local convex
sets are heterogeneous and private-known. Based on differential
Case 3: The load demand of smart grids often changes due inclusions and projections, a distributed algorithm has been
to the uncertainty of consumption. This case will evaluate the designed for our problem. The algorithm is fully distributed, and
performance of the algorithm (30) under load changes. The does not have requirements for the initial conditions. By taking
generation cost functions, initial generations, and parameters advantage of set-valued LaSalle invariance principle, this article
of generating units are the same as in Case 1 when t ∈ [0, 150] has proved the global convergence of the algorithm to the exact
s. Besides, when t ∈ [150, 300] s, Pdi is randomly chosen in optimal resource allocation of nonsmooth RAPs. Finally, our
[60, 70] MW. method is applied to the economic dispatch problems of smart
Figs. 8 and 9 present the simulation results. Figs. 8 and 9 grids, and examples have demonstrated our method.
indicate that generating units with the algorithm (30) can find
the optimal generation even though the load demand changes
REFERENCES
at 150 s. This case explains that the algorithm (30) has the
satisfactory robustness against load changes. [1] W. Wei, F. Liu, S. Mei, and Y. Hou, “Robust energy and reserve dispatch
under variable renewable generation,” IEEE Trans. Smart Grid, vol. 6,
no. 1, pp. 369–380, Jan. 2015.
VI. CONCLUSION [2] D. Yang, S. Zhang, B. Zhou, and S. Bu, “Consensus-based decentralized
optimization for distributed generators power allocation over time-varying
This article has investigated the nonsmooth RAPs of mul- digraphs in microgrids,” IEEE Syst. J., vol. 15, no. 1, pp. 814–825,
tiagent systems. In our problem, every agent has a cost func- Mar. 2021.
[3] H. Lee, K.-J. Lee, H. Kim, B. Clerckx, and I. Lee, “Resource allo-
tion, which is nondifferentiable, and all agents are subject to cation techniques for wireless powered communication networks with
coupling inequality constraints, local inequality constraints, and energy storage constraint,” IEEE Trans. Wireless Commun., vol. 15, no. 4,
local convex set constraints. The coupling and local inequality pp. 2619–2628, Apr. 2016.

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE SYSTEMS JOURNAL

[4] J. Yan, W. Pu, S. Zhou, H. Liu, and M. S. Greco, “Optimal resource [27] S. Boyd, S. P. Boyd, and L. Vandenberghe, Convex Optimization. Cam-
allocation for asynchronous multiple targets tracking in heterogeneous bridge, U. K.: Cambridge Univ. Press, 2004.
radar networks,” IEEE Trans. Signal Process., vol. 68, pp. 4055–4068, [28] D. Kinderlehrer and G. Stampacchia, An Introduction to Variational In-
Jul. 2020. equalities and Their Applications. Philadelphia, USA: SIAM, 2000.
[5] H. Lakshmanan and D. P. De Farias, “Decentralized resource allocation in [29] A. Auslender, Optimisation: Méthodes Numériques. Paris, France: Mas-
dynamic networks of agents,” SIAM J. Optim., vol. 19, no. 2, pp. 911–940, son, 1976.
2008. [30] J. P. Aubin and A. Cellina, Differential Inclusions. Berlin, Germany:
[6] L. Xiao and S. Boyd, “Optimal scaling of a gradient method for distributed Springer-Verlag, 1984.
resource allocation,” J. Optim. Theory Appl., vol. 129, no. 3, pp. 469–488, [31] S. Yang, Q. Liu, and J. Wang, “A multi-agent system with a proportional-
Jun. 2006. integral protocol for distributed constrained optimization,” IEEE Trans.
[7] Z. Deng, S. Liang, and Y. Hong, “Distributed continuous-time algorithms Autom. Control, vol. 62, no. 7, pp. 3461–3467, Jul. 2017.
for resource allocation problems over weight-balanced digraphs,” IEEE [32] C. Godsil and G. F. Royle, Algebraic Graph Theory. London, U.K.:
Trans. Cybern., vol. 48, no. 11, pp. 3116–3125, Nov. 2018. Springer, 2001.
[8] C. Li, X. Yu, T. Huang, and X. He, “Distributed optimal consensus [33] A. P. Ruszczyński, Nonlinear Optimization. Princeton, NJ, USA: Princeton
over resource allocation network and its application to dynamical eco- Univ. Press, 2006.
nomic dispatch,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 6, [34] G. Strang, “The fundamental theorem of linear algebra,” Amer. Math.
pp. 2407–2418, Jun. 2018. Monthly, vol. 100, no. 9, pp. 848–855, 1993.
[9] X.-F. Wang, Y. Hong, X.-M. Sun, and K.-Z. Liu, “Distributed optimization [35] J. Ahmad, M. Tahir, and S. K. Mazumder, “Dynamic economic dispatch
for resource allocation problems under large delays,” IEEE Trans. Ind. and transient control of distributed generators in a microgrid,” IEEE Syst.
Electron., vol. 66, no. 12, pp. 9448–9457, Dec. 2019. J., vol. 13, no. 1, pp. 802–812, Mar. 2019.
[10] S. Liang, X. Zeng, G. Chen, and Y. Hong, “Distributed sub-optimal [36] M. S. Sarafraz and M. S. Tavazoei, “Uncertain multiagent systems with
resource allocation via a projected form of singular perturbation,” Au- distributed constrained optimization missions and event-triggered commu-
tomatica, vol. 121, 2020, Art. no. 109180. nications: Application to resource allocation,” IEEE Syst. J., early access,
[11] E. Ghadimi, I. Shames, and M. Johansson, “Multi-step gradient methods 2022, doi: 10.1109/JSYST.2022.3181937.
for networked optimization,” IEEE Trans. Signal Process., vol. 61, no. 21, [37] Z. Li and G. Chen, “Distributed adaptive control scheme for islanded AC
pp. 5417–5429, Nov. 2013. microgrids with tolerance to uncertain communication links,” IEEE Syst.
[12] T. Anderson, C.-Y. Chang, and S. Martnez, “Distributed approximate J., vol. 16, no. 2, pp. 2741–2752, Jun. 2022.
Newton algorithms and weight design for constrained optimization,” Au- [38] Y. Xu, Z. Dong, Z. Li, Y. Liu, and Z. Ding, “Distributed optimiza-
tomatica, vol. 109, 2019, Art. no. 108538. tion for integrated frequency regulation and economic dispatch in mi-
[13] M. Zargham, A. Ribeiro, A. Ozdaglar, and A. Jadbabaie, “Accelerated crogrids,” IEEE Trans. Smart Grid, vol. 12, no. 6, pp. 4595–4606,
dual descent for network flow optimization,” IEEE Trans. Autom. Control, Nov. 2021.
vol. 59, no. 4, pp. 905–920, Apr. 2014. [39] E. Gonçalves, A. R. Balbo, D. N. da Silva, L. Nepomuceno, E. C. Baptista,
[14] E. Wei, A. Ozdaglar, and A. Jadbabaie, “A distributed newton method and E. M. Soler, “Deterministic approach for solving multi-objective
for network utility maximization—I: Algorithm,” IEEE Trans. Autom. non-smooth environmental and economic dispatch problem,” Int. J. Elect.
Control, vol. 58, no. 9, pp. 2162–2175, Sep. 2013. Power Energy Syst., vol. 104, pp. 880–897, 2019.
[15] Z. Deng, “Distributed algorithm design for resource allocation prob- [40] F. Chen, X. Chen, L. Xiang, and W. Ren, “Distributed economic dispatch
lems of second-order multiagent systems over weight-balanced digraphs,” via a predictive scheme: Heterogeneous delays and privacy preservation,”
IEEE Trans. Syst., Man, Cybern. Syst., vol. 51, no. 6, pp. 3512–3521, Automatica, vol. 123, 2021, Art. no. 109356.
Jun. 2021. [41] X. Wang, S. Yang, Z. Guo, M. Lian, and T. Huang, “A distributed
[16] Z. Deng and T. Chen, “Distributed algorithm design for constrained dynamical system for optimal resource allocation over state-dependent
resource allocation problems with high-order multi-agent systems,” Au- networks,” IEEE Trans. Netw. Sci. Eng., vol. 9, no. 4, pp. 2940–2951,
tomatica, vol. 144, 2022, Art. no. 110492. Jul./Aug. 2022.
[17] J. B. Park, K. S. Lee, J. R. Shin, and K. Y. Lee, “A particle swarm [42] Y. Wan, C. Long, R. Deng, G. Wen, and X. Yu, “Adaptive event-triggered
optimization for economic dispatch with nonsmooth cost functions,” IEEE strategy for economic dispatch in uncertain communication networks,”
Trans. Power Syst., vol. 20, no. 1, pp. 34–42, Feb. 2005. IEEE Trans. Control Netw. Syst., vol. 8, no. 4, pp. 1881–1891, Dec. 2021.
[18] H.-T. Yang, P.-C. Yang, and C.-L. Huang, “Evolutionary programming [43] Y. Yan, Z. Chen, V. Varadharajan, M. J. Hossain, and G. E. Town, “Dis-
based economic dispatch for units with non-smooth fuel cost functions,” tributed consensus-based economic dispatch in power grids using the Pail-
IEEE Trans. Power Syst., vol. 11, no. 1, pp. 112–118, Feb. 1996. lier cryptosystem,” IEEE Trans. Smart Grid, vol. 12, no. 4, pp. 3493–3502,
[19] L. Tan, Z. Zhu, F. Ge, and N. Xiong, “Utility maximization resource Jul. 2021.
allocation in wireless networks: Methods and algorithms,” IEEE Trans. [44] S. Yang, S. Tan, and J.-X. Xu, “Consensus based approach for economic
Syst., Man, Cybern. Syst., vol. 45, no. 7, pp. 1018–1034, Jul. 2015. dispatch problem in a smart grid,” IEEE Trans. Power Syst., vol. 28, no. 4,
[20] P. Mahey and M. de Souza, “Local optimality conditions for multicom- pp. 4416–4426, Nov. 2013.
modity flow problems with separable piecewise convex costs,” Operations [45] B. Xu, F. Guo, W.-A. Zhang, W. Wang, C. Wen, and Z. Li, “Distributed
Res. Lett., vol. 35, no. 2, pp. 221–226, 2007. successive convex approximation for nonconvex economic dispatch in
[21] D. M. Yuan, D. W. C. Ho, and S. Y. Xu, “Regularized primal-dual smart grid,” IEEE Trans. Ind. Informat., vol. 17, no. 12, pp. 8288–8298,
subgradient method for distributed constrained optimization,” IEEE Trans. Dec. 2021.
Cybern., vol. 46, no. 9, pp. 2109–2118, Sep. 2016. [46] Z. Wang, D. Wang, C. Wen, F. Guo, and W. Wang, “Push-based distributed
[22] X. Zeng, J. Chen, S. Liang, and Y. Hong, “Generalized Nash equilibrium economic dispatch in smart grids over time-varying unbalanced directed
seeking strategy for distributed nonsmooth multi-cluster game,” Automat- graphs,” IEEE Trans. Smart Grid, vol. 12, no. 4, pp. 3185–3199, Jul. 2021.
ica, vol. 103, pp. 20–26, 2019. [47] S. Chen, L. Zhang, Z. Yan, and Z. Shen, “A distributed and robust
[23] Z. Deng, “Distributed generalized Nash equilibrium seeking algo- security-constrained economic dispatch algorithm based on blockchain,”
rithm for nonsmooth aggregative games,” Automatica, vol. 132, 2021, IEEE Trans. Power Syst., vol. 37, no. 1, pp. 691–700, Jan. 2022.
Art. no. 109794. [48] A. Cherukuri and J. Cortés, “Distributed generator coordination for ini-
[24] Y. Mo, W. Chen, S. Z. Khong, and L. Qiu, “A structure-tensor approach to tialization and anytime optimization in economic dispatch,” IEEE Trans.
integer matrix completion in indivisible resource allocation,” IEEE Trans. Control Netw. Syst., vol. 2, no. 3, pp. 226–237, Sep. 2015.
Autom. Control, vol. 67, no. 9, pp. 4541–4554, Sep. 2022. [49] H. Xing, Y. Mou, M. Fu, and Z. Lin, “Distributed bisection method for
[25] Z. Zhang and M. Y. Chow, “Convergence analysis of the incremental cost economic power dispatch in smart grid,” IEEE Trans. Power Syst., vol. 30,
consensus algorithm under different communication network topologies no. 6, pp. 3024–3035, Nov. 2015.
in a smart grid,” IEEE Trans. Power Syst., vol. 27, no. 4, pp. 1761–1768, [50] G. Binetti, A. Davoudi, F. L. Lewis, D. Naso, and B. Turchiano, “Dis-
Nov. 2012. tributed consensus-based economic dispatch with transmission losses,”
[26] Y. Jia, X. Lyu, P. Xie, Z. Xu, and M. Chen, “A novel retrospect- IEEE Trans. Power Syst., vol. 29, no. 4, pp. 1711–1720, Jul. 2014.
inspired regime for microgrid real-time energy scheduling with hetero- [51] D. K. Molzahn et al., “A survey of distributed optimization and control
geneous sources,” IEEE Trans. Smart Grid, vol. 11, no. 6, pp. 4614–4625, algorithms for electric power systems,” IEEE Trans. Smart Grid, vol. 8,
Nov. 2020. no. 6, pp. 2941–2962, Nov. 2017.

Authorized licensed use limited to: Indian Institute of Technology- Goa. Downloaded on January 19,2023 at 09:59:32 UTC from IEEE Xplore. Restrictions apply.