You are on page 1of 11

1318 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO.

2, FEBRUARY 2016

Minimum-Time Consensus-Based Approach


for Power System Applications
Tao Yang, Member, IEEE, Di Wu, Member, IEEE, Yannan Sun, Member, IEEE,
and Jianming Lian, Member, IEEE

Abstract—This paper presents minimum-time papers related to the applications considered in this paper are
consensus-based distributed algorithms for power sys- briefly reviewed herein.
tem applications, such as load shedding and economic In [10], the authors propose a novel two-step
dispatch. The proposed algorithms are capable of solving
these problems in a minimum number of time steps instead multi-agent-based load-shedding algorithm. In the first step,
of asymptotically as in most of the existing studies. consensus-based algorithm is developed to obtain the global
Moreover, these algorithms are applicable to both undi- information, including 1) total system overload; and 2) online
rected and directed communication networks. Simulation load index. In the second step, efficient load-shedding decisions
results are used to validate the proposed algorithms. are made based on the discovered global information. For EDP,
Index Terms—Distributed control, load shedding, several distributed algorithms have been proposed by choosing
minimum-time consensus, multi-agent systems, optimal the incremental cost as the consensus variable. For example,
dispatch. leader–follower algorithm is designed in [11], where the
leader is capable to collect the mismatch between demand and
I. I NTRODUCTION generation. To avoid the requirement of a leader, the authors of
[20] propose a two-level consensus-based algorithm, where the
P OWER system operation requires instantaneous balance
between generation and load. This balancing is achieved
through scheduling (e.g., unit commitment and economic dis-
upper level is the consensus and gradient algorithm, and the
lower level executes the classic consensus algorithm by choos-
patch) and real-time control (e.g., automatic generation and ing the local mismatch as the consensus variable. Kar and Hug
primary frequency control). Generation redispatch and load [12] proposed a consensus + innovation algorithm, where an
shedding may also be employed under contingencies. Currently, innovation term ensures that total generation meets the demand.
most of these scheduling and control are designed and deployed In [16], a fully distributed gradient algorithm is proposed and
in centralized schemes, where a single control center accesses implemented in Java Agent DEvelopment framework for
states and provides control signals for the entire system. This simulation studies. These algorithms are applicable to the
centralized control framework may be subjected to performance case where the communication network is undirected, i.e., the
limitations, such as high communication requirement and cost, information exchange is bidirectional. However, it is also desir-
substantial computational burden, limited flexibility and scala- able to develop control and coordination algorithms that only
bility, and disrespect of privacy. It is thus desirable to develop require directed information flow, because directed commu-
distributed approaches to overcome these limitations. nication networks cost less than their undirected counterparts
In recent years, there has been a rapid and sustained devel- [3], [21]. The capability of utilizing low-cost communication
opment in the area of distributed control, multi-agent coor- networks will be desirable in future smart grid. Hence, [13],
dination, and networked control [1]–[7]. Various distributed [14] propose distributed algorithms for more general cases,
algorithms have been developed for consensus, rendezvous, for- where the communication network can be either undirected or
mation, and optimization [8]. Thus, researchers start to apply directed. The distributed algorithm in [13] relies on the analysis
these distributed algorithms or their variations to power system in [22] and a ratio consensus algorithm. The algorithm in [14]
problems, such as load restoration [9], load shedding [10], eco- estimates the mismatch with all the agents being participated
nomic dispatch problem (EDP) [11]–[16], and plug-in electric and its performance over communication networks with time
vehicles (PEV) charging coordination [17]–[19]. Some of these delays is studied in [23]. A distributed algorithm based on
the bisection method and the consensus-like iteration method
Manuscript received February 24, 2015; revised August 28, 2015, is developed in [24] for solving EDP with general type of
October 19, 2015 and November 13, 2015; accepted November 16,
2015. Date of publication November 26, 2015; date of current version cost functions. One common property of these distributed
January 8, 2016. This work was supported by the Laboratory Directed algorithms is that the convergence is achieved asymptotically.
Research and Development (LDRD) program at the Pacific Northwest It could require an arbitrarily long time to obtain results within
National Laboratory.
The authors are with the Pacific Northwest National Laboratory, a reasonable accuracy. Such a performance may not satisfy
Richland, WA 99354 USA (e-mail: Tao.Yang@pnnl.gov; Di.Wu@pnnl. the need in some practical applications with time constraints.
gov; Yannan.Sun@pnnl.gov; Jianming.Lian@pnnl.gov). For example, for the load-shedding application, information
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. discovery process needs to be finished as fast as possible to
Digital Object Identifier 10.1109/TIE.2015.2504050 generate and deploy load-shedding plan in time [10].

0278-0046 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: MINIMUM-TIME CONSENSUS-BASED APPROACH FOR POWER SYSTEM APPLICATIONS 1319

Motivated by the finite-time computation for consensus for all i ∈ V, then the directed graph is balanced. For undi-
[25]–[27], this paper presents minimum-time consensus-based rected graphs, d+ −
i = di always holds. Hence, the superscripts
distributed algorithms for load shedding and EDP. The main are dropped, and di is used to denote the degree of node i.
contributions of this work are summarized as follows. In this paper, variables in boldface represent vectors or
1) For the load-shedding application, developing a dis- matrices. A nonnegative matrix A ∈ Rn×n is a column
tributed algorithm to discover information for directed (row) stochastic matrix if the entries in each column (row) sum
communication networks is a challenging problem, and to 1. A nonnegative matrix A ∈ Rn×n is primitive if Am has
therefore, is identified as an important future work in [10]. only positive elements for some positive integer m. For a matrix
This paper proposes an algorithm to address distributed A ∈ Rn×n , its kernel is defined as ker(A) = {v ∈ Rn |Av =
load-shedding over-directed communication networks. 0}. 0 and 1 denote the column vectors with all entries being 0
2) Minimum-time consensus has been proposed for the clas- and 1, respectively. diag(a1 , . . . , aN ) denotes a diagonal matrix
sical consensus dynamics. However, in addition to the with diagonal entries a1 , . . . , aN . eTi ∈ R1×n denotes the row
consensus dynamics, estimation of local mismatch also vector whose single one entry is at the ith position and all other
appears in the distributed control protocol for the EDP. entries are equal to zero. In denotes the identity matrix of size
Thus, it is unclear whether minimum-time consensus can n × n.
be achieved in this case. In this work, we have shown that
the minimum-time consensus can be extended to address
III. D ISTRIBUTED M INIMUM - TIME A LGORITHM FOR L OAD
the more complicated dynamics.
S HEDDING
3) A key feature that differentiates the proposed algorithms
from existing ones for distributed load shedding and EDP When generation is insufficient to support all loads under
(e.g., [9] and [14]) is that the proposed algorithms are able contingencies, efficient load shedding needs to be deployed to
to solve these problems in a minimum number of time maintain the supply–demand balance. Xu et al. [10] proposed
steps instead of asymptotically, and therefore, accelerate a novel multi-agent-based load-shedding algorithm, which can
the convergence speed and alleviate the communica- make efficient load-shedding decision based on discovered
tion burden. The techniques developed in this paper can global information in undirected communication networks. The
also be applied to other power system applications, such main challenge and essential part of this method is the global
as load restoration [9] and PEV charging coordination information discovery through information exchange in a dis-
[18], [19]. tributed manner. Proper load-shedding decisions are then made
This paper is organized as follows. In Section II, nota- at each agent based on load-priority tables together with the
tions and some preliminaries on graph theory are introduced. discovered total system overload and online load index. The
Sections III and IV present minimum-time consensus-based information discovery algorithms are based on average consen-
distributed algorithms for load shedding and EDP, respectively. sus, which converges asymptotically. A particle swarm opti-
In particular, both undirected and direct communication net- mization approach is also proposed for designing coefficients of
works are considered in load-shedding problem. Section V information exchange and therefore improving the convergence
presents case studies and compares the performance of the speed. In order to further enhance the performance of multi-
proposed algorithms with the ones in existing studies. Finally, agent-based load shedding, in this section, two algorithms are
concluding remarks are offered in Section VI. proposed to realize information discovery in a minimum num-
ber of time steps for undirected and directed communication
networks, respectively. The proposed methods are motivated by
II. P RELIMINARIES AND N OTATIONS the finite-time computation for consensus [25]–[27] and rely on
We first recall some basic concepts in graph theory [28]. the analysis of the rank of a square Hankel matrix constructed
Let G = (V, E) denote a directed graph (digraph) with the set from local states. These local states are obtained by execut-
of nodes (agents) V = {1, . . . , N } and the set of edges E ⊆ ing the classical consensus dynamics, which is asymptotically
V × V. A directed edge from node i to node j is denoted by stable on the consensus manifold.
(i, j) ∈ E. A graph is undirected if (i, j) ∈ E implies (j, i) ∈ The algorithms presented in this section are for average sys-
E. All nodes that can transmit information to node i directly tem overload discovery, but they can also be readily adopted
are said to be in-neighbors of node i and belong to the set for online load discovery by modeling the load index vector as
Ni+ = {j ∈ V | (j, i) ∈ E}. The cardinality of Ni+ is called the a consensus variable in the same way as [10].
in-degree of node i and is denoted by d+ +
i = |Ni |. The nodes
which receive information from agent i belong to the set of out-
A. Undirected Communication Networks
neighbors of node i, denoted by Ni− = {j ∈ V | (i, j) ∈ E}.
The cardinality of Ni− is called the out-degree of node i and The proposed algorithm for undirected communication net-
is denoted by d− −
i = |Ni |. It is assumed that node i belongs to
works relies on the consensus protocol in the following equa-
both its in-neighbor set and out-neighbor set, i.e., i ∈ Ni+ and tion:
i ∈ Ni− . A directed path from node i1 to ik is a sequence of 
n
nodes {i1 , . . . , ik }, such that (ij , ij+1 ) ∈ E for j = 1, . . . , k − wi (k + 1) = wi (k)+ aij (wj (k)−wi (k)), i ∈ {1, . . . , n}
1. A digraph is said to be strongly connected if there exists a j=1
directed path between any pair of distinct nodes. If d+ i = di

(1)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
1320 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 2, FEBRUARY 2016

where wi (k) is the average system overload estimated by bus Algorithm 1. Distributed minimum-time algorithm for infor-
i at time step (iteration) k 1 , n is the number of agents partic- mation discovery (undirected communication networks)
ipating in information discovery, and aij is the edge weight
(coefficient of the information exchange between buses i and j), 1: Data: Edge weight aij that satisfies (2), PGi and PLi for all
with appropriate initial condition wi (0) = PGi − PLi , where i ∈ {1, . . . , n}.
PGi and PLi are the local generation and load of bus i, respec- 2: Result: The average system overload.
tively. The coefficient aij = 0 if i = j or buses i and j are not 3: for k = 0, 1, . . . do
directly connected; otherwise, aij ∈ (0, 1). The nonzero edge 4: for i = 1 to n do
weights aij must satisfy (2) to guarantee convergence [10] 5: Run iteration (1) with initial conditions w(0) =
[w1 (0), . . . , wn (0)]T where wi (0) = PGi − PLi , and
2 store the successive values of wi (k).
0 < aij < (2)
di + dj − 2 6: if k is odd then
7: Compute the vector of differences wi by (3).
where di is agent i’s degree, i.e., the number of agents con-
8: Construct the square Hankel matrix Hji by (4).
nected to node i including itself.
If each agent has the capability to store local states for a 9: if Hji is singular then
few time steps, a Hankel matrix can be constructed. The aver- 10: Compute the normalized kernel β (i) .
age system overload can then be calculated in a minimum 11: Compute the average system overload μ by (5).
number of steps based on the local states and normalized ker- 12: Return μ.
nel of the Hankel matrix, as explained below. At time step 13: end if
k = 2j − 1 (thus k is odd), where j = 1, 2, . . ., agent i stores 14: end if
its successive states wi (0), wi (1), . . . , wi (2j − 1). The differ- 15: end for
ence between these successive values is then calculated in (3) 16: end for
and stored
in [26, Th. 3]. It shows that the required number of time steps
wi (k  )  wi (k  ) − wi (k  − 1), k  = 1, . . . , k. (3)
for each agent is bounded in terms of two combinatorial graph
A square Hankel matrix of dimension j × j is constructed for theoretical properties: the minimum external equitable partition
each agent, as shown below [29] of the graph with respect to that agent and the longest
⎡ ⎤ distance for that agent.
wi (1) wi (2) . . . wi (j) Remark 1: Compared to the asymptotic consensus-based
⎢wi (2) wi (3) . . . wi (j + 1) ⎥
⎢ ⎥ algorithm in [10], the minimum-time consensus-based
Hji  ⎢ . .. . .. ⎥. (4)
⎣ . . . . . . ⎦ Algorithm 1 also executes the same iteration (1). However, the
wi (j) wi (j + 1) . . . wi (2j − 1) minimum-time consensus technique allows agent i to discover
the average system overload using local states within 2Mi time
The rank of Hankel matrix is examined at each time step. steps, even before consensus is achieved within a reasonable
When it becomes singular, its normalized kernel β (i) = accuracy. Note that the total solving time is determined by
(i) (i)
[β1 , . . . , βMi −1 , 1]T is computed, where Mi is the dimension 1) number of iterations; and 2) the required time for each
of the Hankel matrix when it starts to become singular. The iteration. Therefore, users should select appropriate algorithms
average system overload can be calculated as based on 1) their preference with respect to communication cost
and convergence time; and 2) system configurations such as
wMT
i
β (i) communication network settings and capacity, computational
μ = lim wi (k) = (5)
k→∞ 1T β (i) capability at each agent.
Remark 2: Algorithm 1 relies on the analysis of the rank of a
where wMi = [wi (0), . . . , wi (Mi − 1)]T . Following the simi- square Hankel matrix. It is shown in [26] that Hankel matrix is
lar analysis as [26, Th. 1], it can be shown that Mi is equal to guaranteed to lose rank when j = Mi . However, this property
the degree of a minimal polynomial of the matrix pair (Ci , A) cannot be ensured if errors exist in the signals received from
with Ci  eTi ∈ R1×n and A  In − L, where the Laplacian neighboring agents. In such a case, singular-value decomposi-
n
matrix L = [ij ] is defined as ii = j=1 aij and ij = −aij tion method is proposed to find the nearest defective Hankel
for j = i. Therefore, it is impossible to compute the average matrix [30]. Nevertheless, in this paper, we focus on the case
system overload by local states within time steps less than 2Mi . where the received state values are exact and leave the extension
The above procedure to discover the average system over- to the case with noises and disturbances as future work.
load in a minimum number of time steps is summarized in Once the average system overload μ is discovered, the total
Algorithm 1. Since 2Mi is the minimum time steps required to system overload can be calculated at each agent according to
discover the consensus value, Algorithm 1 is called minimum-
time consensus-based algorithm. The graph-theoretical charac- 
n

terization of the required number of time steps is established Pnet = (PGi − PLi ) = nμ (6)
i=1
1 The physical meaning of w (k) is the local estimation at time step k of the
i
average system overload when the system disturbance occurs, rather than the provided that n is known. The value of n and online load index
sample of local system imbalance at time step k. can be obtained by running an algorithm similar to Algorithm 1

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: MINIMUM-TIME CONSENSUS-BASED APPROACH FOR POWER SYSTEM APPLICATIONS 1321

with the load index vector being modeled as the consensus Algorithm 2. Distributed minimum-time algorithm for infor-
variable. Reference [10] provides more details on how to con- mation discovery (directed communication networks)
struct load index vectors that can be used by consensus-based
algorithms. 1: Data: Edge weight qi,j in the ratio consensus algorithm
(7) such that the matrix Q = [qi,j ] ∈ Rn×n is a primi-
tive column stochastic matrix, PGi and PLi for all i ∈
B. Directed Communication Networks {1, . . . , n}.
2: Result: The average system overload.
The algorithm proposed in the previous section is for undi-
3: for k = 0, 1, . . . do
rected communication networks, where every communication
4: for i = 1 to n do
link enables bidirectional information exchange. However, this
5: Run two iterations (7a) and (7b) simultaneously, with
may be restrictive since the information exchange may not
initial conditions given by (8), where ηi (0) = PGi −
be symmetric in practical situation. Therefore, directed com-
PLi , and store the successive values of ηi (k) and ζi (k).
munication networks are needed to model the unidirectional
6: if k is odd then
information exchange. In this section, a distributed algorithm (k )
7: Compute the ratio zi (k  ) = ηζii(k 
 ) for k = 1, . . . , k.
is proposed to discover information in a minimum number of
8: Compute the vectors of differences z i by (3).
time steps for directed communication networks.
9: Construct the Hankel matrix Hji by (4).
In general, average consensus cannot be reached for directed
graphs unless they are strongly connected and balanced [31, Th. 10: if Hji is singular then
2] and [32, Th. 1]. In [33], the authors propose a ratio consensus 11: Compute the normalized kernel β (i) .
algorithm to achieve average consensus for a directed graph that 12: Compute the average system overload μ by (10)
is strongly connected but not necessarily balanced. In this algo-
ηi (k) η T i β (i)
rithm, each agent i in a directed graph G = (V, E) maintains μ = lim = M , (10)
k→∞ ζi (k) T β (i)
ζM
two values ηi and ζi , which are referred as internal states. These i

internal states are updated according to the following equations:


 where ηMi =[ηi (0), . . . , ηi (Mi − 1)]T and ζMi
ηi (k + 1) = qi,j ηj (k) (7a) = [ζi (0), . . . , ζi (Mi − 1)]T .
j∈Ni+
13: Return μ.
 14: end if
ζi (k + 1) = qi,j ζj (k) (7b) 15: end if
j∈Ni+ 16: end for
17: end for
where qi,j is chosen so that matrix Q = [qi,j ] ∈ R|V|×|V| is
column-stochastic and primitive (a specific choice is qi,j =
1/d− −
j for i ∈ Nj and zeros otherwise), and the initial condi- overload in a minimum number of time steps. In contrast to
tions are Algorithm 1, Algorithm 2 is applicable to more general cases,
where information is exchanged unidirectionally.
η(0) = [η1 (0); . . . ; η|V| (0)], ζ(0) = 1 . (8) In Algorithm 2, at each step k, agent i runs iterations (7a)
and (7b) simultaneously, and thus 2d+ i multiplicities are per-
Since Q is column-stochastic and primitive, according to
formed, where d+ i is the in-degree of agent i. As discussed in
Perron–Frobenius theorem [34, pp. 516], limk→∞ Qk = v1T ,
Section III-A, it requires 2Mi steps for agent i to obtain the con-
where v > 0 and 1T v = 1. It can be shown that agent i ∈ V
sensus value. Thus, the computational complexity is 4Mi d+ i .
asymptotically obtains the average by taking the ratio of ηi and
For checking the rank of a square Hankel matrices, the spe-
ζi as
cial structure of Hankel matrices allows for efficient methods to
ηi (k) compute their rank. For example, in [35], a modular method is
lim zi (k) = lim proposed with complexity of O(m2 ), where m is the size of the
k→∞ k→∞ ζi (k)
|V| square matrix.
vi j=1 ηj (0) j∈V ηj (0) Similar to the previous discussion for undirected commu-
= = (9)
|V|
vi j=1 1 |V| nication networks, with the average system overload μ being
discovered, the total system overload can be obtained according
where vi is the ith element of v. to (6) if n is known. The number n and online load index can be
Motivated by this ratio consensus algorithm and the finite- obtained by running an algorithm similar to Algorithm 2 with
time computation for consensus, Algorithm 2 is proposed for the load index vector being modeled as a consensus variable.
the average system overload discovery in a directed com- Remark 3: Note that in Algorithms 1 and 2, after agent
munication network. In this algorithm, each agent constructs i obtains the exact value of the average system overload, it
the square Hankel matrix according to (4) with wi replaced still runs the iteration and thus shares its estimation of the
by z i , where z i are successive values for the ratio of inter- average system overload with its out-neighbors, because other
nal states zi rather than consensus variables. The proposed agents may require more steps to compute the average sys-
algorithm enables each agent to discover the average system tem overload. Nevertheless, all agents stop the iteration and

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
1322 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 2, FEBRUARY 2016

communication when all the agents have obtained the exact [37]. All these methods require a single control center to col-
value, which is upper-bounded by 2n. lect all generators’ cost information, run optimal dispatch, and
Remark 4: Note that Algorithms 1 and 2 can be extended broadcast the optimal output level to each generator. Recently,
to other power system applications, such as load restoration [9] several consensus-based algorithms have been proposed to
and PEV charging coordination [18], [19], by choosing the con- solve EDP in a distributed manner, e.g., [11], [12], [20], and
sensus variables and the initial conditions in the same way as in [14]. Among these algorithms, [14] is applicable to directed
these references. communication networks and fully distributed. For the case
where the directed communication network is strongly con-
nected, the authors of [14] propose a fully distributed algorithm
IV. D ISTRIBUTED M INIMUM - TIME as shown below
A LGORITHM FOR EDP 
Economic dispatch is one of fundamental problems in power λi (k + 1) = pi,j λj (k) + yi (k) (14a)
system, which is used for generation scheduling. It is essentially j∈Ni+

an optimization problem where the objective is to minimize the xi (k + 1) = φi (λi (k + 1)) (14b)
total generation cost while meeting total demand and satisfying 
yi (k + 1) = qi,j yj (k) − (xi (k + 1) − xi (k)) (14c)
individual generator output limits, as described below
j∈Ni+

N
min Ci (xi ) (11a) where the learning gain is a sufficiently small positive con-
xi
i=1 stant, λi (k) is the estimation of optimal incremental cost by
subject to xmin
i ≤ xi ≤ xmax
i (11b) generator i at time step k, xi (k) is the power generation of
generator i at time step k, which is an estimation of optimal

N
xi = D (11c) power generation of generator i, yi (k) is the local estimation
i=1
of the mismatch between demand and total power genera-
tion by generator i at time step k, and the edge weights are
where N is the number of generators in the power system, chosen as
xi is the power generation of generator i, Ci (xi ) is the cost 
function of generator i, xmin and xmax are the lower and
1
+, if j ∈ Ni+
i i pi,j = d i ∀i, j ∈ {1, . . . , N } (15)
upper bounds of the power generation of generator i, respec- 0, otherwise
tively,2 and D is the total demand satisfying i=1 xmin
N
i ≤D≤  1
N max −, if i ∈ Nj−
i=1 xi . Note that there are other power system short-term qi,j = dj ∀i, j ∈ {1, . . . , N } (16)
scheduling problems which consider more practical constraints. 0, otherwise
For instance, power flow and transmission thermal limits are
taken into account in optimal power flow problem, while and the projection operators are defined as
system reserve requirement, generator minimum up/down ⎧
max
cost, and time constraints are modeled in unit commitment ⎪
⎨xi , if λi > λmax
i
problem. φi (λi ) = βi λi + αi , if λmin ≤ λi ≤ λmax (17)

⎩ min
i i
Similar to [11], [12], and [14], we assume that cost functions xi , if λi < λmin
i
are in a quadratic form
where λmini = (xmin
i − αi )/βi and λmax
i = (xmax
i − αi )/βi .
(xi − αi )2 The incremental cost λi (k) in (14a) is the consensus vari-
Ci (xi ) = + γi (12)
2βi able, and all λi (k) converge to a common value. By using
2
the fact that Q = [qi,j ] ∈ RN ×N is a column-stochastic matrix,
where αi ≤ 0, βi > 0, γi ≤ 0, and 2β α
i
+ γi > 0. The incre- N
the authors show that i=1 (yi (k) + xi (k)) is invariant for all
mental cost for the generator i is then given by C· (xi )/x· = k. Furthermore, by appropriately choosing the initial condi-
i i
N
(xi − αi )/βi . According to the equal incremental cost criterion tions xi (0) and yi (0) so that i=1 (xi (0) + yi (0)) = D, all the
N
[36], EDP has a solution as agents collectively learn the mismatch as i=1 yi (k) = D −
⎧ x −α N
∗ min max i=1 xi (k), and thus it is not necessary to have some agents
⎪ i i
⎨ βi = λ , for xi < xi < xi access the entire system information. The local estimated mis-
xi −αi
< λ∗ , for xi = xmax (13) match yi (k) is used as a feedback in (14a) to ensure that all

⎩ xi −α
βi i

βi
i ∗
> λ , for xi = xi min λi (k) converge to the optimal incremental cost λ∗ . Power gen-
eration xi (k) is updated according to the iteration (14b), which
where λ∗ is the optimal incremental cost. converges to the optimal power generation x∗i .
The conventional centralized approaches for EDP include the By adapting the finite-time computation for consensus to
lambda-iteration method and the gradient search method [36], the algorithm in (14), Algorithm 3 is proposed to solve EDP
2 Note that in addition to static operation limits, [xmin , xmax ] could also rep- in a minimum number of time steps. In this algorithm, each
i i
resent feasible power output of generator i in current time period, considering agent i iteratively runs (14), stores the successive values of gen-
the operating level in previous time period and ramping rates. erator incremental cost λi , computes the differences between

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: MINIMUM-TIME CONSENSUS-BASED APPROACH FOR POWER SYSTEM APPLICATIONS 1323

Algorithm 3. Distributed minimum-time algorithm for EDP where λ = [λ1 , . . . , λN ]T , y = [y1 , . . . , yN ]T , P = [Pi,j ] ∈
RN ×N , and B̃ = diag(β̃1 , . . . , β̃N ), with
1: Data: Edge weights pi,j and qi,j given by (15) and (16).

2: Result: Optimal incremental cost λ∗ and optimal power 0, if xi (k) is saturated
generation x∗i . β̃i =
3: for k = 0, 1, . . . do βi , otherwise.
4: for i = 1 to N do
As shown by the eigenvalue perturbation analysis in [14, Th. 3],
5: Run iteration (14), and store the successive values of
when is sufficiently small, the matrix
λi (k).
6: if k is odd then  
P I
7: Compute the vector of differences λi by (3). A
B̃(I − P) Q − I
8: Construct the Hankel matrix Hji by (4).
9: if Hji is singular then has only one eigenvalue at 1 while all other eigenvalues are
10: Compute the normalized kernel β (i) . strictly within the unit circle. Moreover,
11: Compute λ∗ and x∗i according to    
λ(k) 1
converges to span
λTMi β (i) y(k) 0
λ∗ = lim λi (k) = , (18a)
k→∞ 1T β (i)  
x∗i = φi (λ∗ ), (18b) as k → ∞. Next define Ci  eTi 0T . Thus, the same
minimum-time consensus technique which was developed for
where λTMi = [λi (0), . . . , λi (Mi − 1)]. weighted average consensus based on the minimal polynomial
12: Return λ∗i and x∗i . of a matrix pair of the conventional consensus in [26] also
13: end if works with the matrix pair (Ci , A) in our case.
14: end if Remark 7: Note that distributed algorithms have been pro-
15: end for posed to determine generator operating point with transmission
16: end for losses recently, e.g., [15], [38]–[40]. We leave the extension of
our study to handle these constraints as future work.
Remark 8: The minimum-time consensus technique used in
this paper was developed for the classical single-integrator con-
these successive values λi according to (3) with wi replaced by sensus dynamics. This technique is essentially an identification
λi , and constructs the square Hankel matrix according to (4) of linear time-invariant systems, which requires a sufficient
with wi replaced by λi . The rank of the Hankel matrix is then number of local states to construct the system. The identifica-
checked by each agent at each odd-number time step. When tion for the nonlinear case based on fuzzy dynamic models [41],
it becomes singular, its normalized kernel β (i) is calculated. [42] is more challenging and interesting, and thus left as future
Finally, each agent computes the optimal incremental cost by work.
(18a) and the optimal power generation by (18b). The proposed Remark 9: Recent works [43], [44] proposed a novel two-
distributed algorithm enables each agent to compute the incre- layer structure to solve the setpoints compensation problem
mental cost and optimal generation in a minimum number of industrial processes. In these works, the optimization problem
time steps. The technique developed here can also be applied to at larger operation involves system dynamics constraints, which
other asymptotic consensus-based distributed algorithms, e.g., is different from the optimization problem considered in this
[11]–[13]. paper where system dynamics are not involved. We leave the
Remark 5: Note that Algorithm 3 is a distributed method to extension of the minimum-time computation concept to the
solve an operational scheduling problem (e.g., 5-min economic optimization problem with system dynamics as future work.
dispatch), rather than to control generators’ instantaneous out-
put in real time. However, it solves the problem in a minimum
number of time steps rather than asymptotically as in existing V. C ASE S TUDIES
studies [11]–[13]. The iteration process in (14) only happens
This section presents case studies and results using the pro-
within each virtual agent, rather than being physically real-
posed algorithms. The test systems and parameters are adopted
ized in each generator. Once the optimal solution is obtained,
from [10] and [14] for the purpose of comparison. Numerical
each generator will be dispatched the same as existing practice
experiments are used to illustrate and validate the proposed
except the generator output level is determined in a distributed
algorithms. Simulation results are compared to those in the
manner rather than determined and broadcasted at a control
previous studies to demonstrate effectiveness and speed of the
center.
proposed algorithms.
Remark 6: Note that the iteration (14) in Algorithm 3 can be
written in the following composite form:

     A. Load Shedding
λ(k + 1) P I λ(k) For the load-shedding application, the proposed distributed
=
y(k + 1) B̃(I − P) Q − I y(k) algorithms have been tested for both undirected and directed

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
1324 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 2, FEBRUARY 2016

TABLE I
DATA OF THE IEEE 5-B US S YSTEM

Fig. 1. IEEE 5-bus power system.

Fig. 2. IEEE 14-bus power system. Fig. 3. IEEE 5-bus with bidirectional communication: average sys-
tem overload discovery process using the algorithm in [10] versus
Algorithm 2 (stopping time is indicated by the vertical blue line).
communication networks. In both cases, the proposed algo-
rithms are used to discover the average system overload. Online TABLE II
load index is discovered similarly by modeling the load index DATA OF THE IEEE 14-B US S YSTEM
vector as the consensus variable in the same way as [10]. The
detailed procedure is omitted due to space limitation.
1) Undirected Communication Networks: Both IEEE 5-
bus and 14-bus systems adopted from [10] are used to test
Algorithm 1, which are shown in Figs. 1 and 2, respectively.
The communication network is assumed to be the same as the
topology of the physical system, and thus it is undirected.
a) IEEE 5-bus system: The system information is sum-
marized in Table I. For the purpose of comparison, the same
edge weights from [10] are used as here: a12 = 13 , a13 = 13 , which is the average system overload. Similarly, agents 2–5
a23 = 13 , a24 = 1/2, and a35 = 12 , which are designed using discover the same average system overload at k = 9. Fig. 3
a particle swarm optimization method to speed up the conver- visualizes the discovery processes by using the algorithm pro-
gence. For the average system overload discovery, the initial posed in [10] and Algorithm 1. As can be seen, with the
conditions are listed in the fifth column of Table I. Following algorithm proposed in [10], all agents converge to −10 asymp-
Algorithm 1, each agent executes the average consensus pro- totically. However, the errors are still big at time step 9, and
tocol, stores local states for a few time steps, constructs the gradually decrease to less than 0.1 when k = 20. In contrast,
square Hankel matrix, and examines the rank of the Hankel with Algorithm 1, all the agents discover the exact average
matrix. It should be noted that the Hankel matrices for dif- system overload in 9 time steps, with about 11 time steps
ferent agents may become singular at different time steps. For saved. The IEEE-14 bus system shown in Fig. 2 is used to test
example, the Hankel matrix at agents 1 and 2 starts to become Algorithm 1. The system information is summarized in Table II.
singular at time step k = 5 and k = 9, respectively. In partic- For the purpose of comparison, the optimized edge weights
ular, the Hankel matrix and its normalized kernel at k = 5 for from [10] are used and shown in Table III. The comparison of
agent 1 are given in (19) and (20), respectively, the discovery processes using the algorithm proposed in [10]
⎡ ⎤
−50.000 83.333 −8.333 and Algorithm 1 is shown in Fig. 4. As can be seen, with the
H31 = ⎣ 83.333 −8.333 13.889 ⎦ (19) algorithm proposed in [10], it takes 50 time steps for all agents
−8.333 13.889 −1.389 to converge to −10 with an accuracy of 0.3. On the other hand,
using the proposed algorithm in this paper, all the agents dis-
β (1) = [−0.167, 0, 1]T . (20) cover the exact average system overload in 27 time steps, with
Agent 1 then performs calculation according to (5) about 23 time steps saved.
2) Directed Communication Networks: In this case, the
[−50.000, −100.000, −16.667][−0.167, 0, 1]T communication network is assumed to be different from the
μ= = −10
[1, 1, 1][−0.167, 0, 1]T physical network topology and is modeled as a directed

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: MINIMUM-TIME CONSENSUS-BASED APPROACH FOR POWER SYSTEM APPLICATIONS 1325

TABLE III
E DGE W EIGHTS OF THE IEEE 14-B US S YSTEM

Fig. 5. IEEE 14-bus with unidirectional communication: average system


overload discovery process using the ratio consensus algorithm versus
Algorithm 2 (stopping time is indicated by the vertical blue line).

TABLE IV
G ENERATOR PARAMETERS

Fig. 4. IEEE 14-bus with bidirectional communication: average sys-


tem overload discovery process using the algorithm in [10] versus
Algorithm 1 (stopping time is indicated by the vertical blue line).

graph G = (V, E), with the edge set E = {(i, i + 1), (i, i +
2)|1 ≤ i ≤ 12} ∪ {(13, 14), (13, 1), (14, 1)}. In this example,
the directed network is strongly connected, but not balanced
since for node 2, the in-degree d+ 2 = 2 which is not equal to
its out-degree d− 2 = 3. The edge weights are chosen as qi,j =
1/d− j for all i ∈ N −
j (zeros otherwise) for the linear itera-
tion (7). Fig. 5 plots the results using both the existing ratio
consensus algorithm and Algorithm 2 proposed in this paper.
Following Algorithm 2, all the agents discover the average
system overload μ = −10, but using different number of time Fig. 6. Communication topology and command agent for initialization.
steps: 55 steps (agents 4 and 12), 57 steps (agents 8, 9, and 13),
59 steps (agents 1, 10, 11 and 14), 61 steps (agents 2, 3, and 7), TABLE V
I NITIALIZATIONS
63 steps (agents 5), and 65 steps (agents 6). Nevertheless, all the
agents are able to discover the exact average system overload by
time step 65. On the other hand, using only the ratio consensus
algorithm (7), the estimated average system overload converges
−10 with an accuracy of 0.3 until time step 90. It is also inter-
esting to observe that the discovery process for this directed
network (cf., Fig. 5) is slower than the undirected network case The directed graph associated with the network is strongly
(cf., Fig. 4). connected, and agents 1 and 3 receive information from the
command agent 0. The initial conditions are listed in Table V.
The total demand is D = 1500 MW, and the learning gain is
B. EDP = 7.026 × 10−4 , which is also the value used in [14] designed
For economic dispatch, the proposed algorithm has been via a systematic method therein.
tested for two cases: without and with generator capacity con- 1) Without Generator Capacity Constraints: Following
straints. Three types of generators are considered: Type A Algorithm 3 proposed in this paper, each agent iteratively
(coal-fired steam unit), Type B (oil-fired steam unit), and Type executes (14), stores the successive values of generator incre-
C (oil-fired steam unit). The parameters of the cost functions mental cost, constructs the square Hankel matrix, and exam-
are specified in terms of α, β, and γ as given in [14, Table I] ines the rank of the Hankel matrix. Once the Hankel matrix
and listed in Table IV in this paper for readers’ convenience. becomes singular, each agent computes its normalized ker-
Consider a power network consisting of four generators, where nel. With this information being obtained, each agent finally
generators 1 and 2 are Type A generators, generator 3 is a Type computes the optimal incremental cost as (18a) and finds that
B generator, and generator 4 is a Type C generator. The com- λ∗ = 8.8405 $ /MWh. In particular, agents 1, 3 and 4 run 16
munication network is adopted from [14], as shown in Fig. 6. iterations, while agent 2 runs 18 iterations. Once the optimal

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
1326 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 2, FEBRUARY 2016

Fig. 7. Solving EDP for the unconstrained case using the algorithm in Fig. 8. Solving EDP for the constrained case using the algorithm in
[14] versus Algorithm 3 whose stopping time is indicated by the verti- [14] versus Algorithm 3 whose stopping time is indicated by the verti-
cal blue line. (a) Generator incremental cost ($/MWh). (b) Dynamics of cal blue line. (a) Generator incremental cost ($/MWh). (b) Dynamics of
incremental cost at agent 3. incremental cost at agent 1.

incremental cost λ∗ is computed, each agent then finds the all the agents are able to compute the optimal incremental cost
optimal generations x∗1 = 577.5279 MW, x∗2 = 577.5279 MW, at time step k = 27 as indicated by the thick vertical lines in
x∗3 = 254.9880 MW, and x∗4 = 89.9563 MW x∗i = βi λ∗ + αi . Fig. 8(a), while the solution errors using the algorithm in [14]
These results are the same as the ones obtained using the are about 0.002 $ /MWh at time step 27, resulting about 1.8
algorithm proposed in [14]. As can be seen from Fig. 7(a), MW mismatch between generation and load. Fig. 8(b) high-
with Algorithm 3, all agents find the optimal incremental lights the different dynamics of the two algorithms at agent 1.
cost λ∗ = 8.8405 $ /MWh at time step 17. In contrast, with Compared to the unconstrained case, the required number of
the algorithm proposed in [14], the solution errors are about time steps increases, but is still smaller than the number of time
0.02 $/MWh at time step 17 and the corresponding genera- steps needed when using the algorithm without the minimum-
tion errors are as large as 6 MW. The incremental cost and time computation technique. Moreover, the exact values are
generation output gradually converge to the optimal values obtained rather than their approximations.
with a reasonable accuracy roughly at time step 50. Since
Algorithm 3 executes the same iteration as the algorithm in [14]
to obtain the required local states for Hankel matrix construc- VI. C ONCLUSION
tion, the two algorithms result in the same dynamics before In this paper, we presented minimum-time consensus-based
Algorithm 3 computes the exact consensus value at the min- distributed algorithms for power system applications including
imum time step for each agent. As an example, Fig. 7(b) load shedding and economic dispatch. The proposed algorithms
highlights the different dynamics of the two algorithms at solved these problems in a minimum number of time steps
agent 3, where the vertical axis is zoomed into the range of rather than asymptotically, by using local states with a min-
[8.8, 8.85]. imum number of time steps. Simulation results showed that
2) With Generator Capacity Constraints: The genera- the solution can be obtained in a minimum time steps using
tor capacity limits are imposed in this case. It is found that the proposed algorithms. There are several interesting future
Algorithm 3 is still capable of obtaining the optimal incremen- directions: 1) extend the existing methodology by considering
tal cost λ∗ = 8.8405 $ /MWh. In particular, agents 1 and 2 use potential impacts of the practical communication network, such
24 time steps, while agents 3 and 4 use 28 time steps. Thus, as time delays and packet dropouts; 2) design (minimum-time)

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
YANG et al.: MINIMUM-TIME CONSENSUS-BASED APPROACH FOR POWER SYSTEM APPLICATIONS 1327

distributed algorithms to accommodate additional physical con- [21] E. Klavins, “Communication complexity of multi-robot systems,” in
straints as in optimal power flow and unit commitment prob- Algorithmic Foundations of Robotics V, J.-D. Boissonnat, J. Burdick,
K. Goldberg, and S. Hutchinson, Eds. New York, NY, USA: Springer,
lems; and 3) extend the concept of minimum-time computation 2004, vol. 7, pp. 275–291.
to nonlinear systems and wide practical network-control sys- [22] M. Madrigal and V. Quintana, “An analytical solution to the economic
tems. dispatch problem,” IEEE Power Eng. Rev., vol. 20, no. 9, pp. 52–55, Sep.
2000.
[23] T. Yang, D. Wu, Y. Sun, and J. Lian, “Impacts of time delays on dis-
tributed algorithms for economic dispatch,” in Proc. IEEE Power Energy
R EFERENCES Soc. Gen. Meeting, Jul. 2015, pp. 1–5.
[24] H. Xing, Y. Mou, M. Fu, and Z. Lin, “Distributed bisection method
[1] J. Tsitsiklis, “Problems in decentralized decision making and computa- for economic power dispatch in smart grid,” IEEE Trans. Power Syst.,
tion,” Ph.D. dissertation, Dept. Elect. Eng., Massachusetts Inst. Technol., vol. 30, no. 6, pp. 3024–3035, Nov. 2015.
Cambridge, MA, USA, 1984. [25] S. Sundaram and C. N. Hadjicostis, “Finite-time distributed consensus
[2] A. Jadbabaie, J. Lin, and A. Morse, “Coordination of groups of mobile in graphs with time-invariant topologies,” in Proc. IEEE Amer. Control
autonomous agents using nearest neighbor rules,” IEEE Trans. Autom. Conf., Jul. 2007, pp. 711–716.
Control, vol. 48, no. 6, pp. 988–1001, Jun. 2003. [26] Y. Yuan, G. Stan, L. Shi, M. Barahona, and J. Gonçalves, “Decentralised
[3] R. Olfati-Saber and R. Murray, “Consensus problems in networks of minimum-time consensus,” Automatica, vol. 49, no. 5, pp. 1227–1235,
agents with switching topology and time-delays,” IEEE Trans. Autom. May 2013.
Control, vol. 49, no. 9, pp. 1520–1533, Sep. 2004. [27] T. Charalambous, Y. Yuan, T. Yang, W. Pan, C. N. Hadjicostis, and
[4] L. Moreau, “Stability of multiagent systems with time-dependent com- M. Johansson, “Distributed finite-time average consensus in digraphs
munication links,” IEEE Trans. Autom. Control, vol. 50, no. 2, pp. 169– in the presence of time-delays,” IEEE Trans. Control Netw. Syst., to be
182, Feb. 2005. published.
[5] W. Ren and R. Beard, “Consensus seeking in multiagent systems [28] C. Godsi and G. Royle, Algebraic Graph Theory. Berlin, Germany:
under dynamically changing interaction topologies,” IEEE Trans. Autom. Springer-Verlag, 2001, vol. 207.
Control, vol. 50, no. 5, pp. 655–661, May 2005. [29] M. Egerstedt, S. Martini, M. Cao, K. Camlibel, and A. Bicchi,
[6] R. Gupta and M.-Y. Chow, “Networked control system: Overview and “Interacting with networks: How does structure relate to controllability
research trends,” IEEE Trans. Ind. Electron., vol. 57, no. 7, pp. 2527– in single-leader, consensus networks?” IEEE Control Syst. Mag., vol. 32,
2535, Jul. 2010. no. 4, pp. 66–73, Aug. 2012.
[7] J. Qin, C. Yu, and H. Gao, “Coordination for linear multiagent systems [30] Y. Yuan, J. Liu, R. Murray, and J. Gonçalves, “Decentralised minimal-
with dynamic interaction topology in the leader-following framework,” time dynamic consensus,” in Proc. IEEE Amer. Control Conf., Montréal,
IEEE Trans. Ind. Electron., vol. 61, no. 5, pp. 2412–2422, May 2014. QC, Canada, 2012, pp. 800–805.
[8] Y. Cao, W. Yu, W. Ren, and G. Chen, “An overview of recent progress [31] R. Olfati-Saber, J. Fax, and R. Murray, “Consensus and cooperation in
in the study of distributed multi-agent coordination,” IEEE Trans. Ind. networked multi-agent systems,” Proc. IEEE, vol. 95, no. 1, pp. 215–233,
Informat., vol. 9, no. 1, pp. 427–438, Feb. 2013. Jan. 2007.
[9] Y. Xu and W. Liu, “Novel multiagent based load restoration algorithm for [32] L. Xiao and S. Boyd, “Fast linear iterations for distributed averaging,”
microgrids,” IEEE Trans. Smart Grid, vol. 2, no. 1, pp. 152–161, Mar. Syst. Control Lett., vol. 53, no. 1, pp. 65–78, Sep. 2004.
2011. [33] A. Dominguez-Garcia and C. N. Hadjicostis, “Coordination and control
[10] Y. Xu, W. Liu, and J. Gong, “Stable multi-agent-based load shedding of distributed energy resources for provision of ancillary services,” in
algorithm for power systems,” IEEE Trans. Power Syst., vol. 26, no. 4, Proc. 1st IEEE Int. Conf. Smart Grid Commun. (SmartGridComm), Oct.
pp. 2006–2014, Nov. 2011. 2010, pp. 537–542.
[11] Z. Zhang and M.-Y. Chow, “Convergence analysis of the incremental cost [34] R. Horn and C. Johnson, Matrix Analysis. Cambridge, U.K.: Cambridge
consensus algorithm under different communication network topologies Univ. Press, 1985.
in a smart grid,” IEEE Trans. Power Syst., vol. 27, no. 4, pp. 1761–1768, [35] J. Sendra and J. Llovet, “Rank of a Hankel matrix over, Z[x1 , · · · , xr ]”
Nov. 2012. Appl. Algebra Eng. Commun. Comput., vol. 3, no. 4, pp. 245–256, 1992.
[12] S. Kar and G. Hug, “Distributed robust economic dispatch in power sys- [36] A. Wood and B. Wollenberg, Power Generation, Operation and Control,
tems: A consensus+innovations approach,” in Proc. IEEE Power Energy 2nd ed. Hoboken, NJ, USA: Wiley, 1996.
Soc. Gen. Meeting, Jul. 2012, pp. 1–8. [37] C. Lin and G. Viviani, “Hierarchical economic dispatch for piecewise
[13] A. Dominguez-Garcia, S. Cady, and C. N. Hadjicostis, “Decentralized quadratic cost functions,” IEEE Trans. Power Appl. Syst., vol. PAS-103,
optimal dispatch of distributed energy resources,” in Proc. 51st IEEE no. 6, pp. 1170–1175, Jun. 1984.
Conf. Decision Control (CDC), Dec. 2012, pp. 3688–3693. [38] V. Loia and A. Vaccaro, “Decentralized economic dispatch in smart grids
[14] S. Yang, S. Tan, and J.-X. Xu, “Consensus based approach for economic by self-organizing dynamic agents,” IEEE Trans. Syst., Man, Cybern.,
dispatch problem in a smart grid,” IEEE Trans. Power Syst., vol. 28, no. 4, vol. 44, no. 4, pp. 397–408, Apr. 2014.
pp. 4416–4426, Nov. 2013. [39] G. Binetti, A. Davoudi, F. Lewis, D. Naso, and B. Turchiano, “Distributed
[15] W. T. Elsayed and E. F. El-Saadany, “A fully decentralized approach consensus-based economic dispatch with transmission losses,” IEEE
for solving the economic dispatch problem,” IEEE Trans. Power Syst., Trans. Power Syst., vol. 29, no. 4, pp. 1711–1720, Jul. 2014.
vol. 30, no. 4, pp. 2179–2189, Jul. 2015. [40] G. Binetti, A. Davoudi, D. Naso, B. Turchiano, and F. Lewis, “A dis-
[16] W. Zhang, W. Liu, X. Wang, L. Liu, and F. Ferrese, “Online optimal tributed auction-based algorithm for the nonconvex economic dispatch
generation control based on constrained distributed gradient algorithm,” problem,” IEEE Trans. Ind. Informat., vol. 10, no. 2, pp. 1124–1132, May
IEEE Trans. Power Syst., vol. 30, no. 1, pp. 35–45, Jan. 2015. 2014.
[17] L. Gan, U. Topcu, and S. Low, “Optimal decentralized protocol for [41] J. Qiu, G. Feng, and H. Gao, “Fuzzy-model-based piecewise static-
electric vehicle charging,” IEEE Trans. Power Syst., vol. 28, no. 2, output-feedback controller design for networked nonlinear systems,”
pp. 940–951, May 2013. IEEE Trans. Fuzzy Syst., vol. 18, no. 5, pp. 919–934, Oct. 2010.
[18] N. Rahbari-Asr, M.-Y. Chow, Z. Yang, and J. Chen, “Network coopera- [42] J. Qiu, G. Feng, and H. Gao, “Static-output-feedback control of
tive distributed pricing control system for large-scale optimal charging of continuous-time T-S fuzzy affine systems via piecewise Lyapunov func-
PHEVs/PEVs,” in Proc. 39th Conf. IEEE Ind. Electron. Soc. (IECON), tions,” IEEE Trans. Fuzzy Syst., vol. 21, no. 2, pp. 245–261, Apr.
Nov. 2013, pp. 6148–6153. 2013.
[19] N. Rahbari-Asr and M.-Y. Chow, “Cooperative distributed demand man- [43] F. Liu, H. Gao, J. Qiu, S. Yin, J. Fan, and T. Chai, “Networked multirate
agement for community charging of PHEV/PEVs based on KKT condi- output feedback control for setpoints compensation and its application
tions and consensus networks,” IEEE Trans. Ind. Informat., vol. 10, no. 3, to rougher flotation process,” IEEE Trans. Ind. Electron., vol. 61, no. 1,
pp. 1907–1916, Aug. 2014. pp. 460–468, Jan. 2014.
[20] Z. Zhang, X. Ying, and M.-Y. Chow, “Decentralizing the economic dis- [44] T. Wang, H. Gao, and J. Qiu, “A combined adaptive neural network
patch problem using a two-level incremental cost consensus algorithm in and nonlinear model predictive control for multirate networked industrial
a smart grid environment,” in Proc. North Amer. Power Symp. (NAPS), process control,” IEEE Trans. Neural Netw. Learn. Syst., to be published.
Aug. 2011, pp. 1–7.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.
1328 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 63, NO. 2, FEBRUARY 2016

Tao Yang (M’12) received the M.S. degree Yannan Sun (M’11) received the B.S. degree
(with distinction) in control engineering from in mathematics from the University of Science
City University, London, U.K., in 2004, and and Technology of China, Hefei, China, in 2004,
the Ph.D. degree in electrical engineering from and the M.S. degree in statistics and the Ph.D.
Washington State University, Pullman, WA, USA, degree in mathematics from Washington State
in 2012. University, Pullman, WA, USA, in 2007 and
From 2012 to 2014, he was an ACCESS 2010, respectively.
Postdoctoral Researcher with the ACCESS She joined the Pacific Northwest National
Linnaeus Centre, Royal Institute of Technology, Laboratory, Richland, WA, USA, in 2010, and
Stockholm, Sweden. He joined the Pacific is currently a Senior Research Scientist with
Northwest National Laboratory, Richland, WA, the Electricity Infrastructure Group. Her research
USA, in 2014, and is currently a Research Scientist with the Energy interests include multivariate risk (multivariate extremes and risk mea-
and Environment Directorate. His research interests include distributed sures), statistical modeling, machine learning, and applying mathe-
control and optimization in power systems, cyber physical systems, matical and statistical methods in power engineering and building
networked control systems, and multi-agent systems. control.

Di Wu (M’12) received the B.S. and M.S. Jianming Lian (M’10) received the B.S.
degrees in electrical engineering from Shanghai degree (highest honors) in automation from the
Jiao Tong University, Shanghai, China, in 2003 University of Science and Technology of China,
and 2006, respectively, and the Ph.D. degree in Hefei, China, in 2004, and the M.S. and Ph.D.
electrical and computer engineering from Iowa degrees in electrical engineering from Purdue
State University, Ames, IA, USA, in 2012. University, West Lafayette, IN, USA, in 2007 and
He is currently a Senior Power System 2009, respectively.
Engineer with the Electricity Infrastructure He is currently a Senior Research Engineer
Group, Pacific Northwest National Laboratory with the Electricity Infrastructure Group, Pacific
(PNNL), Richland, WA, USA. His research inter- Northwest National Laboratory, Richland, WA,
ests include distributed control and coordination USA. From 2010 to 2011, he was a Postdoctoral
in power systems, operational planning and optimization of energy Research Associate with the Center for Advanced Power Systems,
storage, power system dynamic simulation, impacts of plug-in elec- Florida State University, Tallahassee, FL, USA, where he was involved
tric vehicles on power systems, and planning of national energy and in various projects related to the development of a future all-electric
transportation infrastructure. ship supported by the Office of Naval Research. His research inter-
Dr. Wu serves as an Associate Editor for the IEEE P OWER AND ests include the application of distributed control to the power grid and
E NERGY T ECHNOLOGY S YSTEMS J OURNAL. building systems, and the use of responsive loads for providing grid
services.

Authorized licensed use limited to: NATIONAL INSTITUTE OF TECHNOLOGY SILCHAR. Downloaded on August 22,2022 at 14:16:47 UTC from IEEE Xplore. Restrictions apply.

You might also like