1124 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 10, NO.

2, MAY 2014

A Distributed Auction-Based Algorithm for the

Nonconvex Economic Dispatch Problem
Giulio Binetti, Student Member, IEEE, Ali Davoudi, Member, IEEE, David Naso,
Biagio Turchiano, Senior Member, IEEE, and Frank L. Lewis, Fellow, IEEE

Abstract—This paper presents a distributed algorithm based on multiple fuel options, and prohibited operating zones are preva-
auction techniques and consensus protocols to solve the nonconvex lent in power systems and result in nonconvex ED problems
economic dispatch problem. The optimization problem of the [11].
nonconvex economic dispatch includes several constraints such
as valve-point loading effect, multiple fuel option, and prohibited
The most efficient techniques to address the nonconvex ED
operating zones. Each generating unit locally evaluates quantities problem are computational intelligence methods, mainly ge-
used as bids in the auction mechanism. These units send their netic algorithm (GA) [11]–[13], particle swarm optimization
bids to their neighbors in a communication graph that supports (PSO) [14]–[19], and differential evolution [20], [21]. These
the power system and which provides the required information heuristic algorithms have been shown to be very effective in
flow. A consensus procedure is used to share the bids among the handling nonconvex ED by acting centrally on the system-wide
network agents and resolves the auction. As a result, the power
distribution of generating units is updated and the generation cost
gathered information. The centralized controller requires a high-
is minimized. The effectiveness of this approach is demonstrated bandwidth communication infrastructure and a high level of
by simulations on standard test systems. connectivity, poses reliability concerns due to the presence of
a single point-of-failure, and is prone to modeling error. More-
Index Terms—Auction algorithm, distributed algorithm, eco- over, both the future power grid and the communication network
nomic dispatch (ED), nonconvex optimization, smart grid.
are likely to have a variable topology, which further undermines
the efficacy of centralized mechanisms.
In the smart grid environment, the decision-making proce-
I. INTRODUCTION dures are moving from centralized to distributed frameworks
[22]. Distributed algorithms can be immune to topological

T HE transformation of the legacy electric grid into a smart variations and accommodate desired plug-and-play features
grid [1], [2], which includes renewable energy resources [23]. Moreover, they are scalable and can accommodate a
[3], [4] and electrified transportation fleets [5], requires real large number of distributed generation units. Distributed
time and optimal operations of the power system [6]. One of decision-making can also effectively utilize sparse commu-
the fundamental optimization challenges is the economic dis- nication infrastructures with limited message passing among
patch (ED) problem that aims at minimizing the overall genera- participating agents. Recently, distributed approaches, based
tion cost, allocating power generation among available units [7]. on consensus protocols and iterative procedures, have been
This optimization framework acts on the cost functions of the proposed to solve the ED problem [24]–[26] considering only
generating units. Conventional centralized approaches, e.g., the convex cost functions and neglecting the transmission loss.
lambda iteration method [8] and interior point method [9], re- This paper proposes an optimization technique based on the
quire the cost function to be convex. Although convex cost func- market-based paradigm for multi-agent systems [27] and auc-
tions are suitable for preliminary analysis, they can lead to poor tion theory [28], [29]. The main contributions of this paper are
solutions since, in practical cases, cost functions are often non- given here.
smooth and nonconvex [10]. The valve-point loading effects, • The nonconvex ED problem, including transmission losses,
valve-point loading effect, multiple fuel option, and pro-
hibited operating zones is considered.
Manuscript received June 13, 2013; revised September 03, 2013; accepted
• The proposed approach is fully distributed, i.e., it obvi-
October 17, 2013. Date of publication November 01, 2013; date of current ates the requirement for a central authority, thus making
version May 02, 2014. This work was supported in part by the National the system more fault-tolerant (in centralized systems, the
Science Foundation under Grant ECCS-1137354 and Grant ECCS-1128050, failure of the supervisor node implies the blocking of the
the Office of Naval Research under Grant N00014-13-1-0562 and ARO Grant
W911NF-11-D-0001. Paper no. TII-13-0385. entire system). It also requires only a sparse communica-
G. Binetti is with the Polytechnic of Bari, Bari 70125, Italy, and also with tion structure which is cheaper and can be more reliable.
the University of Texas at Arlington, Arlington, TX 76015 USA (e-mail: To the best knowledge of the authors, no other distributed
[email protected]; [email protected]).
D. Naso and B. Turchiano are with the Polytechnic of Bari, Bari 70125, Italy approach has yet been proposed to solve the nonconvex ED
(e-mail: [email protected]; [email protected]). problem.
A. Davoudi and F. L. Lewis are with the University of Texas at Arlington, • The proposed distributed technique is leaderless, i.e., it
Arlington, TX 76015 USA (e-mail: [email protected]; [email protected]).
Color versions of one or more of the figures in this paper are available online
does not rely on a leader node which could jeopardize a
at http://ieeexplore.ieee.org. solution’s configurability and pose reliability concerns if
Digital Object Identifier 10.1109/TII.2013.2287807 the leader node is compromised.

1551-3203 © 2014 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.
BINETTI et al.: DISTRIBUTED AUCTION-BASED ALGORITHM FOR THE NONCONVEX ECONOMIC DISPATCH PROBLEM 1125

• The proposed auction algorithm is deterministic, i.e., The transmission losses are represented by the Kron’s for-
repeated algorithm executions yield the same identical mula [7]
results. This is in direct contrast with majority of existing
centralized algorithms based on computational intelli-
gence (e.g., PSO and its variants) that perform a heuristic (6)
search with no guarantees on the uniqueness of the final
solution found by a single run.
This paper is organized as follows. In Section II, the non- where , , and are the loss coefficients.
convex ED problem problem with transmission losses is dis- The constraints in (3) bound the output powers to admis-
cussed. Section III introduces some preliminaries on market- sible ranges defined by the lower and upper limits for each
based approach, auctions, and consensus protocols. The pro- unit .
posed distributed approach is presented in Section IV, and the
simulation results in Section V show its effectiveness. B. Nonconvex Cost Function
A practical cost function encompasses a series of nonsmooth
II. NONCONVEX ECONOMIC DISPATCH curves to represent the nondifferentiable points due to the pres-
The following notation is adopted here: denotes the ele- ence of the valve-point loading effects and the multiple fuel op-
ment of vector , refers to element of the matrix , tion. To include the valve-point loading effects, an additional
denotes the variable carried by the unit , means that rectified sinusoidal term is considered in the quadratic cost func-
is a function of , and denotes the variable at the dis- tion [11], [30]
crete time instant .
(7)
A. Economic Dispatch Problem
where , , , , and are the cost coefficients of unit .
The ED minimizes the total generation cost given by the Since generating units are practically supplied with multiple
sum of the generation costs for each unit , , as follows: fuel sources, each generating unit uses the most economical fuel
option for a given interval of their operating range. Once both
(1) the valve-point loading effect and the multiple fuel option are
considered, a practical cost function is formulated as [11], [15]
subject to

(2)

(3)

where , , and
denote the vectors of output powers , lower (8)
power bound , and upper power bound , for each unit , re- where , , , , and are the cost coefficients of unit
spectively. Conventionally, the generation cost function is ex- for fuel .
pressed in the following quadratic form [7]:
III. PRELIMINARIES
(4)
A. Market-Based Approach
where , , and are the cost coefficients of unit . Equation
The principles of market economics can be used to coordinate
(2) represents the generation-demand equality constraint, i.e.,
multi-agent systems. In general, market-based approaches share
the sum of the generated powers is equal to the active
some underlying elements [27] that can also be recognized for
load demand plus the transmission losses, .
the ED. The problem formulation should be decomposable into
To ensure stable operation of generation units, prohibited op-
subtasks that each agent can perform. A global objective func-
erating zones (POZs) are considered [14]. This concept is in-
tion measures a solution’s quality. Each agent has a local ob-
cluded in the ED problem as the following constraints:
jective function describing its preferences for performing tasks,
and a mapping exists between the global and the local objec-
tive functions. For the ED problem, the global objective is to
generate , which is decomposed into individual unit powers
(5) , each assigned to a unit. The global objection function is the
where and are the lower and upper bounds for the pro- total generation cost , the local objective functions are the
hibited zone of unit , is the number of POZs for unit , and the mapping between global and local functions is
, and is the number of units with POZs. Note that when provided in (1). Thus, such problem formulations are suitable to
POZs are considered, the fuel-cost characteristics will have dis- be solved with trading mechanisms such as auctions.
continuities leading to a nonconvex solution space even with An auction is a structured negotiation mechanism described
quadratic cost functions. by a set of explicit rules that specify how tasks are assigned and
1126 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 10, NO. 2, MAY 2014

how the winner is determined [28], [29]. The auction mech- its output power . Each unit computes the following two
anism is characterized by an announcement phase listing the bids:
tasks to be performed, a bidding phase where agents can submit
bids and make offer for the listed tasks, and a clearing phase
in which the winner for each task is chosen. In a well-designed if is admissible
mechanism, the cost minimization (or reward maximization) of otherwise
a single agent should contribute to the cost minimization (or (9)
reward maximization) of the overall system. Interested readers if is admissible
can refer to [27]–[29] for a detailed discussion. otherwise.
(10)
B. Graph Theory and Consensus Protocols
The value in (9) represents the amount that unit
A distributed solution must be supported by a communication
has to spend to increase its generation from the current value
network that provides its information flow, i.e., each generating
to a new value . The value in (10) rep-
unit exchanges information with a subset of other units to make
resents the amount that unit can save by reducing its genera-
autonomous decisions. The information flow is represented by
tion from the current value, , to a new value, . Each
modeling the communication network by a graph ,
unit considers the constraints on prohibited operating zones and
where is the set of nodes, and
lower and upper generation limits, placing a zero bid when in-
is the set of edges. An edge from node to node
creasing or reducing the power generation violates operational
exists if node can receive information from node . Node
constraints.
is the neighbor of node if . The set of node
2) Consensus Procedure: Each unit uses the quantities in (9)
neighbors is denoted by . The graph is assumed to be simple,
and (10), and , as bids in the auction procedure. Using
i.e., there are neither self-loops nor multiple edges between the
a sparse communication infrastructure that only communicates
same pair of nodes. The units are represented by nodes on the
with neighbors, the units perform a consensus procedure on the
communication graph, while the communication links among
highest bid and the corresponding bidding unit for both and
units are modeled by edges. Each unit is endowed with a state
. Thus, each unit reaches a consensus on final values denoted
characterized by certain dynamics. In consensus problems [31],
by , , , and , which are the highest bid for adding an amount
all units reach the same state, i.e., converge to the same value
of power, the highest bid for removing an amount of power,
by using only local communications with their neighbors.
the unit placing the bid , and the unit placing the bid , respec-
tively. The consensus procedure is detailed in Section IV-D.
IV. DISTRIBUTED AUCTION-BASED ALGORITHM 3) Auction Resolution and Swap Operation: Each unit eval-
FOR THE NONCONVEX ED uates the difference after the consensus. It
should be noted that , with and , implies that
The main principles of the proposed auction algorithm are exchanging amount of power leads to saving an amount equal
presented in Section IV-A. Multistep and multiswap options are to . Thus, after the consensus procedure, units and can be
detailed in Sections IV-B and IV-C. The required consensus pro- recognized as the winning units. If , these units can update
cedures are described in Section IV-D. Finally, the transmission their powers as
losses are handled in Section IV-E.
(11)
A. Distributed Auction-Based Algorithm
This operation is referred to as a swap. The consensus proce-
The proposed algorithm belongs to an important category of
dure in step 2 ensures that, given the current power allocation,
auction procedures, called exchanges [29], where each agent
the pair provides the highest global saving when the swap
acts as both the buyer and the seller. The general idea is to design
operation is performed. The algorithm iterates through steps
two-sided auctions, also known as double auctions, where the
1–3, leading to a new solution with a lower cost at each swap op-
generating units can change their output power by negotiating
eration. The algorithm reaches the final solution when no further
with other units. The choice of each unit is driven by the min-
swaps are possible, i.e., when no exchange of power between
imization of the overall generation cost, while the constraints
any pair of units provides a lower-cost solution. Finally, note
are satisfied. Consensus protocols, inspired by multirobot task
that a swap operation does not change the amount of allocated
allocation [32] implement this idea in a distributed fashion. The
power in the system. Since the algorithm starts from a solu-
system is initialized with a guess solution that allocates the re-
tion fulfilling the generation-demand constraint ,
quired power and avoids the prohibited zones, with no consid-
such an equality constraint is always accommodated during AA.
erations concerning the cost. The initial guess generation is a
simple task that can be performed with an initialization proce-
B. Multistep Option
dure that is omitted due to space limitation (a short discussion of
the initialization criteria is given in Section V). The basic steps The basic AA relies on the scalar parameter that defines the
of the algorithm, named Auction Algorithm (AA), are described discretized step for the search process. The algorithm may get
here. stuck in a local minima since a fixed does not permit a flexible
1) Bids Evaluation: Assume that a parameter specifies a search in the solution space. Therefore, the basic algorithm in
predefined amount of power. Suppose that each unit carries only Section IV-A is improved with a multistep option, as shown in
BINETTI et al.: DISTRIBUTED AUCTION-BASED ALGORITHM FOR THE NONCONVEX ECONOMIC DISPATCH PROBLEM 1127

Algorithm 1. Assume that each unit knows a pre-defined param- representing the bids of unit using the parameter . Step
eter vector, , named percentage vector, with elements 3 is then updated as follows. Each unit can compute the values

Algorithm 1: AA with multistep option Algorithm 2: AA with multistep and multiswap option

Input: , , , Input: , , ,
Set: , , ; Set: , , ;
for do for do
Broadcast to neighbors. Broadcast to neighbors.
Perform (15)–(17) for each element of to reach Receive from neighbors.
consensus. Perform (18) for each element of to reach
if is multiple of then consensus.
Consensus on . Evaluate . if is multiple of then
Find maximum with and . Consensus on . Evaluate
Units perform swap in (14) with . .
Update bids using (12), (13). Create a list of swap operations in descending order
end if of .
end for Perform swap in (14) for each element in the swap list.
Update bids using (12), (13).
. Each unit can calculate the vector end if
defined as , with elements , i.e., is the end for
pre-defined percentage of the current load demand . For
example, means that is 10
percent of the demand. Accordingly, the bid evaluation in step corresponding to the cost saving when units
1 is updated to consider the vector . Each unit computes the bid and increase and decrease their output powers by , re-
vectors as spectively. In general, many elements can be positive, with
and . Thus, each unit identifies the set of pos-
sible swap operations and creates an ordered list in a descending
if is admissible order of . It is important that each unit is involved in no
otherwise more than one swap for each step 3. Suppose that a unit per-
(12) forms a swap operation for , and then the previous power
if is admissible is updated, say . The bids and with
otherwise are no longer meaningful since they were evaluated considering
(13) and not the updated value . Each unit must be involved
in no more than one swap for each step 3 to perform mean-
with . Then, the consensus procedure in step 2 is ingful swaps. Thus, the swaps operation list needs to be eval-
repeated for each element , i.e., the consensus is reached on uated accordingly. Finally, each unit modifies its output power
the final vectors with elements , , , and according to the swap operations in the list.
. In step 3, all units evaluate the vector and
can eventually identify different swaps with and leading D. Consensus Procedure
to the cost saving for different values of The AA is composed of two levels, as shown in Fig. 1.
. Then, the swap operation is performed between the units The first level represents the bid evaluation and swap opera-
and giving the maximum cost saving for the tions, while the second level is the consensus procedure. This
corresponding as follows: two-level architecture operates on two time scales, as shown in
Fig. 2. At each time instant , all units communicate with each
(14) other to perform the consensus procedure (second level), while
only at time instant the units apply the swaps and evaluate
The design parameter should include small elements to refine the bids for the next iteration (first level). The time instant
the search near the solution and large elements to explore the is a multiple of a certain , which is defined in the following
solution space and go beyond the POZs. section.
1) Consensus for the Multistep Option: This consensus is
used by the AA with multistep option to find the highest bid and
C. Multiswap Option
the corresponding bidding unit for each and in (12) and
The multiswap option can be used to speed up the AA, as (13), as shown in Algorithm 1. Each unit carries two scalars:
shown in Algorithm 2. Assume that the multistep option is used is the unit index currently winning the auction according to
and suppose that step 2 leads to a consensus on all of the bids the knowledge of unit , and is the corresponding bid. At
placed by all of the units (the details of this consensus procedure each time instant of the first level, , each unit initializes
are described in Section IV-D). Thus, the consensus is reached with the local bid and with the unit index . Then, at each
on some bids matrices , with elements and time instant of the second level , each unit sends the values
1128 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 10, NO. 2, MAY 2014

properly when , because such a choice leads to con-

sensus when the first level of the algorithm has meaningful in-
formation to resolve the auction. In practical application, if the
graph diameter is unknown, one can use the conservative choice
of , since is the greatest possible diameter for
a given number of units.
2) Consensus for Multistep and Multiswap Option: This con-
sensus is used by the AA with multistep and multiswap option
to obtain matrices and with all of the bids placed in the
network, as shown in Algorithm 2. In this case, each unit car-
ries a vector where each element is the bid of unit as
known by unit . Starting with a zero vector at each instant
time , each unit performs the following update at each time
instant :

(18)

that is, each unit saves, in the th element of vector , the

maximum bid of unit known in its neighborhood. The con-
sensus is reached within a number of iterations equal to the
graph diameter, . Thus, it is possible to choose or,
alternatively, if the graph diameter is unknown.

E. Incorporating the Transmission Losses

The ED problem with the transmission losses is analytically
Fig. 1. General flowchart for a generating unit.
intractable. An approximation is proposed that leads to a very
satisfactory solution (less than 0.1% is reported in case studies).
Kron’s formula in (6) requires that the entire solution is known
in order to evaluate losses. Thus, the AA is modified assuming
that each unit carries an output power vector , where
is the output power of unit known by the unit . Each unit
locally updates its output power (at the first level), while the
consensus is reached on the entire vector using the consensus
Fig. 2. Timeline for the two-level architecture. procedure in Section IV-D (at the second level).
The output powers in does not include the transmission
losses. Then, assume that each unit evaluates the losses associ-
and to the neighbors, receives the values and from ated with its output power as
its neighbors, and performs the following updates:
(19)

(15)
based on the power vector and the losses coefficients in (6).
(16) The proposed approximation defines the output power including
the losses, denoted by , as
with
(20)
(17)
Equation (20) uses the losses calculated by the current solution
This means that each unit saves the highest value among the to approximate the actual powers when losses are consid-
received bids and its local information as in (15), and the corre- ered. The rationale for this choice is that the losses for the cur-
sponding bidding unit index as in (16) and (17). The units reach rently known solution are very close to the losses for the
consensus on both the highest bid and the bidding unit within unknown solution , that is, . Thus, the
a number of iterations equal to the communication graph di- approximation (20) is very satisfactory. In addition, to include
ameter, . At each iteration, , the information of each unit is the available loss information in the decision process, it is con-
shared with its neighbors. Thus, the number of iterations needed venient to modify the bids evaluation by using the approximated
to share the information among the entire network is equal to the output power including the losses instead of the values
greater minimum distance between any two nodes in the graph, in (12) and s(13). This minimizes the generation cost with the
i.e., the graph diameter. This means that the algorithm works losses considered.
BINETTI et al.: DISTRIBUTED AUCTION-BASED ALGORITHM FOR THE NONCONVEX ECONOMIC DISPATCH PROBLEM 1129

TABLE I TABLE II
SOLUTION FOR THE 10-GENERATOR SYSTEM 2700 MW WITH SOLUTION FOR THE 15-GENERATOR SYSTEM 2630 MW WITH
VALVE-POINT LOADING EFFECT AND MULTIPLE FUEL OPTIONS PROHIBITED OPERATING ZONES AND TRANSMISSION LOSSES

V. CASE STUDIES
Here, the effectiveness of the proposed distributed auction- TABLE III
based algorithm is verified by numerical simulation on three test SOLUTION FOR THE 40-GENERATOR SYSTEM 10 500 MW WITH
systems with nonconvex solution spaces. VALVE-POINT LOADING EFFECT

A. Description of the Test Systems

The first system is composed of ten units with cost functions
modeling both valve-point loading effect and multiple fuel op-
tion. The first unit has two fuel options, while the remaining
units have three fuel options. The detailed characteristics of all
units are given in [11]. The load demand is 2700 MW.
The second test system is composed of 15 units with quadratic
cost functions, POZs for four units, and transmission-line losses.
The detailed characteristics of all units are given in [14]. The
load demand is 2630 MW.
The third test system is composed of 40 units with cost func-
tions modeling the valve-point loading effect. The detailed char-
TABLE IV
acteristics of all units are given in [30]. The load demand is COST FOR THE 10-GENERATOR TEST SYSTEM
10 500 MW.

B. Simulation Setting
The AA search mechanism is characterized by the per-
centage vector that should include small enough values to
refine the search near the final solution, and large enough
values both to explore the solution space and to go beyond
the POZs. Thus, in both the 10- and 15-generator test cases,
the percentage vector is set to .
In the 40-generator test case, the percentage vector is set to
. The latter vector
contains more elements to improve the quality of the search
process since the solution space is much bigger than other test
cases. In both vectors, the biggest element (equal to 10%) has C. Solution Quality
been chosen considering that the biggest POZ is 40 MW, that Tables I–III report the complete power allocation obtained by
is, the 2.7% of the maximum admissible load demand. the proposed AA with multistep and multiswap options for the
All of the generating units are initialized using the average 10-, 15- and 40-generator test systems, respectively. In Table I,
values between the lower and upper limits, scaled to allocate the fuel type is reported for each unit. To validate the effec-
the demand , that is, the initial output powers are set to tiveness of the proposed approach, Tables IV–VI present the
, where is the scale factor that leads to al- comparison of the total cost obtained by the AA and the best
locating the demanded power without considering the losses. If available centralized algorithms. The results show that the so-
the initial condition for a unit lies in a prohibited zone, then its lution obtained by the proposed AA provides a lower cost with
output power is reduced to the lower bound of that prohibited respect to many centralized algorithms and it is very close to the
zone, and an equal amount of power is added to the next unit in cost obtained by the best algorithm (the percentage difference
an increasing index order. This initialization criterion has been is 0.022%, 0.094%, and 0.302% in the three test cases). Thus,
chosen due to its suitability for a distributed implementation. Fi- in addition to the benefits of the distributed nature, the AA per-
nally, a preliminary simulation campaign suggests that the spe- formance can be considered very satisfactory. Moreover, all the
cific initialization values do not affect the search mechanism of centralized algorithms are reported in terms of minimum, av-
the AA if the percentage vector contains sufficiently large values erage, and maximum costs among a certain number of trials. As
to go beyond the POZs. opposed to these heuristic approaches, the AA is deterministic
1130 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 10, NO. 2, MAY 2014

TABLE V TABLE VIII

COST FOR THE 15-GENERATOR TEST SYSTEM NUMBER OF ITERATIONS FOR DIFFERENT TOPOLOGIES OF THE COMMUNICATION
GRAPH SUPPORTING THE POWER SYSTEM

TABLE VI
COST FOR THE 40-GENERATOR TEST SYSTEM

Fig. 3. (a) Line and (b) fully connected topologies for the communication graph
supporting the power system in an illustrative example with 4 units.

TABLE VII
CPU TIME PER ITERATION FOR AA WITH AND WITHOUT MULTISWAP OPTION

Fig. 4. Generation cost for the test systems with (a) 10-, (b) 15-, and (c) 40-gen-
erating units in case of AA with multistep option (dashed line) and AA with
multistep and multiswap options (continuous line), in case of a line network
(blue line) and a fully-connected network (red line). The circles denote the time
in the sense that, given the initial parameters, repeated execu- instant when the algorithm reaches the final solution.
tions of the algorithm always give the same solution, that is,
the best AA solution can be achieved by a single run. One can
also examine the approximation in handling the transmission D. Computational Efficiency and Convergence Analysis
losses. In Table II, the final solution provides 2660.42 MW of The distributed nature of the AA affects both computational
generated power with 31.38 MW of losses, i.e., a net amount of and convergence times. It should be noted that the AA has a
2629.04 MW. This means an approximation error of 0.96 MW two-level architecture characterized by that has to be equal to
equal to 0.036% of the total demand. However, the solution pro- the diameter of the communication network among the gener-
viding the exact amount of required power is shown in Table II ators. The computational efficiency analysis is carried out based
and compared with the other algorithms in Table V. on CPU time elapsed for first- and second-level iterations, while
BINETTI et al.: DISTRIBUTED AUCTION-BASED ALGORITHM FOR THE NONCONVEX ECONOMIC DISPATCH PROBLEM 1131

the convergence time is analyzed considering its dependence on [2] V. C. Gungor, D. Sahin, T. Kocak, S. Ergut, C. Buccella, C. Cecati,
the communication network. Both a line and a fully-connected and G. P. Hancke, “A survey on smart grid potential applications and
communication requirements,” IEEE Trans. Ind. Inf., vol. 9, no. 1, pp.
topologies are considered for the communication network since 28–42, Feb. 2013.
they are characterized by minimum and maximum diameters, as [3] F. Yao, Z. Y. Dong, K. Meng, Z. Xu, H. H.-C. Iu, and K. P. Wong,
shown in Fig. 3. “Quantum-inspired particle swarm optimization for power system op-
erations considering wind power uncertainty and carbon tax in aus-
The computer simulations are carried out using MATLAB on tralia,” IEEE Trans. Ind. Inf., vol. 8, no. 4, pp. 880–888, Nov. 2012.
a PC with a Core2 Duo processor (3 GHz) and 4 GB RAM. [4] J. Zhao, F. Wen, Z. Y. Dong, Y. Xue, and K. P. Wong, “Optimal
The results for the AA with and without multiswap option are dispatch of electric vehicles and wind power using enhanced particle
shown in Table VII. The CPU time for the second-level iteration swarm optimization,” IEEE Trans. Ind. Inf., vol. 8, no. 4, pp. 889–899,
Nov. 2012.
is not affected by the multiswap option, while the CPU time for [5] W. Su, H. R. Eichi, W. Zeng, and M.-Y. Chow, “A survey on the elec-
the first-level iteration is longer in the case of multiswap op- trification of transportation in a smart grid environment,” IEEE Trans.
tion since the AA performs more swap operations per iteration. Ind. Inf., vol. 8, no. 1, pp. 1–10, Feb. 2012.
[6] P. Siano, C. Cecati, H. Yu, and J. Kolbusz, “Real time operation of
Moreover, the second-level iteration requires more CPU time smart grids via FCN networks and optimal power flow,” IEEE Trans.
due to the necessary operations on the exchanged data among Ind. Inf., vol. 8, no. 4, pp. 944–952, Nov. 2012.
the generators. The CPU time increases with the size of the [7] A. J. Wood and B. F. Wollenberg, Power Generation, Operation, Con-
trol. New York, NY, USA: Wiley, 1996.
system but remains low even for the large-scale test system with [8] C. E. Lin and G. L. Viviani, “Hierarchical economic dispatch for piece-
40 generators. wise quadratic cost functions,” IEEE Trans. Power App. Syst., vol.
Considering the AA with and without the multiswap option, PAS-103, no. 6, pp. 1170–1175, Jun. 1984.
Table VIII shows the number of iterations required to reach the [9] W.-M. Lin and S.-J. Chen, “Bid-based dynamic economic dispatch
with an efficient interior point algorithm,” Int. J. Electr. Power Energy
final solution, the 98% and 95% of cost optimality in case of Syst., vol. 24, no. 1, pp. 51–57, Jan. 2002.
both line and fully connected network. Fig. 4 shows the cost de- [10] H.-T. Yang, P.-C. Yang, and C.-L. Huang, “Evolutionary program-
creasing over time. It is evident that the multiswap option can ming based economic dispatch for units with non-smooth fuel cost
functions,” IEEE Trans. Power Syst., vol. 11, no. 1, pp. 112–118,
notably speed up the algorithm. Such an option lets the algo- Feb. 1996.
rithm quickly find a satisfactory solution within a few iterations [11] C.-L. Chiang, “Improved genetic algorithm for power economic dis-
before reaching the optimal solution. The number of first-level patch of units with valve-point effects and multiple fuels,” IEEE Trans.
Power Syst., vol. 20, no. 4, pp. 1690–1699, Nov. 2005.
iterations does not depend on the topology of the communica- [12] N. Amjady and H. Nasiri-Rad, “Nonconvex economic dispatch with ac
tion graph, while the number of second-level iterations is re- constraints by a new real coded genetic algorithm,” IEEE Trans. Power
lated to its diameter. Thus, one can design the communication Syst., vol. 24, no. 3, pp. 1489–1502, Aug. 2009.
network, which is different from the underlying power system, [13] N. Amjady and H. Nasiri-Rad, “Economic dispatch using an efficient
real-coded genetic algorithm,” IET Gener., Transm. Distrib., vol. 3, no.
to speed up the converge time. 3, pp. 266–278, Mar. 2009.
[14] Z.-L. Gaing, “Particle swarm optimization to solving the economic dis-
VI. CONCLUSION patch considering the generator constraints,” IEEE Trans. Power Syst.,
vol. 18, no. 3, pp. 1187–1195, Aug. 2003.
A distributed algorithm to solve the nonconvex economic dis- [15] A. I. Selvakumar and K. Thanushkodi, “A new particle swarm opti-
patch problem is presented. The problem model incorporates mization solution to nonconvex economic dispatch problems,” IEEE
transmission losses, power generation constraints, prohibited Trans. Power Syst., vol. 22, no. 1, pp. 42–51, Feb. 2007.
[16] K. T. Chaturvedi, M. Pandit, and L. Srivastava, “Self-organizing hi-
operating zones, valve-point loading effect, and multiple fuel erarchical particle swarm optimization for non-convex economic dis-
option. These constraints lead to a nonconvex solution space patch,” IEEE Trans. Power Syst., vol. 23, no. 3, pp. 1079–1087, Aug.
with nondifferentiable points and discontinuities. The proposed 2008.
[17] B. K. Panigrahi, V. R. Pandi, and S. Das, “Adaptive particle swarm
method tackles this problem by introducing an auction-based al- optimization approach for static and dynamic economic load dispatch,”
gorithm to handle the combinatorial and nonconvex problem na- Energy Convers. Manag., vol. 49, no. 6, pp. 1407–1415, Jun. 2008.
ture, and by relying on leaderless consensus protocols to obtain a [18] A. I. Selvakumara and Thanushkodi, “Optimization using civilized
swarm: Solution to economic dispatch with multiple minima,” Electr.
distributed procedure. The decentralization of the decision algo- Power Syst. Res., vol. 79, no. 1, pp. 8–16, Jan. 2009.
rithm bears two fundamental advantages, namely it distributes [19] J. Sun, V. Palade, X.-J. Wu, W. Fang, and Z. Wang, “Solving the power
the computation load among a network of interconnected units economic dispatch problem with generator constraints by random drift
with much smaller computational power, and it makes the de- particle swarm optimization,” IEEE Trans. Ind. Inf., vol. PP, no. 99,
DOI: 10.1109/TII.2013.2267392.
cision system much more robust and fault tolerant, since it re- [20] A. Bhattacharya and P. K. Chattopadhyay, “Hybrid differential evo-
moves the criticality of the central node. The effectiveness of the lution with biogeography based optimization for solution of economic
proposed approach is verified by numerical simulations on dif- load dispatch,” IEEE Trans. Power Syst., vol. 25, no. 4, pp. 1955–1964,
Nov. 2010.
ferent test systems. The approach is deterministic (i.e., repeat- [21] N. Nomana and H. Iba, “Differential evolution for economic load
able) and has shown very satisfactory performance when com- dispatch problems,” Electr. Power Syst. Res., vol. 78, no. 8, pp.
pared to the conventional centralized algorithms. Future work 1322–1331, Aug. 2008.
[22] Q. Yang, J. A. Barria, and T. C. Green, “Communication infrastructures
will extend the problem formulation to address the integration of for distributed control of power distribution networks,” IEEE Trans.
renewable energy resources and the presence of energy storage Ind. Inf., vol. 7, no. 2, pp. 316–327, May 2011.
units. [23] S. Dasgupta, S. N. Mohan, S. K. Sahoo, and S. K. Panda, “A plug
and play operational approach for implementation of an autonomous-
micro-grid system,” IEEE Trans. Ind. Inf., vol. 8, no. 3, pp. 615–629,
REFERENCES Aug. 2012.
[1] V. C. Gungor, D. Sahin, T. Kocak, S. Ergut, C. Buccella, C. Cecati, and [24] R. Mudumbai, S. Dasgupta, and B. B. Cho, “Distributed control for
G. P. Hancke, “Smart grid technologies: Communication technologies optimal economic dispatch of a network of heterogeneous power gen-
and standards,” IEEE Trans. Ind. Inf., vol. 7, no. 4, pp. 529–539, Nov. erators,” IEEE Trans. Power Syst., vol. 27, no. 4, pp. 1750–1760, Nov.
2011. 2012.
1132 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 10, NO. 2, MAY 2014

[25] Z. Zhang and M.-Y. Chow, “Convergence analysis of the incremental David Naso was born in Salerno, Italy, on April 29,
cost consensus algorithm under different communication network 1967. He received the Laurea degree (with highest
topologies,” IEEE Trans. Power Syst., vol. 27, no. 4, pp. 1761–1768, honors) in electronic engineering and Ph.D. degree in
Nov. 2012. electrical engineering from the Polytechnic of Bari,
[26] G. Binetti, M. I. Abouheaf, F. L. Lewis, D. Naso, A. Davoudi, and B. Bari, Italy, in 1994 and 1998, respectively.
Turchiano, “Distributed solution for the economic dispatch problem,” During his Ph.D. studies, he was also a Guest
in Proc. 21st Mediterranean Conf. Control Autom.(MED), 2013, pp. Researcher with the Operation Research Institute,
243–250. Technical University of Aachen, Aachen, Germany,
[27] M. B. Dias, R. Zlot, N. Kalra, and A. Stentz, “Market-based multirobot in 1997. Since 1999, he has been an Assistant
coordination: A survey and analysis,” Proc. IEEE, vol. 94, no. 7, pp. Professor of automatic control with the Department
1257–1270, Jul. 2006. of Electric and Electronic Engineering, Polytechnic
[28] E. Wolfstetter, “Auctions: An introduction,” J. Econ. Surveys, vol. 10, of Bari, Bari, Italy, where he currently serves as the Technical Head of the
no. 4, pp. 367–420, Dec. 1996. Robotics Laboratory. Since 2010, he is also a member of the Executive Com-
[29] Y. Shoham and K. Leyton-Brown, Multiagent Systems: Algorithmic, mittee of his University. He has been a Visiting Professor with the University
Game-Theoretic, Logical Foundations. Cambridge, U.K.: Cam- of Saarland, Saarbrcken, Germany, during the summer of 2013. He is one of the
bridge Univ., 2008. cofounders of Energy Factory Bari, an integrated multidisciplinary laboratory
[30] N. Sinha, R. Chakrabarti, and P. K. Chattopadhyay, “Evolutionary pro- for research in aerospace and energy based on the partnership between Avio
gramming technique for economic load dispatch,” IEEE Trans. Evol. S.p.A. and the Polytechnic of Bari. His current research interests focus on
Computation, vol. 7, no. 1, pp. 83–94, Feb. 2003. precision motion control, control of innovative actuators based on smart
[31] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mo- materials, distributed automation, and multi-agent systems.
bile autonomous agents using nearest neighbor rules,” IEEE Trans.
Autom. Control, vol. 48, no. 6, pp. 988–1001, Jun. 2003.
[32] G. Binetti, D. Naso, and B. Turchiano, “Decentralized task allocation
for surveillance systems with critical tasks,” Robot. Autonomous Syst. Biagio Turchiano (SM’01) was born in Bitetto, Bari,
, vol. 61, no. 12, pp. 1653–1664, Dec. 2013. Italy. He received the Laurea degree in electrical en-
[33] P. S. Manoharan, P. S. Kannan, S. Baskar, and M. W. Ruthayarajan, gineering (with honors) from the University of Bari,
“Penalty parameter-less constraint handling scheme based evolu- Bari, Italy, in 1979.
tionary algorithm solutions to economic dispatch,” IET Gener., He joined the Department of Electrical and Elec-
Transm. Distrib., vol. 2, no. 4, pp. 478–490, Jul. 2008. tronic Engineering, Polytechnic of Bari, Bari, Italy,
[34] S. Pothiya, I. Ngamroo, and W. Kongprawechnon, “Application of in 1984, as an Assistant Researcher. He is currently a
multiple tabu search algorithm to solve dynamic economic dispatch Full Professor of automatic control at the same uni-
considering generator constraints,” Energy Conv. Manag., vol. 49, no. versity. His research and teaching interests are in the
4, pp. 506–516, Apr. 2008. areas of production automation, systems and control
[35] S. Khamsawang and S. Jiriwibhakorn, “Dspso-tsa for economic dis- theory, and modeling and control of discrete event
patch problem with nonsmooth and noncontinuous cost functions,” En- systems.
ergy Conv. Manag., vol. 51, no. 2, pp. 365–375, Feb. 2010. Prof. Turchiano is an associate editor of the IEEE TRANSACTIONS ON
[36] J.-B. Park, Y.-W. Jeong, J.-R. Shin, and K. Y. Lee, “An improved par- AUTOMATION SCIENCE AND ENGINEERING.
ticle swarm optimization for nonconvex economic dispatch problems,”
IEEE Trans. Power Syst., vol. 25, no. 1, pp. 156–166, Feb. 2010.

Giulio Binetti was born in Bari, Italy, on November

Frank L. Lewis (F’94) received the B.S. degree
2, 1982. He received the B.S. and M.S. degrees in
in physics and electrical engineering and M.S.E.E.
computer science engineering from the Polytechnic
degree from Rice University, Houston, TX, USA,
of Bari, Bari, Italy, in 2004 and 2006, respectively,
the M.S. degree in aeronautical engineering from
where he is currently working toward the Ph.D.
the University of West Florida, Pensacola, FL, USA,
degree in electrical and information engineering
and the Ph.D. degree from the Georgia Institute of
through a joint program at the “Scuola Interpolitec-
Technology, Atlanta, GA, USA.
nica di Dottorato”.
He is a Distinguished Scholar Professor, Distin-
During his Ph.D. studies, he has been also a Vis-
guished Teaching Professor, and Moncrief-ODonnell
iting Research Scholar and a Research Assistant with
Chair with the University of Texas at Arlington Re-
the University of Texas at Arlington, Arlington, TX,
search Institute, Arlington, TX, USA. He is the author
USA. His current research interests include consensus protocols, distributed
or coauthor of six U.S. patents, numerous journal special issues, journal papers,
computation, auction theory, multirobot systems, and smart grid.
and 14 books, including Optimal Control, Aircraft Control, Optimal Estima-
tion, and Robot Manipulator Control , which are used as university textbooks
worldwide. He is a Distinguished Visiting Professor with Nanjing University
of Science and Technology and a Project 111 Professor with Northeastern Uni-
Ali Davoudi (S’04–M’11) received the Ph.D. degree versity, Shenyang, China. He works in feedback control, intelligent systems,
in electrical and computer engineering from the Uni- cooperative control systems, and nonlinear systems.
versity of Illinois at Urbana-Champaign, Urbana, IL, Dr. Lewis is a Fellow of the IFAC and the U.K. Institute of Measurement
USA, in 2010. and Control, and he is a U.K. Chartered Engineer. He was the recipient of the
He is currently an Assistant Professor with the Fulbright Research Award, the National Science Foundation Research Initia-
Electrical Engineering Department, University of tion Grant, the ASEE Terman Award, the International Neural Network Society
Texas at Arlington, Arlington, TX, USA. He worked Gabor Award, the U.K. Institute of Measurement and Control Honeywell Field
for Solar Bridge Technologies, Texas Instruments Engineering Medal, and the IEEE Computational Intelligence Society Neural
Inc., and Royal Philips Electronics. His research Networks Pioneer Award. He was also the recipient of the Outstanding Service
interests are various aspects of modeling and control Award from the Dallas IEEE Section and selected as Engineer of the Year by
of power electronics, energy conversion systems, the Ft. Worth IEEE Section. He was listed in the Ft. Worth Business Press Top
and finite-inertia power systems. 200 Leaders in Manufacturing and is a Founding Member of the Board of Gov-
Prof. Davoidi is an associate editor for the IEEE TRANSACTIONS ernors of the Mediterranean Control Association.
ON VEHICULAR TECHNOLOGY, the IEEE TRANSACTIONS ON INDUSTRY
APPLICATIONS, and the IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS.