Received September 19, 2016, accepted October 20, 2016, date of publication October 27, 2016, date of current

version November 18, 2016.

Digital Object Identifier 10.1109/ACCESS.2016.2621338

Robust Optimization for Dynamic Economic

Dispatch Under Wind Power Uncertainty With
Different Levels of Uncertainty Budget
HUIFENG ZHANG, DONG YUE, AND XIANGPENG XIE
Institute of Advanced Technology, Nanjing University of Posts and Telecommunications, Nanjing, China
Corresponding author: H. Zhang (zhanghuifeng_520@163.com)
This work was supported in part by the National Natural Fund under Grant 61503199 and Grant 61503193, in part by the National Natural
Science Key Fund under Grant 61533010, in part by the Ph.D. Programs Foundation of the Ministry of Education of China under
Grant 20110142110036, in part by the Jiangsu Province Natural Science Fund under Grant BK20150853, in part by the Jiangsu
Province Post-Doctoral Fund under Grant 1501042C, in part by the Jiangsu Province High School Natural Science Fund under
Grant 15KJB120009, in part by NUPTSF under Grant NY214206, and in part by the Open Fund under Grant XJKY14019.

ABSTRACT Since dynamic economic dispatch with wind power uncertainty poses great challenges for
power system operation due to its non-linear and uncertain characteristics, this paper proposes a robust
optimization model with different levels of uncertainty budget. The dynamic economic dispatch problem
is converted into the robust optimization model with an uncertainty budget, which transforms the non-
deterministic model into a deterministic optimization problem. Differential evolution is improved by the
sequential quadratic programming method and utilized to solve the robust optimization model. Due to the
complex-coupled constraints among thermal units, several constraint-handling procedures are proposed to
address those constraint limits, which have a significant impact on the efficiency of the whole optimization.
The robust optimization model with an adjustable uncertainty budget is implemented in two test systems.
The results obtained for the first test system prove the efficiency of differential evolution-based sequential
quadratic programming and the constraint-handling procedures; the performance of the second test system
reveals that the robust optimization method with different levels of uncertainty budget provides a promising
method for solving the dynamic economic dispatch problem with wind power uncertainty.

INDEX TERMS Dynamic economic dispatch, wind power, robust optimization, differential evolution,
sequential quadratic programming, uncertainty budget.

I. INTRODUCTION Reference [3] takes the stochastic approach, using sce-

To decrease fuel costs and carbon emissions, much effort has
been made to integrate renewable generation (especially wind
power generation) into the power grid. High penetration of
wind power generation can result in significant operational
challenges due to its uncertainty [1]. Since a large deviation
may exist between actual output and forecasted output, the
dispatch problem for wind power generation cannot be man-
aged by traditional deterministic optimization methods. Cur-
rently, the forecasting of wind power generation on temporal
and spatial scales remains elusive [2]; therefore, stochas-
tic optimization methods are naturally utilized to solve the
economic problem associated with wind power. Generally,
uncertainty optimization can be broadly divided into three
categories: 1) Stochastic programming, 2) Fuzzy program-
ming, and 3) Robust optimization.
narios to describe the uncertainty of wind power output
and generating a scheduling solution for each scenario.
Reference [4] utilizes chance-constrained programming to
model the impacts of fluctuating wind output on the whole
operation system. Reference [5] presents an interesting risk-
based scheduling strategy using a fuzzy method to model
the uncertainty of wind power generation. However, the
probability distribution function obtained from limited sam-
ples in stochastic programming can cause large deviations
in real-world applications [6]–[9]. The personal experiences
of decision makers in the membership function can be
subjective, which does not ensure that accurate values are
obtained [10], [11]. Recently, the robust optimization (RO)
method has received growing attention for being able to han-
dle uncertainty optimization problems. All of the uncertain

2169-3536
2016 IEEE. Translations and content mining are permitted for academic research only.
VOLUME 4, 2016 Personal use is also permitted, but republication/redistribution requires IEEE permission. 7633
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

parameters are predetermined in the uncertain set, which
gives RO two advantages [12]: 1) The uncertain bound is
easier to construct than the probability distribution function
and membership function and 2) The typical extreme cases
can be included in the uncertainty set. Reference [13] presents
a robust optimization approach that considers the uncertainty
of wind power output and demand response. Reference [14]
employs a two-stage robust optimization method with bi-level
mixed-integer optimization to consider the uncertainty under
nodal net injection. Reference [15] proposes, with some suc-
cess, a robust interval method for look-ahead dispatch with
wind power.
To improve conventional robust optimization, the
adjustable uncertainty budget (AUB) is utilized to min-
imize the conservation problem of RO; the uncertain
domain can be controlled by adjusting the uncertainty
budget to flexibly obtain the optimal solution [14].
Reference [16] proposes a flexible robust optimization with
an adjustable uncertainty budget dispatch model for hybrid
wind/photovoltaic/hydro/thermal power systems, which is
built to achieve coordination between reliability and econ-
omy. In the flexible robust optimization, the adjustable
uncertainty budget is proposed to cover the shortcoming of
conventional robust optimization.
In this paper, the robust optimization with AUB in liter-
ature [16] is utilized to convert non-deterministic dynamic
economic dispatch (DED) model into an easily-solving deter-
ministic model. On the basis of literature [16], it is assumed
that forecast errors of wind power follow the normal distri-
bution, uncertainty domain of wind power generation can be
divided into several levels, in which probability is calculated
with probability density function (PDF). Since probability
density function of wind power generation is taken into
consideration, the conservation of robust optimization can
be decrease in comparison to robust optimization method
especially when it is assumed that the uncertainty of wind
power generation is completely unknown. Several typical lev-
els of wind power uncertainty are further analyzed for certain
extreme cases in real-world application, the optimal schemes
can be taken as representative optimal schemes when
corresponding extreme cases occur. Moreover, differential
evolution with sequential quadratic programming (DE-SQP)
is utilized to optimize the deterministic optimization model
under different uncertainty levels, and some constraint-
handling procedures are proposed to assign the deviation
increment of system load balance to thermal units with wind
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch
robust optimization model. The implementation in two test
systems and conclusions are presented in Section 6 and
Section 7, respectively.
II. PROBLEM FORMULATION
With the integration of wind power generation, DED is highly
random in comparison to conventional DED of thermal units.
The uncertainty budget for wind power output is utilized to
obtain the extreme values in extreme cases, which can provide
the maximum boundary for a robust optimization model with
an adjustable uncertainty budget. The goal of DED with an
adjustable uncertainty budget is to optimize the economic
cost of thermal units while considering system load balance,
output limits, ramp rate limits, and uncertainty budget limits.
A. UNCERTAINTY BUDGET OF INTERMITTENT
POWER OUTPUT
For further analysis on the uncertainty of wind power, the
wind power output can be described as follows:
(
Pwj,t = P̄wj,t + P̃wj,t
max (1)
P̃min
wj,t ≤ P̃wj,t ≤ P̃wj,t
Literature [16] proposes an adjustable robust interval,
which can be described as:
wj,t , P̄wj,t + rwj,t P̃wj,t ]
Pwj,t ∈ [P̄wj,t + rwj,t P̃min max
(2)
where rwj,t ∈ [0, 1] is the dispatch interval coefficient of
the jth wind farm at tth time interval, P̄wj,t is the actual
output at the jth wind farm at tth time interval, P̃wj,t is
the forecasted error at the jth wind farm at tth time inter-
wj,t , P̃wj,t are the minimum and maximum output
val, and P̃min max
fluctuations, respectively, at the jth wind farm at tth time
interval. The dispatch interval coefficient reflects the toler-
able range for the uncertainty of wind power output, a large
dispatch interval coefficient means that the obtained robust
dispatch scheme has large tolerated range in the economic
problem [17]. When rwj,t = 0, the robust optimization of
dynamic dispatch with intermittent uncertainty reduces to a
deterministic optimization dispatch.
According to the scenario-based method in reference [18],
nine levels of deviation between actual output and forecast
output are taken to describe the adjustable dispatch interval
coefficient rwj,t , which satisfies rwj,t ∈ {0, 1/4, 1/2, 3/4, 1}
and P̃min max
wj,t = −P̃wj,t . The deviation between actual output and
power uncertainty. Furthermore, the optimal results of two forecasted output can be dividedinto nine levels,  it generates
test systems by the proposed method reveal that DE-SQP can nine possible output intervals if Pwj,t − P̄wj,t  ≤ δε (δε > 0
optimize the DED problem well, and DED with wind power is a small number) is also taken as an interval, which can
uncertainty can be properly tackled by robust optimization be seen in Figure 1. It is assumed that forecast errors of
with different levels of uncertainty budget. wind power follow the normal distribution, the size
 of each
interval (expect the interval when Pwj,t − P̄wj,t  ≤ δε ) is

The remainder of this paper is organized as follows:
Section 2 presents the problem of DED with wind power,
Section 3 converts the DED problem into a robust opti-
mization model, Section 4 shows the DE-SQP method, and
Section 5 describes the implementation of DE-SQP in the
7634
δ = 1/8∗(P̃max min
wj,t − P̃wj,t ), which is the standard deviation of the
forecast error, and each interval l is associated with the prob-
ability pl [19], [20]. The uncertainty budget for wind power
generation at the tth time period is defined as the maximum
VOLUME 4, 2016
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch

(1) Power balance constraint

Nc
X Nw
X
Pci,t + Pwj,t = PD,t + Ploss,t (5)
i=1 j=1

where Pwj,t is the output of the jth wind farm at tth time
period, PD,t , Ploss,t are, respectively, the system load require-
ment and transmission loss at tth time period, and Nw is
the number of wind farms. The transmission loss between
different power generators is considered; it can be generally
expressed as follows:
c +Nw NX
NX c +Nw c +Nw
NX
FIGURE 1. The division of output deviation between actual output and Ploss,t = <it Bij <jt + B0i <it + B00 (6)
forecast output.
i=1 j=1 i=1

Here, the output of different power generators is labeled

limit on the summation of the dispatch interval coefficient of <it , and Bij , B0i , B00 are the coefficients of transmission loss.
each wind power output, which can be described as follows: (2) Output limits
Nw
(
X Pci,min ≤ Pci,t ≤ Pci,max
rwj,t ≤ 1t (3) max (7)
P̄wj,t + rwj,t P̃min
wj,t ≤ Pwj,t ≤ P̄wj,t + rwj,t P̃wj,t
j=1

It can be seen that the uncertainty budget 1t is in the where Pci,max is the maximum output of the ith thermal unit.
range [0, Nw ]. The uncertainty budget 1t can be adjusted (3) Spinning reserve constraint
according to the robustness of dynamic economic dispatch. Nc
X
The allowed range of wind power generation can be enlarged (Pci,max − Pci,t ) ≥ PD,t ∗ L% (8)
when the uncertainty budget increases and the dispatch prob- i=1
lem is reduced to a deterministic optimization problem, which where L% is the proportion expressed in the range [0, 1].
means that only the expected values of outputs from wind For the stability of a hybrid energy resource system, some
power are considered. output needs to be reserved for special cases. Here, a part
of the thermal power is reserved based on the propor-
B. ECONOMIC OBJECTIVE tion of the system load requirement, which is described in
Since wind power generally has a relatively low cost of equation (8).
power generation, the cost of the thermal-wind system can be (4) Ramp rate limits
represented by the fuel cost of the thermal units, which can
be expressed by a quadratic function. Moreover, the valve- DRci ≤ Pci,t − Pci,t−1 ≤ URci (9)
point effect can also be taken into consideration: due to the where DRci , URci are, respectively, the down-ramp and up-
sudden opening of the intake valve of a steam turbine, the ramp rates of the ith thermal unit.
fuel cost of a thermal unit can be described as the summation (5) Uncertainty budget limits
of a quadratic function and a sinusoidal function. The total
Nw
cost of thermal units can be expressed as follows: X
rwj,t ≤ 1t (10)
Nc
T X
X j=1
f1 = (ai P2ci,t + bi Pci,t + ci
t=1 i=1
As requirements vary, different uncertainty budgets are
 
+ di sin(hi (Pci,t − Pci,min )))
where T is the total time period, Nc is the number of thermal
units, Pci,t is the output of the ith thermal unit at tth time step,
Pci,min is the minimum output of the ith thermal unit, and ai ,
bi , ci , di , ei are the cost coefficients of the ith thermal unit.
C. CONSTRAINTS
The main goal of dynamic economic dispatch of thermal units
with wind power uncertainty is to minimize the economic cost
with the valve-point effect, while considering the system load
balance constraint, output limits, the spinning reserve con-
straint, ramp rate limits and the uncertainty budget constraint.
VOLUME 4, 2016
(4) required to decrease the conservatism of robust optimization,
which is mainly satisfied by adjusting the dispatch interval
coefficient rwj,t , which divides the uncertainty domain into
different levels.
III. ROBUST OPTIMIZATION MODEL WITH
ADJUSTABLE UNCERTAINTY BUDGET
Combined with robustness and an adjustable robust interval,
the uncertainty constraints can be approximated by deter-
ministic constraints in extreme cases. Using the uncertainty
budget and dispatch interval coefficient of intermittent power
output in Section 2.1, the economic dispatch model can be
converted into the robust optimization model [16]. To obtain
7635
 H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch

the extreme value of deviation between the actual output and The inequality of total thermal power output can also be
the forecast output, linear duality theory is utilized with the obtained:
Lagrange function.
Nw Nc
X Nc
X Nw
X
P
According to linear duality theory, max P̃wj,t = Pci,t ≤ [ Pci,max + L% · Ploss,t − L% · ( P̄wj,t
j=1 i=1 i=1 j=1
Nw
P
− min(− P̃wj,t ), and the Lagrange function of Nw
X Nw
X
j=1 + λ+
t rwj,t P̃min −
wj,t −λt rwj,t P̃max
wj,t )]/(1+L%)
Nw
j=1 j=1
P
min(− P̃wj,t ) can be obtained with relaxing the constraint
j=1 (15)
limits of wind power output:
Then, the robust optimal model with an adjustable uncer-
Nw
X tainty budget can be transformed into the deterministic model
L( P̃wj,t , λt , λ+
t , λt )
−
as follows:
j=1
T P Nc

Nw Nw Nw Nw
(ai P2ci,t + bi Pci,t + ci
P
X X X X 
 min f =
P̃wj,t +λ+ rwj,t P̃min

=− P̃wj,t +λt t ( P̃wj,t − wj,t )



 t=1
 i=1 
j=1 j=1 j=1 j=1



 + di sin(hi (Pci,t − Pci,min )))

Nw Nw


 PNc PNc
Pci,t ≤ [ Pci,max + L% · Ploss,t
X X
+ λ− rwj,t P̃max

t ( wj,t − P̃wj,t ) (11)


i=1 i=1



j=1 j=1 
 Nw Nw
− L% · ( P̄wj,t + λ+ rwj,t P̃min

 P P

 t wj,t
With the Karush-Kuhn-Tucker (KKT) optimal condition, j=1 j=1



Nw

it can be simplified as follows:

λ −P
rwj,t P̃max

− wj,t )]/(1 + L%)

t



  j=1
Nw Nw Nw 
 Nc Nw
P̃wj,t = min(λ+ rwj,t P̃min −λ− rwj,t P̃max
P P P P
max wj,t )
 

 t wj,t t  Pci,t + P Pwj,t = PD,t + Ploss,t


 j=1 j=1 j=1 (16)
 i=1 j=1
s.t. λt + λ+ t − 1 − λt = 0
−
  NcP +Nw NcP +Nw NcP+Nw
(12) 

 Ploss,t = <it Bij <jt + B0i <it + B00
λ +
 


 t ≥ 0 


 i=1 j=1 i=1
 
λ− ≥ 0

Pci,min ≤ Pci,t ≤ Pci,max
 

t 





 DRci ≤ Pci,t − Pci,t−1 ≤ URci
Simultaneously, by substituting equation (5) into


λt + λ+ t − 1 − λt = 0
−


equation (8), the maximum value of total wind power output 


can be obtained: λ+

t ≥0




λ−

t ≥0
Nw


X 

max Pwj,t

Nw


 rwj,t ≤ 1t

 P
j=1 
Nc Nc j=1
P P
Pci,max + L% · Ploss,t − (1 + L%) Pci,t
i=1 i=1 According to the above optimization model, the uncertain
≤ (13) domain of wind power output can be expanded by adjusting
L%
the dispatch interval coefficient rwj,t , which leads to adaptive
By replacing Pwj,t with equation (1) and equation (2), the optimization on the robustness of the optimal solution with
extreme value of the deviation between forecast output and an adjustable uncertainty budget.
actual output can be obtained:
Nw IV. SEQUENTIAL QUADRATIC PROGRAMMING-BASED
DIFFERENTIAL EVOLUTION METHOD
X
max P̃wj,t
j=1
To properly solve the robust optimization model, an improved
Nw
differential evolution (DE) with sequential quadratic pro-
≤
X
P̄wj,t gramming (SQP) is proposed. The SQP is embedded into
the mutation of DE, which can provide the optimal search
j=1
direction for DE. In comparison to conventional DE, DE with
Nc Nc
P
Pci,max + L% · Ploss,t − (1 + L%)
P
Pci,t SQP has more powerful search abilities, which increase its
i=1 i=1 efficiency for solving the robust optimization model with an
+ (14)
L% adaptive uncertainty budget.
7636 VOLUME 4, 2016
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch

A. BACKGROUND OF DIFFERENTIAL EVOLUTION optimal solution. Generally, the optimization problem can be
DE is well-known as a simple yet powerful evolutionary opti- described as follows:
mization method; it is widely used in real-world applications 
due to its low number of parameters in comparison to other min

x
f (x)
evolutionary algorithms [21]. DE consists of three operations:  gi (x) = 0, i = 1, 2, . . . me (20)
gi (x) ≤ 0, i = me + 1, me + 2, . . . m

mutation, crossover and selection. During mutation, a new
individual is generated for the next generation from three
where me is the number of equality constraints, and m is
distinct individuals, which are selected from the evolutionary
the total number of constraints. The objective function f (x)
population. The mutation operator can be generally described
is the economic cost objective in equation (16) and gi (x)
as follows [22]:
represents the equality constraints and inequality constraints
vk+1 = xr1
k k
+ F(xr2 k
− xr3 ) (17) in equation (16). A quadratic program is solved to obtain
the search direction at each iteration to update the control
where vk+1 is the trial parameter generated at the k + 1th variables; the quadratic program can be described as follows:
generation, and F ∈ [0, 2) is the scaling parameter of the 
k when F = 0. T T
mutation operator. vk+1 converges to xr1 min[∇f (xk ) + 1/2dk Hk dk

In the crossover operator, candidate individual uk+1
i,j is gen- gi (xk ) + [∇gi (xk )]T = 0 i = 1, 2, . . . , me (21)
erated with the trial parameter, which is generally described i = me + 1, . . . , m
 T
gi (xk ) + [∇gi (xk )] ≤ 0

as follows:
( where Hk is the Hessian matrix of the Lagrangian function
k+1 vk+1 , if (rand() ≤ pc ) or j = random(1, n) L(x, λ) and dk is the direction basis at the kth iteration.
ui,j = i,j k+1
xi,j , otherwise Generally, Hk can be approximated by Bk using the quasi-
(18) Newton method.
q qT (B S )(B S )T

where rand() represents a random number in [0, 1], pc is the Bk+1 = Bk + k k − k k k k

qTk Sk SkT Bk Sk

crossover probability, vk+1

i,j represents the jth variable of vec-


tor vik+1 , and random() generates an integer between 1 and n. Sk = xk+1 − xk (22)
qk = ∇L(xk+1 , λk+1 ) − ∇L(xk , λk+1 )


The selection operator chooses the better individual between 

the candidate individual uk+1 and xik . A new individual of the L(x, λ) = F(x) + β T g(x)

i
next generation can be generated with the selection operator:
( where β is the vector of the Lagrangian operator and ∇ is the
k+1 uk+1 , uk+1 is better than xik gradient operator. According to the above steps for solving
xi = ik i (19) the QP sub-problem, the search direction dk can be obtained
xi , otherwise
for the new iteration:
All optimization problems can be optimized with the above
= xik + αk dk
 k+1
xi
three differential operators by DE regardless of whether the


 m
L(x, β, ρ) = f (x) − βi (gi (x) − si )
 P
optimization problem is non-convex, high-dimensional or

i=1 (23)
multi-modal. However, conventional differential evolution 
 m
+ 12 ρi (gi (x) − si )2
P
generally optimizes with a fixed scaling parameter; it can-


not adaptively adjust the optimization process, which means i=1
that conventional DE cannot properly guide the differential where αk is the step length at the kth iteration, which can
evolution and is apt to fall into local optima. be calculated with an augmented Lagrangian merit function
L(x, λ, ρ) by considerable reduction, ρi is a non-negative
B. SEQUENTIAL QUADRATIC PROGRAMMING-BASED penalty parameter, and si is a slack variable (si = 0 for i =
DIFFERENTIAL EVOLUTION 1, 2, . . . , me , and si ≥ 0 for i = me + 1, me + 2, . . . , m). The
To improve on conventional differential evolution, the SQP provides a local minimum for the optimization problem.
sequential quadratic programming (SQP) method is utilized It can be embedded into the mutation operator of DE, which
to provide the optimal search direction for the mutation oper- can improve the convergence ability for nonlinear program-
ator. SQP is widely known as one of the best non-linear pro- ming. This paper replaces equation (17) with equation (23),
gramming methods for constrained optimization problems. and the iteration process can also be taken into the population
In comparison to other nonlinear programming methods, SQP evolution, which improves the efficiency of SQP.
performs with more efficiency, accuracy and a higher percent-
age of successful solutions for a large number of test prob- V. IMPLEMENTATION OF THE PROPOSED DE-SQP
lems [23]. In SQP, the Hessian of the Lagrangian function IN THE ROBUST OPTIMIZATION MODEL WITH AN
is utilized for approximation with the BFGS quasi-Newton ADAPTIVE UNCERTAINTY BUDGET
updating method, generating the quadratic programming sub- Since the robust optimization model with adaptive uncer-
problem, which can obtain the optimal search direction to the tainty is a high-dimension and complex-constraint problem,

VOLUME 4, 2016 7637

 H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch

λ0 λ+ λ−
 
Pc1,0 Pc2,0 ··· PcNc ,0 0 0 rw1,0 rw2,0 ··· rwNw ,0
 Pc1,1 Pc2,1 ··· PcNc ,1 λ1 λ+1 λ−1 rw1,1 rw2,1 ··· rwNw ,1 
X = .. .. .. .. .. .. .. .. .. .. .. (24)
 
. . . . . . . . . . .

 
Pc1,T −1 Pc2,T −1 ··· PcNc ,T −1 λT −1 λ+
T −1 λ−
T −1 rw1,T −1 rw2,T −1 ··· rwNw ,T −1

proper implementation primarily depends on the constraint-

handling technique during the optimization process. The
proposed algorithm optimizes the DED problem by prun-
ing the infeasible regions and promoting the promising
regions. Here, the implementation strategy focuses on the
solution encoding strategy, population initialization strategy
and constraint-handling strategy.

A. SOLUTION ENCODING STRATEGY

Due to the high-dimensional and complex-constrained char-
acteristics of the DED problem, a set of thermal outputs and
some other parameters are used as the decision variables.
For T scheduling periods on the schedule horizon, there are
T schedule periods of thermal output Pci,t in Nc thermal FIGURE 2. Output details of 10 thermal units with transmission loss.

units, T schedule periods of parameters λt , λ+ t , λt , and

−

T schedule periods of the dispatch interval coefficient rwjt in

Nw wind farms, which can be encoded as (24), as shown at defined as follows:
the top of this page Y XNc Nw
X
= Pci,t + Pwj,t − Lload,t − Lloss,t (27)
B. POPULATION INITIALIZATION AND i=1 j=1
VIOLATION FORCE STRATEGY According to the expression for Lloss,t presented in equa-
Before the evolutionary population proceeds in differential tion (6), the variation of the deviation can be found as follows:
evolution, the individuals in the evolutionary population need c +Nw
NX c +Nw NX
NX c +Nw
to be initialized in the feasible area. Generally, the initial-
Y
1 = (1 − B0i )1<it − 2 Bij <it 1<jt
ization of the evolutionary population can be carried out as i=1 i=1 j=1
follows: (28)
xi = li + U (0, 1) · (ui − li ) (25) Assuming that all of the thermal output is adjusted with the
where li , ui are the lower and upper boundary of the ith same increment of deviation, or ℵ = 1Pit = 1Pjt (i 6 = j), the
decision variable, respectively. In the differential evolution, following can be obtained:
1
Q
individual values cannot avoid violating the permitted con-
straint limits; therefore, some measures need to be taken to ℵ= (29)
+Nw
NcP NcP+Nw NcP
+Nw
force those infeasible values into the feasible domain. The (1 − B0i ) − 2 Bij <it
i=1 i=1 j=1
violation force strategy is as follows:
 The above variation of deviation is used to adjust the out-
0
li , if xi < li
 put of the thermal unit with the coupled constraint-handling
xi = xi , if li ≤ xi ≤ ui (26) technique in [24], which handles the system load balance with
ui , if xi > ui

 coarse adjustments and a fine tuning method and defines the
total violation to control the constraint limits for screening
C. CONSTRAINT-HANDLING TECHNIQUE feasible individuals during the population evolution.
FOR DED PROBLEM
Since DED has various complex-coupled constraints, the D. IMPLEMENTATION FLOWCHART FOR ROBUST
constraint-handling technique can affect the DED problem- OPTIMIZATION OF THE DED PROBLEM WITH AN
solving efficiency to a great extent. The system load balance ADJUSTABLE UNCERTAINTY BUDGET
constraint can be taken as the most important constraint in Since robust optimization for DED with an adjustable uncer-
solving the DED problem. To address this constraint properly, tainty budget is a complex coupled optimization problem,
a constraint handling technique is proposed to decrease the the procedures for solving the robust optimization problem
deviation of system load balance constraint, which can be need to be properly followed, and the whole optimization

7638 VOLUME 4, 2016

H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch
TABLE 1. Comparison with alternatives in literature [25] and literature [26].
procedure must be substantial and intact, with all possible
cases being considered in the implementation process. The
steps in the optimization model can be taken as follows:
Step 1: The value of the dispatch interval coefficient rwj,t
is taken from the set {0, 1/4, 1/2, 3/4, 1} in ascending order,
Nw
rwj,t ≤ 1t . After setting the current
P
which needs to satisfy
j=1
order number I = 0, go to Step 2;
Step 2: Initialize all of the individuals in the evolutionary
population and parameters. If the value of each individual is
in the feasible domain of equation (25), the maximum value
of uncertainty budget 1t is given. Set the generation number
g = 0 and go to Step 3;
Step 3: Apply the DE-SQP method to the evolutionary
population; the information of each individual can be updated
as the evolution proceeds. In each generation, if the value of
any individual violates the feasible domain, the force viola-
tion strategy is applied to force the individual into the feasible
area, which is introduced in Section 5.2. The procedure goes
to Step 4;
Step 4: Use the constraint-handling technique on the pop-
ulation evolution of current generation: if the violation error
1p,t < εp , the individual can be taken as a feasible individual.
Calculate the economic cost of all of the thermal units, select
the best individual as the current optimal individual Xg∗ , and
go to Step 5;
Step 5: If g < gmax , then g = g + 1, go to Step 2;
Otherwise, the obtained individual Xg∗ is taken as the optimal
individual at the I th order, store it in the optimal set, and go
to Step 6;
Step 6: If I < Imax , then I = I +1, go to Step 1; Otherwise,
select the optimal individual from the optimal set according
to the required extreme cases, and output the optimal scheme
of the selected optimal individual.
VI. CASE STUDY
To verify the efficiency of robust optimization on DED,
two test systems are presented: 1) Test system 1 consists
of 10 thermal units without wind power generation; the
valve-point effect and transmission loss are considered;
2) Test system 2 consists of 10 thermal units and 4 wind
farms; the valve-point effect and transmission loss are con-
sidered. The DED is first converted into a robust optimiza-
tion problem and is then optimized with DE-SQP using the
robust optimization model while dealing with various differ-
ent equality constraints and inequality constraints.
VOLUME 4, 2016
A. TEST SYSTEM 1
Test system 1 utilizes the conventional DED model for
10 thermal units over 24 time periods at hourly intervals,
and the valve-point effect and transmission loss among those
thermal units are properly taken into consideration. This test
system mainly demonstrates the efficiency of DE-SQP with
several constraint- handling procedures, and the obtained
results are used for comparison with other alternatives. All the
details about thermal units, output limits, system load balance
and transmission loss can be found in [25].
The population size is set to 100, the mutation probability
is set to 0.4, the maximum generation number is 400, and the
permitted total violation is set to 0.1. The thermal output of
each thermal unit at each time period is taken as the decision
variable, DE-SQP is utilized to optimize the DED problem
with transmission loss, and the constraint-handling method
for transmission loss is provided in literature [25]. Here,
the obtained results are compared with other alternatives in
literature [26], which can be seen in Table 1.
In comparison with other alternatives, it can be seen that
the DE with SQP generates the minimum fuel cost within
a shorter CPU time. Evolutionary programming (EP) has
the highest fuel cost and takes the most time, while Parti-
cle swarm optimization (PSO) and Artificial immune sys-
tem (AIS) in literature [26] have relative high fuel costs
with high time consumption. Non-dominated sorting genetic
algorithm-II (NSGA-II) in literature [25] optimizes the fuel
cost and emission rate under a different scaling factor. The
economic load scheduling (ELD) result has a relatively low
fuel cost in comparison with EP, PSO and AIS in [26]. Since
the DE-SQP method is based on DE, DE is also utilized for
comparison with DE-SQP. The obtained fuel cost is lower
than that of DE yet consumes more time, which is mainly
caused by the iteration procedures in SQP.
For further analysis of the optimal scheme by DE-SQP, the
output details of the 10 thermal units are shown in Figure 2.
Each colored bar represents the output of a different ther-
mal unit at each time period. The outputs of Units #6-#10
are relatively low in comparison to Units #1-#5 due to
their low output capacity. The output of each thermal unit
at each time period properly satisfies the output limits,
ramp rate limits, and system load balance constraint, which
has been properly handled. From the time point of view,
the total output of the 10 thermal units properly satisfies
the system load requirements, which are low in the time
intervals 01:00-10:00, 15:00-18:00 and 22:00-24:00.
7639

TABLE 2. Transmission losses obtained by DE-SQP for DED.
FIGURE 3. Wind power output of wind farm #1 with different dispatch
interval coefficients.
FIGURE 4. Wind power output of wind farm #2 with different dispatch
interval coefficients.
Since transmission loss among thermal units is taken into
consideration, transmission loss can also be taken as an
important criterion for the efficiency of solving DED. All of
the transmission losses are presented in Table 2.
It can be seen that transmission loss changes with the
system load requirement: if the system load is large at a
time period, the transmission loss can also be relatively large.
However, the output does not exceed 5% of the system load
at each time period, which means that the system load bal-
ance constraint has been properly handled with the proposed
constraint-handling procedures. According to the above anal-
ysis on the DED, it can be found that the proposed DE-DQP
is efficient for the DED optimization problem and that the
7640
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch
FIGURE 5. Wind power output of wind farm #3 with different dispatch
interval coefficients.
FIGURE 6. Wind power output of wind farm #4 with different dispatch
interval coefficients.
constraint-handling procedures can handle the output limits,
ramp rate limits and system load limits properly. Further-
more, the proposed DE-SQP and constraint-handling tech-
nique show a promising approach for solving DED with wind
power generation, which provides great impetus for solving
the robust optimization model with an adjustable uncertainty
budget.
B. TEST SYSTEM 2
Four wind farms are added to the 10 thermal units to verify the
robust optimization with uncertainty budget. Nine dispatch
interval coefficients are utilized to divide the uncertainty
of wind power generation. In this test system, the value
VOLUME 4, 2016
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch

TABLE 3. Scheduling results with nine dispatch interval coefficients.

TABLE 4. The output (MW) details of 10 thermal units in extreme case 1.

of dispatch interval coefficient rwjt is taken from the set power generation are shown in Figures 3-6. It can be seen
{0, 1/4, 1/2, 3/4, 1}. Based on the robust optimization model that the nine levels of wind power generation cannot overlap
with adjustable uncertainty budget, the nine levels of wind with each other, which divides the uncertainty domain into

VOLUME 4, 2016 7641


TABLE 5. The output (MW) details of 10 thermal units in deterministic optimization.
eight intervals. Combined with the wind power output in
Figures 3-6, they have similar variation tendencies at several
time periods: wind power output falls until 6:00 in the morn-
ing, rises at 10:00 and 16:00, and falls again until 20:00 at
night. The upper and lower levels represent two extreme cases
in the feasible domain: wind power output at its maximum
and wind power output at its minimum.
Based on different uncertainty budget requirements, differ-
ent intervals can be obtained by adjusting the dispatch interval
coefficient. After solving the robust model with the DE-SQP
method, the obtained results for different levels are presented
in Table 3. The fuel cost, transmission loss, uncertainty bud-
get and total output violation are listed, with nearly all the data
in ascending order. When wind power generation is small, the
output of the thermal units is reduced, which causes further
reduction in the fuel cost. Otherwise, the fuel cost increases.
The total output violation is set to 1.0 MW, the violation of
different levels do not exceed 1.0 MW, which also means
that all the individuals in the evolutionary population are fea-
sible individuals, and the constraint-handling technique has
been properly implemented in the robust optimization model.
7642
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch
For further analysis on the efficiency of robust optimization
with an uncertainty budget, three typical optimal schemes
have been selected based on their dispatch interval coeffi-
cients. The first typical scheme is at the lower bound of
deviation between the forecast output and actual output, the
second typical scheme represents the deterministic scheme
when the forecast output is equal to the actual power output,
and the third typical scheme is at the upper bound of deviation
between the forecast output and actual output.
The results of the three typical cases are shown in Table.4,
Table.5 and Table.6. It can be seen in these tables that the
thermal output has similar variation tendencies during the
whole period, which means that the optimal DED scheme has
been properly obtained under all of the different cases. The
first and third typical schemes can be considered the optimal
schemes under extreme cases: extreme case 1 and extreme
case 2. In extreme case 1, the first optimal scheme can be
considered the optimal scheme for the worst case in which
actual output is smaller than forecast output, and it avoids all
of the possible risks caused by the actual output being smaller
than the forecast output. Similarly, the third optimal scheme
VOLUME 4, 2016
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch
TABLE 6. The output (MW) details of 10 thermal units in extreme case 2.
can be considered the robust optimal solution for the worst
case in which actual power output is larger than actual output
and avoids all of the possible risks in this situation.
VII. CONCLUSION
When wind power is integrated into the DED model, the opti-
mization of the economic problem becomes a great challenge
due to the uncertainty of wind power. To solve this opti-
mization problem, this paper proposes a robust optimization
method with an adjustable uncertainty budget with different
levels. When the DED model with wind power is converted
into a robust optimization model with an uncertainty budget,
the nondeterministic optimization problem becomes a deter-
ministic problem, which can be more conveniently optimized.
The uncertainty domain is divided into nine levels of extreme
cases, and the optimal scheme can be selected according
to the extent of risk in real-world applications. Moreover, a
constraint-handling technique is proposed to handle various
constraints during the optimization process and ensure the
feasibility of individuals in the evolutionary population. After
implementing the proposed method in two test systems, the
VOLUME 4, 2016
obtained results are better than other alternatives, and the
robust optimal scheme avoids different degrees of possible
risks, which also verifies the efficiency of the robust opti-
mization with an uncertainty budget.
REFERENCES
[1] L. Xie et al., ‘‘Wind integration in power systems: Operational challenges
and possible solutions,’’ Proc. IEEE, vol. 99, no. 1, pp. 214–232, Jan. 2011.
[2] K. Porter and J. Rogers, ‘‘Survey of variable generation forecast-
ing in the west,’’ Nat. Renew. Energy Lab, Golden, CO, USA,
Tech. Rep. SR-5500-54457, Apr. 2012.
[3] B. Bahmani-Firouzi, E. Farjah, and R. Azizipanah-Abarghooee, ‘‘An
efficient scenario-based and fuzzy self-adaptive learning particle swarm
optimization approach for dynamic economic emission dispatch consid-
ering load and wind power uncertainties,’’ Energy, vol. 50, pp. 232–244,
Feb. 2013.
[4] X. Yuan, B. Ji, S. Zhang, H. Tian, and Z. Chen, ‘‘An improved artificial
physical optimization algorithm for dynamic dispatch of generators with
valve-point effects and wind power,’’ Energy Convers. Manage., vol. 82,
pp. 92–105, Jun. 2014.
[5] M. R. Haghifam, H. Falaghi, and O. P. Malik, ‘‘Risk-based distributed
generation placement,’’ IET Generat. Transmiss. Distrib., vol. 2, no. 2,
pp. 252–260, 2008.
[6] X. Fei, L. Zhang, and D. James, ‘‘Risk-based security and economy
tradeoff analysis for real-time operation,’’ IEEE Trans. Power Syst., vol. 22,
no. 4, pp. 2287–2288, Nov. 2007.
7643

[7] K. Wang, Z. Ouyang, R. Krishnan, L. Shu, and L. He, ‘‘A game theory-
based energy management system using price elasticity for smart grids,’’
IEEE Trans. Ind. Informat., vol. 11, no. 6, pp. 1607–1616, Dec. 2015.
[8] K. Wang, H. Gao, X. Xu, J. Jiang, and D. Yue, ‘‘An energy-efficient
reliable data transmission scheme for complex environmental monitoring
in underwater acoustic sensor networks,’’ IEEE Sensors J., vol. 16, no. 11,
pp. 4051–4062, Nov. 2016.
[9] H. Jiang, K. Wang, Y. Wang, M. Gao, and Y. Zhang, ‘‘Energy big
data: A survey,’’ IEEE Access, vol. 4, pp. 3844–3861, Aug. 2016,
doi: 10.1109/ACCESS.2016.2580581.
[10] R. H. Liang and J. H. Liao, ‘‘A fuzzy-optimization approach for generation
scheduling with wind and solar energy systems,’’ IEEE Trans. Power Syst.,
vol. 22, no. 4, pp. 1665–1674, Nov. 2007.
[11] Y.-K. Liu, ‘‘Convergent results about the use of fuzzy simulation in
fuzzy optimization problems,’’ IEEE Trans. Fuzzy Syst., vol. 14, no. 2,
pp. 295–304, Apr. 2006.
[12] X. Bai and W. Qiao, ‘‘Robust optimization for bidirectional dispatch
coordination of large-scale V2G,’’ IEEE Trans. Smart Grid, vol. 6, no. 4,
pp. 1944–1954, Jul. 2015.
[13] C. Zhao, J. Wang, J.-P. Watson, and Y. Guan, ‘‘Multi-stage robust unit
commitment considering wind and demand response uncertainties,’’ IEEE
Trans. Power Syst., vol. 28, no. 3, pp. 2708–2717, Aug. 2013.
[14] D. Bertsimas, E. Litvinov, X. A. Sun, J. Zhao, and T. Zheng, ‘‘Adaptive
robust optimization for the security constrained unit commitment prob-
lem,’’ IEEE Trans. Power Syst., vol. 28, no. 1, pp. 52–63, Feb. 2013.
[15] W. Wu, J. Chen, B. Zhang, and H. Sun, ‘‘A robust wind power optimization
method for look-ahead power dispatch,’’ IEEE Trans. Sust. Energ., vol. 5,
no. 2, pp. 507–515, Apr. 2014.
[16] C. Peng, P. Xie, L. Pan, and R. Yu, ‘‘Flexible robust optimization dispatch
for hybrid wind/photovoltaic/hydro/thermal power system,’’ IEEE Trans.
Smart Grid, vol. 7, no. 2, pp. 751–762, Mar. 2016.
[17] K. Wang, L. Zhuo, Y. Shao, D. Yue, and K. F. Tsang, ‘‘Towards dis-
tributed data processing on intelligent leakpoints prediction in petro-
chemical industries,’’ IEEE Trans. Ind. Informat., to be published,
doi: 10.1109/TII.2016.2537788.
[18] J. Aghaei, T. Niknam, R. Azizipanah-Abarghooee, and J. M. Arroyo,
‘‘Scenario-based dynamic economic emission dispatch considering load
and wind power uncertainties,’’ Int. J. Elect. Power Energy Syst., vol. 47,
pp. 351–367, May 2013.
[19] K. Wang, Y. Shao, L. Shu, Y. Zhang, and C. Zhu, ‘‘Mobile big data fault-
tolerant processing for eHealth networks,’’ IEEE Netw., vol. 30, no. 1,
pp. 1–7, Feb. 2016.
[20] K. Wang, Y. Shao, L. Shu, G. Han, and C. Zhu, ‘‘LDPA: A local data
processing architecture in ambient assisted living communications,’’ IEEE
Commun. Mag., vol. 53, no. 1, pp. 56–63, Jan. 2015.
[21] R. Storn and K. Price, ‘‘Differential evolution—A simple and efficient
adaptive scheme for global optimization over continuous spaces,’’ Int.
Comput. Sci. Inst., Univ. California, Berkeley, Berkeley, CA, USA, Tech.
Rep. TR-95-012, 1995.
[22] L. H. Wua and Y. N. Wanga, ‘‘Environmental/economic power dispatch
problem using multi-objective differential evolution algorithm,’’ Elect.
Power Syst. Res., vol. 80, no. 9, pp. 1171–1181, 2010.
[23] A. M. Elaiw, X. Xia, and A. M. Shehata, ‘‘Hybrid DE-SQP and hybrid
PSO-SQP methods for solving dynamic economic emission dispatch
problem with valve-point effects,’’ Elect. Power Syst. Res., vol. 103,
pp. 192–200, Oct. 2013.
7644
H. Zhang et al.: Robust Optimization for Dynamic Economic Dispatch
[24] H. Zhang, D. Yue, X. Xie, S. Hu, and S. Weng, ‘‘Multi-elite guide hybrid
differential evolution with simulated annealing technique for dynamic
economic emission dispatch,’’ Appl. Soft Comput., vol. 34, pp. 312–323,
Sep. 2015.
[25] M. Basu, ‘‘Dynamic economic emission dispatch using nondominated
sorting genetic algorithm-II,’’ Int. J. Elect. Power Energy Syst., vol. 30,
NO. 2, pp. 140–149, 2008.
[26] M. Basu, ‘‘Artificial immune system for dynamic economic dispatch,’’ Int.
J. Elet. Power Energy Syst., vol. 33, no. 1, pp. 131–136, 2011.
HUIFENG ZHANG received the M.S. and Ph.D.
degrees from the Huazhong University of Sci-
ence and Technology, Wuhan, China, in 2009 and
2013, respectively. He is currently with the Insti-
tute of Advanced Technology, Nanjing Univer-
sity of Posts and Telecommunications, Nanjing,
China. His current research interest includes elec-
trical power management and optimal scheduling,
optimal operation of power system, and multi-
objective optimization.
DONG YUE received the Ph.D. degree from
the South China University of Technology,
Guangzhou, China, in 1995. He is currently a Pro-
fessor and the Dean with the Institute of Advanced
Technology, Nanjing University of Posts and
Telecommunications, Nanjing, China. He is also
a Changjiang Professor with the Department of
Control Science and Engineering, Huazhong Uni-
versity of Science and Technology, Wuhan, China.
He has authored over 100 papers in international
journals, domestic journals, and international conferences. His current
research interests include the analysis and synthesis of networked control
systems, multiagent systems, optimal control of power systems, and Internet
of Things. He is an Associate Editor of the IEEE Control Systems Society
Conference Editorial Board. He is an Associate Editor of the International
Journal of Systems Science.
XIANGPENG XIE received the B.S. and Ph.D.
degrees in engineering from Northeastern Univer-
sity, Shenyang, China, in 2004 and 2010, respec-
tively. From 2012 to 2014, he was a Post-Doctoral
Fellow with the Department of Control Science
and Engineering, Huazhong University of Sci-
ence and Technology, Wuhan, China. He is cur-
rently an Associate Professor with the Institute
of Advanced Technology, Nanjing University of
Posts and Telecommunications, Nanjing, China.
His current research interests include fuzzy modeling and control synthesis,
state estimations, optimization in process industries, and intelligent opti-
mization algorithms.
VOLUME 4, 2016