Electrical Power and Energy Systems 96 (2018) 132–142

Contents lists available at ScienceDirect

Electrical Power and Energy Systems

journal homepage: www.elsevier.com/locate/ijepes

Adaptive robust AC optimal power flow considering load and wind power MARK
uncertainties
⁎
Ahmad Attarhaa, Nima Amjadya, , Antonio J. Conejob
a
Department of Electrical Engineering, Semnan University, Semnan, Iran
b
ISE and ECE Departments, The Ohio State University, Columbus, OH 43210, USA

A R T I C L E I N F O A B S T R A C T

Keywords: This paper proposes a tri-level adaptive robust AC optimal power flow (AR-ACOPF) model incorporating wind
ACOPF units. The uncertain wind power production as well as the system demand are characterized in terms of bounded
Adaptive robust optimization intervals forming a polyhedral uncertainty set. The proposed model is robust against any realization of uncertain
Budget of uncertainty parameters (i.e. wind power production and system demand) within the uncertainty set. Also the robustness of
Initialization
the solutions is controlled through a parameter denominated budget of uncertainty. Since the proposed tri-level
Wind power
model is not solvable via an off-the-shelf optimization package, a decomposition strategy relying on primal and
dual cuts is proposed to solve it. To reduce the computation burden of the proposed AR-ACOPF problem, an
effective initialization process is also presented. The proposed AR-ACOPF model and solution approach are
illustrated using the well-known IEEE 300-bus and Polish 2746-bus test systems.

1. Introduction knowledge of uncertain parameters, information that is rarely available

in practice [12].
The optimal power flow (OPF) is a common and frequently used To overcome SP difficulties, robust optimization [12] has recently
operational tool in power systems. Mainly, OPF aims at obtaining the attracted an increasing attention. However, previous robust OPF re-
optimal operating state of a power system based on a specific objective search works consider DC power flow equations instead of AC ones
function (commonly minimum cost) while both units and system con- [13]. In a DC representation, voltage magnitudes and reactive powers
straints are satisfied. The OPF problem is usually modeled as a mixed as well as their constraints are neglected. As a result, the obtained so-
integer and nonlinear (MINLP) optimization problem which is hard to lutions might be inaccurate or even insecure.
solve [1–3]. Recently, the increasing penetration of intermittent re- To overcome the aforementioned limitations, this paper proposes an
newable energy sources, such as wind power, has made this problem adaptive robust OPF model with AC constraints (AR-ACOPF), char-
even more challenging. As the OPF is a large-scale and complex opti- acterizing the uncertainties pertaining to wind power productions and
mization problem, some of the previous research works have neglected system demands in terms of bounded intervals rather than scenarios.
the uncertainty sources and developed deterministic models [1–5]. Also, the effect of contingencies is considered in this work using a pre-
However, deterministic OPF results may be non-optimal or even in- specified set of these contingencies. In comparison to the robust AC unit
secure when the power system deviates from the forecasted conditions commitment (ACUC) presented in [14], the following differences can be
such as forecasted wind power production or system demand. There- mentioned:
fore, to ensure security while considering uncertainties in the OPF
problem, stochastic programming (SP) has been widely used [4–11]. (1) In the ACUC work in [14], the second level considers DC power
Note, however, that SP characterizes the uncertain parameters by flow equations. However, in this OPF work, a new linearization
means of scenarios. Accordingly, the optimal solution of a SP problem is process is applied to the AC power flow equations to find the worst-
only guaranteed to be feasible for the scenarios considered in the case realization of the uncertain parameters. Unlike DC power flow
model. Moreover, the solution space of a SP problem depends on the equations, the linearized AC power flow equations consider reactive
number of uncertain parameters and the number of scenarios. Ac- power and voltage magnitude, which increases the accuracy of the
cordingly, SP faces two challenges: (1) becoming intractable for large- second level.
scale optimization problems, and (2) requiring a detailed distributional (2) In the ACUC work in [14], only the uncertainty of wind power is

⁎
Corresponding author.
E-mail address: [email protected] (N. Amjady).

http://dx.doi.org/10.1016/j.ijepes.2017.09.037
Received 14 July 2017; Received in revised form 2 September 2017; Accepted 26 September 2017
0142-0615/ © 2017 Elsevier Ltd. All rights reserved.
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

Nomenclature Variables

s
Indices Γik tap setting of the tap-changing transformer i -k ; Γik in-
dicates Γik at iteration s
s
n index of thermal units θikphase setting of the phase shifter i -k ; θikphase indicates θikphase at
i/k index of buses iteration s
s
s index of iterations QiC setting of the shunt capacitor/reactor of bus i ; QiC in-
l index of branches C
dicates Qi at iteration s
m index of break points Pn / Qn active/reactive power of thermal unit n ; Pns /Qns indicate
s′ index of out-of-sample scenarios Pn /Qn at iteration s
Pnm active power output of thermal unit n in segment m of the
Parameters piece-wise linear approximation of its cost function; Pns m
m
indicates Pn at iteration s
An , Bn , Cn the coefficients of the quadratic cost function pertaining to Pr nsU / Pr D up/down deployed reserve of thermal unit n in scenario s′
′ ns′
thermal unit n of the out-of-sample analysis
PLCnm the production cost of power segment Pnm Pisp spillage of the wind unit at bus i ; Pissp indicates Pisp at
Cwi the marginal cost of the wind unit at bus i iteration s ; P issp′ indicates Pisp in scenario s′ of the out-of-
C Sh cost of unserved load sample analysis
CnU / CnD cost of the up/down deployed reserve of thermal unit n PiSh unserved load at bus i ; PisSh indicates PiSh at iteration s ; P isSh′
Gik / Bik conductance/susceptance of line i -k indicates PiSh in scenario s′ of the out-of-sample analysis
μ budget of uncertainty ΔVk deviation of voltage magnitude from 1 p.u. at bus k ;ΔVks
max min
Γik /Γik maximum/minimum limits of Γik indicates ΔVk at iteration s
θikphase max / θikphase min maximum/minimum limits of θikphase Vk voltage magnitude of bus k
QiC max / QiC min maximum/minimum limits of QiC θk voltage angle at bus k ; θks indicates θk at iteration s
Pnmax / Pnmin maximum/minimum active power output of thermal unit BFl apparent power flow of branch l ; BFls indicates BFl at
n ; Pnmax min
,m / Pn,m are maximum/minimum limits for segment m iteration s
of thermal unit n lBFl linear approximation of the apparent power flow of
Qnmax / Qnmin maximum/minimum reactive power output of thermal branch l ; lBFls indicates lBFl at iteration s
unit n LB ACOPF lower bound of the objective function of the ACOPF pro-
RnU ,max / RnD,max maximum up/down reserve capacity of thermal unit blem
n LBLACOPF lower bound of the objective function of the LACOPF
Pwi forecasted wind power output of the wind unit at bus i problem
Pdi forecasted load demand of bus i UB ACOPF upper bound of the objective function of the ACOPF pro-
∼ ∼
Pwi/
 Pdi deviation of Pwi and Pwi from Pdi and Pwi , respectively blem
∼s ∼ s
Pwi / Pdi the worst case realization of Pwi/ Pdi at iteration s UBLACOPF upper bound of the objective function of the LACOPF
′ s′
Pwis / Pdi realizations of Pwi/ Pdi at scenario s′ problem
pfk power factor of the demand at bus k ρ,τ ,φ auxiliary continuous variable
ΔVkmax /ΔVkmin maximum/minimum value for ΔVk λikI , λikII , λikIII dual variables
BFlmax maximum capacity of branch l ΨI objective function of the first-level problem
I sufficiently large constant ΨII objective function of the second-level problem
ΨIII objective function of the third-level problem, where ΨsIII
Sets indicates ΨIII at iteration s
s′
Ψ OS objective function of scenario s′ pertaining to the out-of-
ω DC set of dual cuts generated by the third-level problem sample analysis
ω PC set of primal cuts generated by the second-level problem
ω ACOPF set of ACOPF constraints for continuous decision variables Uncertain parameters
ωLACOPF set of linear ACOPF (LACOPF) constraints for continuous
∼
decision variables Pwi uncertain wind power of the wind unit at bus i
∼
ωI−OPF set of OPF constraints for integer decision variables Pdi uncertain load demand of bus i
ωI set of buses
ωw set of wind units Vectors
ω NG set of thermal units
ω NGk set of thermal units connected to node k E, F vector of operation costs
ωL set of lines U vector of uncertain parameters
ωnBP set of break points for thermal unit n M vector of integer decision variables
ωs
′
set of out-of-sample scenarios N vector of continuous decision variables
ωUS uncertainty set

considered. However, in this OPF work both the uncertainties of delivery, a more accurate cost function is usually required. For this
load and wind power are taken into account, which makes the OPF reason, we consider piece-wise linear approximation of the quad-
solution immune against load forecast errors in addition to wind ratic cost function in the second level and the exact quadratic cost
power forecast errors. function in the third level of the proposed OPF model.
(3) The ACUC work in [14] considers linear approximation of the (4) Since the results obtained from a linearized AC power flow are close
quadratic cost function of units. However, as OPF is closer to power to the results of a nonlinear AC power flow, we have used an

133
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

initialization process for the third level in this work to enhance the is equal to the total number of uncertain parameters. If μ = 100%, all
convergence of the proposed OPF model instead of the con- uncertain parameters of the case study can deviate from their forecast
vexification used in the ACUC work in [14]. values and take their worst-case realizations, which leads to the most
(5) The effect of branch contingencies has been considered in this robust model. The above considerations have been modelled through
work, which has been neglected in the ACUC work in [14]. (1). The correlations between the uncertain parameters have not been
(6) The optimum settings of the proposed AR-OPF approach are de- considered in this study to focus on the proposed AR-ACOPF approach.
termined through a long-run out-of-sample analysis, which helps However, we can easily consider these correlations through, e.g., the k-
the operators make sound decisions about the robustness level and means clustering technique [17].
to avoid over/under-conservativeness. This analysis is not available It is straightforward to consider two separate budgets of uncertainty
in [14]. for the uncertainty sources pertaining to loads and wind power pro-
(7) The application of this work is on OPF, which has different decision ductions. However, such approach increases the adjustable settings of
variables and constraints with respect to the ACUC work in [14]. the proposed OPF tool and complicates its practical application. Note
that the uncertainties of both loads and wind power productions are
Therefore, the contributions of this paper are: forecast type uncertainties. Thus, one budget of uncertainty is con-
sidered for these uncertain parameters aiming at reducing the number
(1) Proposing an adaptive robust min-max-min model for the OPF of the adjustable settings of the proposed OPF tool and consequently its
problem immunized against demand and wind power uncertainty, complexity for decision makers.
which can be used for both ring and radial networks. The proposed AR-ACOPF model minimizes the objective function
(2) Contrary to previous robust optimization frameworks [13–16] over the operating variables while maximizes it over the uncertain
which incorporate DC power flow equations to find the worst case parameters, which is represented through a min-max-min model as
realizations, this paper considers a linearized AC power flow model, follows:
which includes constraints pertaining to both voltage magnitude
min max min F ′·N
and reactive power. Moreover, unlike previous robust OPF works in M ∈ ωI−OPFU ∈ ωUS N ∈ ωACOPF (M ,U ) (2)
the literature, quadratic cost functions are considered here.
In (2), the outermost minimization problem finds the optimal values
(3) Applying a tractable tri-level decomposition strategy via primal and
of integer decision variables including the transformer tap settings,
dual cuts to decompose the proposed adaptive robust min-max-min
phase shifter settings, and shunt capacitor/reactor settings. The middle
OPF problem into a mixed-integer linear minimization master
maximization problem finds the worst-case realization of the uncertain
problem, a linear maximization sub-problem and a non-linear
wind power and uncertain loads to immunize the proposed AR-ACOPF
minimization sub-problem.
model against these uncertainty sources. The innermost minimization
(4) Applying a novel initialization process to reduce the computation
problem finds the optimal values of the continuous decision variables
burden of the non-linear problem of the third level.
including the active power and voltage set points. Accordingly, the
proposed model (2) is tri-level and cannot be solved using an off-the-
The rest of this paper is organized as follows. In Section 2, the
shelf optimization package. To cope with this limitation, a decom-
framework of the proposed AR-ACOPF model is introduced. Section 3
position strategy is proposed in this paper, which breaks the whole
presents the details of the proposed tri-level min-max-min AR-ACOPF
problem into a linear master problem at the first level and two sub-
solution approach. In Section 4, the proposed approach is illustrated
problems at the second and third levels communicating with the master
using the IEEE 300-bus and Polish 2746-bus test systems under different
problem using primal and dual cuts, respectively. The proposed solu-
conditions. Section 5 concludes the paper.
tion approach can be described through the following steps:

2. Framework of the proposed AR-ACOPF model Step (1) The master problem at the first level of the proposed de-
composition strategy solves a mixed-integer linear OPF problem
To apply the robust optimization to an ACOPF problem, the un- together with primal and dual cuts to obtain optimal integer solu-
certain parameters are first modeled through bounded intervals within tions. The obtained optimal integer solutions are then fixed in the
a polyhedral uncertainty set as follows: second- and third-level problems.
∼ Step (2) Provided that the first- and second-level problems have not
⎧∼ ∼ Pwi−Pwi
Pwi, |Pdi−Pdi | ⩽ 
ωUS = |Pwi−Pwi | ⩽  Pdi, i ∈ ω I ; ∑ converged yet, a max-min OPF problem is solved at the second level
⎨ ∀ i ∈ωI

Pwi to find the worst case realization of the uncertain parameters and to
⎩
∼ create one primal cut to be sent to the first-level problem. It is
Pdi−Pdi ⎫
+ ⩽μ worthy to mention that the AC power flow equations at the second

Pdi ⎬
⎭ (1) level are linearized aiming at obtaining a linear max-min ACOPF
problem which enables us to dualize the min problem and recast it
In (1), the budget of uncertainty μ is used to control the robustness
as a tractable maximization problem. since the second-level problem
level of the proposed approach through limiting the size of the un-
is linear (due to the proposed linearization), its associated primal
certainty set. In other words, the budget of uncertainty is a setting,
cuts [18] are linear and so the problem of the first level keeps its
which limits and controls the number of the uncertain parameters that
linearity. Having the second-level problem solved, the convergence
can deviate from their forecast values. As μ increases/decreases, a
criterion between the first and second-level problems, i.e.
higher/lower number of uncertain parameters can deviate from their
|UBLACOPF −LBLACOPF | ⩽ ε , is evaluated. In case that the convergence
forecast values leading to a comparatively more/less robust model. The
criterion between the first and second-level problems is not satisfied,
minimum value for the budget of uncertainty is zero. When μ = 0 , no
the obtained worst case realizations are sent to the first level to add
uncertain parameter can deviate from its forecast value. Therefore, the
∼ ∼ the associated primal cut and the solution strategy goes back to Step
uncertain parameters Pwi and Pdi are equal to their forecasted values,
1. On the contrary, if the convergence criterion
i.e. Pwi and Pdi , respectively, which leads to a deterministic re-
|UBLACOPF −LBLACOPF | ⩽ ε is satisfied, the second level has converged.
presentation with no robustness. For μ > 0 , a robust model is obtained.
Subsequently, the optimal integer variables and the worst-case
Additionally, as μ increases, an increasing number of uncertain para-
realization of the uncertain parameters, obtained, respectively, from
meters can deviate from their forecast values resulting in an increas-
the first-level and second-level problems, are fixed within the third-
ingly robust model. The maximum value for the budget of uncertainty μ

134
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

level problem and the solution algorithm goes to Step 3. Since equations at the third-level problem. The final solution of the proposed
considering an AC network model at the third-level problem leads to OPF approach is obtained from the third level and thus it satisfies AC
nonlinearity, high computation burden, especially for a large-scale power flow constraints. Regarding the convergence of the iterative
power system, is expected. To overcome this issue and make the procedure between the first and second levels, since the second level is
third-level problem tractable, the obtained results from the second- linear and thus convex, the convergence is guaranteed [18]. Regarding
level problem (i.e. the one incorporating linearized AC power flow the convergence of the iterative procedure between the first and third
equations) are used as initial points for the third-level problem, levels, we have applied an effective initialization approach, which en-
which accelerates convergence. hances the convergence of this iterative procedure. Since the solution
Step (3) The third level solves a nonlinear ACOPF problem con- obtained using linearized AC power flow equations is pretty close to the
sidering a full AC network representation to obtain the optimal results obtained using nonlinear AC power flow equations, this can be a
values of the continuous decision variables and to generate dual good starting point for the third level. With the aid of the proposed
cuts. Since the third-level problem incorporates nonlinear AC power initialization approach, the iteration between the first and third levels
flow equations, sending primal cuts from the third level to the first has converged in all of our numerical simulations.
level leads to a MINLP first-level problem which is hard to solve.
Thus, to preserve the linearity of the first-level problem, the third- 3. Detailed formulation of the proposed model
level problem generates and sends dual cuts to the first level. Having
the convergence criterion between the first and third levels satisfied Based on the general framework introduced in the previous section,
(i.e. |UB ACOPF −LB ACOPF | ⩽ ε ), the proposed solution approach has the detailed formulations of the first-level, second-level and third-level
converged and an optimal solution is obtained. Otherwise, the as- problems are presented in the following.
sociated dual cut is added to the first-level problem and the solution
algorithm goes back to step 1. The flowchart of the proposed AR- 3.1. First level
ACOPF approach is given in Fig. 1.
The first-level problem optimizes the system operating state enfor-
Note that unlike previous robust OPF works, we use AC power flow cing primal cuts obtained in the second-level problem and dual cuts

Fig. 1. The flowchart of the proposed AR-ACOPF algorithm.

Start

First level: The OPF problem incorporating primal and dual cuts is solved
and the lower bounds of the second and third-level problems are updated

The integer variables are fixed in the second Primal cut is generated
level problem. and sent to the first level

NO
Second level: A max-min linear ACOPF is
solved using duality theory, and the worst
Convergence
case realization of the uncertain parameters
Check
and the upper bound of the second level
problem are obtained
YES

The integer variables obtained from the first-

level problem and the worst case realizations
Dual cut is generated
obtained from the second-level problem as
well as the initial values for the continuous and sent to the first level
variables are sent to the third-level problem.
NO

Third level: The nonlinear ACOPF problem is

Convergence
solved and the upper bound of the third-level
Check
problem is updated

Report the results

YES

Finish

135
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

∼
obtained in the third-level problem as follows: ∑ ∑ PLCnm·Pnm + ∑ Cwi·(Pwi−Pisp) + ∑ (C ShPiSh)
n ∈ ω NG m ∈ ωnBP i ∈ ωw i ∈ωI
minΨI = ρ (3a)
= C Operation: O ∈  (4b)
min max
Γik ⩽ Γik ⩽ Γik , i,k ∈ ω I (3b)

θikphase min ⩽ θikphase ⩽ θikphase max i,k ∈ ω I Pnmin ⩽ Pn(m = 1) ⩽ Pnmax min max
,(m = 1) : f n (m = 1) ,f n (m = 1) n ∈ ω NG (4c)
(3c)

QiC min ⩽ QiC ⩽ QiC max , i ∈ ω I (3d)

0 ⩽ Pnm ⩽ Pnmax max max
,m −Pn,(m − 1) : f nm n ∈ ω NG, m ∈ ωnBP , m ≠ 1 (4d)
∼s
τ⩾ ∑ ∑ PLCnm·Pns
m
+ ∑ Cwi·(Pwk −Pissp) + ∑ (C ShPisSh) s
n ∈ ω NG m ∈ ωnBP i ∈ ωw i ∈ωI
Qnmin ⩽ Qn ⩽ Qnmax : gnmin ⩾ 0,gnmax ⩾ 0 n ∈ ω NG (4e)
∈ ω PC (3e)

Pnmin ⩽ Pns
(m = 1)
⩽ Pnmax n ∈ ω NG, s ∈ ω PC ∼
,(m = 1) (3f) 0 ⩽ Pksp ⩽ Pwk : hkmax ⩾ 0 k ∈ ωw (4f)
m
0 ⩽ Pns ⩽ Pnmax max
,m −Pn,(m − 1) n ∈ ω NG, m ∈ ωnBP , m ≠ 1, s ∈ ω PC (3g) ∼
0 ⩽ PkSh ⩽ Pdk : Tkmax ⩾ 0 k ∈ ω I (4g)
Pnmax
,M = Pnmax n∈ ω NG (3h)

Qnmin ⩽ Qns ⩽ Qnmax n ∈ ω NG, s ∈ ω PC (3i) ΔVkmin ⩽ ΔVk ⩽ ΔVkmax : pkmin ⩾ 0,pkmax ⩾ 0 k ∈ ω I (4h)

∼s
0 ⩽ Pkssp ⩽ Pwk k ∈ ωw, s ∈ ω PC (3j)
−π ⩽ θk ⩽ π : gkmin ⩾ 0,gkmax ⩾ 0 k ∈ ω I (4i)
∼s
0 ⩽ PksSh ⩽ Pdk k ∈ ω I , s ∈ ω PC (3k)
∼ ∼
∼s ∼s ∑ ∑ Pnm + Pwk−Pksp + PkSh−Pdk− ∑ ((1 + ΔVk + ΔVi )(Gki )
∑ ∑ m
Pns + Pwk −Pkssp + PksSh−Pdk − ∑ ((1 + ΔVks + ΔVis )(Gki ) nω NGk m ∈ ωnBP i ∈ωI
n ∈ ω NGk m ∈ ωnBP i ∈ωI
+ Bki (θk−θi )) = 0: m k ∈  k ∈ ω I (4j)
+ Bki (θks−θis )) = 0 k ∈ ω I , s ∈ ω PC (3l)
∼s s ∼
∑ Qns + QkC + tan−1 (pfk ) PksSh−tan−1 (pfk )· Pdk ∑ Qn + QkC + tan−1 (pfk )·PkSh−tan−1 (pfk )· Pdk− ∑ (Gki (θk−θi )
n ∈ ω NGk nω NGk i ∈ωI

− ∑ ((Gki (θks−θis )−(1 + ΔVks + ΔVis ) Bki ) = 0 k ∈ ω I , s ∈ ω PC −(1 + ΔVk + ΔVi ) Bki ) = 0: n k ∈  k ∈ ω I (4k)
i ∈ωI

(3m)
−BFlmax ⩽ lBFl ⩽ BFlmax : rlmin ⩾ 0,rlmax ⩾ 0 l ∈ ω L (4l)
ΔVkmin ⩽ ΔVks ⩽ ΔVkmax , k∈ ωI , s∈ ω PC (3n)
The objective function (4a), to be minimized, is the operation cost
−BFlmax ⩽ lBFls ⩽ BFlmax , l ∈ ω L, s ∈ ω PC (3o) presented in (4b) in three separate terms. The first term is the pro-
duction cost of each thermal unit based on a piecewise linear approx-
−π ⩽ θks ⩽ π k ∈ ω I , s ∈ ω PC (3p) imation of its quadratic production cost function. The second and third
s s terms of the objective function are wind power production cost and
φ ⩾ Ψ sIII + λikIs (Γik−Γik
s
) + λikIIs (θikphase−θikphase ) + λikIIIs (QkC−QkC ) s ∈ ω DC
unserved load cost, respectively. Eqs. (4c) and (4d) in addition to (3h)
(3q) limit the active power of each segment of the piecewise linear cost
ρ ⩾ max(τ ,φ) function. Constraint (4e) bounds the reactive power output of each
(3r)
thermal unit. Constraints (4f) and (4g) limit the wind power spillage
Constraints (3b)–(3d) limit the transformer tap settings, phase and unserved load, respectively. Constraint (4h) limits the deviation of
shifter settings, and shunt capacitor/reactor settings, respectively. voltage magnitude from 1 p.u. and constraint (4i) limits the voltage
Constraints (3e)–(3p) are the primal cuts added from the second-level angle. Linearized active and reactive power balance constraints, i.e.
problem. Constraint (3q) is the dual cut added from the third-level linearized AC power flow equations, are enforced through (4j) and (4k),
problem. Since τ and φ represent the production costs of the second- respectively. The linearization of nonlinear AC power flow equations
level and third-level problems, respectively, adding both of them to the leading to (4j) and (4k) is described in the Appendix A. We assume that
first-level objective function may add extra weight to the production the power factor of wind units is 1 [15]. Constraint (4l) applies branch
cost and result in a non-optimal operating state. To overcome this issue flow limits. To include the impact of contingencies in the second-level
a new auxiliary variable ρ is introduced and constraint (3r) is added to formulation, constraints (4h), (4j), (4k) and (4l) should be added for the
the first-level problem. When the first-level problem is solved, post-contingent states of the considered contingencies. Additionally,
LBLACOPF = τ and LBACOPF = φ are updated and the obtained optimal ℧ = {O , f nmin
(m = 1)
max
, f nm , gnmin , gnmax , hkmax , Tkmax , pkmin , pkmax , qkmin , qkmax ,
values for the integer variables are fixed within the second-level pro- mk , n k , rlmin , rlmax } is the set of dual variables pertaining to constraints
s s
s
blem (i.e., Γik , θikphase , and QkC ). (4b)–(4l).
The max-min second-level problem (4a)–(4l) cannot be solved di-
3.2. Second level rectly. To make it solvable, we dualize the inner minimization problem
using the dual variables in ℧ . As a result, the max-min problem at the
Since the integer variables of the second-level problem are fixed to second level is recast into a single maximization problem as follows:
the obtained values of the first level, the second-level problem can be
written as a linear ACOPF problem as:
max min C Operation
∼ ∼ m sp Sh
Pwi,PdiPn ,Pi ,Pi (4a)

136
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

T1
   T2
   T9 = (Pdk + 
Pdk ·V +k −
Pdk ·V −k )·m k−(Pwk + 
Pwk ·U+k −
Pwk ·U−k )·m k
∼
maxΨII ≡ ∑ w Cwi· Pwi·O + ∑ NG ∑ Pnmin·f nmin
(m = 1) + Gki·m k (7d)
i∈ω n∈ω (m = 1)
T3
  
− ∑ NG ∑ (Pnmax max T10 = tan−1 (pfk )·(Pdk + 
Pdk ·V +k − Cs
Pdk ·V −k )·n k−Qk ·n k−Bki ·n k (7e)
n∈ω,(m = 1) )· f n (m = 1)
m=1

 T4
 T5 Nonlinearities stemmed from the products of binary and continuous
− ∑ NG ∑ (Pnmax max max
,m −Pn,(m − 1) )· f nm + ∑ NG Qnmin·gnmin variables in (7a)–(7e) can be easily recast into linear terms using the
n∈ω m≠1 n∈ω
T6 T7
big-M method as follows:
   
∼ T1 = Cwi·Pwi·O + 
Pwi·(O +−O −)
− ∑ NG Qnmax ·gnmax − ∑ w Pwk ·hkmax (8a)
n∈ω k∈ω
T8 T9
  T7 = Pwi·himax + 
Pwi·(himax +−himax −) (8b)
∼ ∼ ∼
− ∑ I Pdk ·Tkmax + ∑ I ∑ I (Pdk−Pwk + Gki)·mk
k∈ω k∈ω i∈ω

 T10
   T11 T8 = Pdi ·Timax + 
Pdi ·(Timax +−Timax −) (8c)

∼ s
+ ∑ I ∑ I (tan−1 (pfk )·Pdk−QkC −Bki )·n k + ∑ I ΔVkmin·pkmin
k∈ω i∈ω k∈ω T9 = Pdk ·m k + 
Pdk ·(m kI+−m kI−)−Pwk ·m k−
Pwk ·(m kII +−m kII −) + Gki·m k
T12

   (8d)
− ∑ I ΔVkmax ·pkmax
k∈ω
T10 = tan−1 (pfk )·(Pdk ·n k +  s
Pdk ·(nk+−nk−))−QkC ·n k−Bki ·n k (8e)
T13 T14
  
− ∑ I (qkmin + qkmax )·π − ∑ L (rlmin + rlmax )·BFlmax where O + and O − in (8a) are:
k∈ω l∈ω

(5a) −I·U+i ⩽ O + ⩽ I·U+i (8f)

−O ⩽ 1 (5b) −I·U−i ⩽ O − ⩽ I·U−i (8g)
f nmin
(m = 1)
max
−f nm + m k | nω NGk + PLCnm·O ⩽0 m∈ ωnBP , n∈ ω NG (5c) O −I·(1−U+i ) ⩽ O+ ⩽ O + I·(1−U+i ) (8h)

−hkmax −m k + Cwk ·O ⩽ 0, k ∈ ωw (5d) O −I·(1−U−i ) ⩽ O − ⩽ O + I·(1−U−i ) (8i)

gnmin−gnmax + n k | nω NGk ⩽ 0, n ∈ ω NG The I value of the big M method in (8f)–(8i) should be chosen
(5e)
lager than |O|, i.e.,  I⩾ |O|. The minimum value for this I can be
mk + tan−1 (pfk )·n k−Tkmax + C Sh·O ⩽ 0 k ∈ ω I (5f)  I= |O| which has been used in this paper for better convergence.
Considering  I⩾ |O|, and (8f)–(8i), the linear terms in (8a) work si-
−Gki (m k + m i ) + Bki (n k + n i ) + pkmin−pkmax ⩽ 0, k ,i ∈ ω I (5g) milarly as the nonlinear terms in (7a). Similar constraints can be con-
sidered for (8b)–(8e) to make the second-level problem linear. This
−Bki (m k−m i )−Gki (n k−n i ) + qkmin−qkmax ⩽ 0, k ,i ∈ ω I (5h) second-level maximization problem incorporating linear terms in
(8a)–(8e) and the additional constraints (8f)–(8i) is solved to obtain the
∼ ∼
rlmin−rlmax ⩽ 0, l ∈ ω L (5i) worst case realization of the uncertain parameters (i.e., Pwi and Pdi ) in
terms of U+i ,U−i ,V +i ,V −i as given in (6a) and (6b). Also, UBLACOPF = ΨII is
∼
Pwi−
Pwi ⩽ Pwi ⩽ Pwi + 
Pwi, i ∈ ωw (5j) updated. If |UBLACOPF −LBLACOPF | ⩽ ε , the second-level problem has
converged and the third-level problem should be solved. Otherwise, the
∼
Pdi−
Pdi ⩽ Pdi ⩽ Pdi + 
Pdi i ∈ ωI (5k) worst case realization of the uncertain parameters and their associated
primal cut given in (3e)–(3p) are sent to the first-level problem.
∼ ∼
Pwi−Pwi Pdi−Pdi
∑ 
+

⩽μ
∀ i ∈ωI Pwi Pdi (5l) 3.3. Third level
∼ ∼ max ∼ max
Note that the products of Pwi·O in T1, Pwk ·hk in T7, Pdk ·Tk in Having the loop between the first-level and second-level problems
∼ ∼ ∼
T8, Pdk ·m k and Pwk ·m k in T9, and Pdk ·n k in T10 made the second-level converged, the obtained values for the integer variables as well as the
problem (5a)–(5l) a bilinear maximization problem. However, since the worst case realization of the uncertain parameters (wind production
solution of the proposed model is on the extreme points of the poly- and demand), obtained from the first and second-level problems, re-
hedral uncertainty set, we only need to search the corners of the spectively, are fixed at the third-level ACOPF problem. Additionally,
polyhedron to find the optimal solution. Thus, new binary variables since the third level is a large-scale NLP problem (due to considering AC
∼ ∼
U+i ,U−i ,V +i and V −i can be introduced to model Pwi and Pdi as [16]: power flow equations), we use the obtained results for the continuous
∼
Pwi = Pwi + 
Pwi·U+i −
Pwi·U−i , U+i + U−i ⩽ 1. U+i , U−i ∈ {0,1}, i ∈ ωw variables Pnm , PiSh , Pisp , Qn , Vk , θk at the second level as initial points for
this third-level problem to accelerate the convergence of the third level.
(6a) The third-level problem can be mathematically modeled as follows:
∼
Pdi = Pdi + 
Pdi ·V +i −
Pdi ·V −i , V +i + V −i ⩽ 1. V +i , V −i ∈ {0,1}, i ∈ ω I minΨIII ≡ ∑ (An Pn2 + Bn Pn + Cn ) + ∑ ∼s
Cwi·(Pwi −Pisp)
(6b) n ∈ ω NG i ∈ωI

+ ∑ (C ShPiSh)
∑ (U+i + U−i + V +i + V −i ) ⩽μ i ∈ωI (9a)
i ∈ωI (6c)

Substituting (6a) and (6b) in T1, T7, T8, T9 and T10 leads to: Pnmin ⩽ Pn ⩽ Pnmax n∈ ω NG (9b)

T1 = Cwi·(Pwi + 
Pwi·U+i −
Pwi·U−i )·O (7a) Qnmin ⩽ Qn ⩽ Qnmax n ∈ ω NG (9c)
∼s
T7 = (Pwi + 
Pwi·U+i −
Pwi·U−i )·himax (7b) 0 ⩽ Pksp ⩽ Pwk k ∈ ωI (9d)
∼s
T8 = (Pdi + 
Pdi ·V +i −
Pdi ·V −i )·Timax (7c) 0 ⩽ PkSh ⩽ Pdk k ∈ ωI (9e)

137
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

Vkmin ⩽ Vk ⩽ Vkmax k ∈ ωI , (9f) solvable.

(2) As a result of the proposed linearization, the primal cuts that are
−π ⩽ θk ⩽ π k ∈ ωI (9g) sent from the second level to the first level are linear and thus the
∼s ∼s first level problem remains linear.
∑ Pn + Pwk + PkSh−Pksp−Pdk − ∑ Vk Vi (Gki cos(θk−θi )
nω NGk i ∈ωI
After merging the second-level problem and the dual problem of the
+ Bkisin(θk−θi )) = 0 k ∈ ω I (9h)
linearized third-level problem, a big-M linearization technique is used
∼s to make linear the bilinear terms in this single-level maximization
∑ Qn + QkC + tan−1 (pfk )·PkSh−tan−1 (pfk )· Pdk − ∑ Vk Vi (Gkisin(θk−θi )
nω NGk i ∈ωI problem. As a result of the proposed linearization, the second level
includes an MILP problem which can be easily solved using the CPLEX
−Bki cos(θk−θi )) = 0 k ∈ ω I (9i)
solver in GAMS [19].
|BFl | ⩽ BFlmax l ∈ ωL (9j) How to solve the third level of the tri-level min–max-min optimization
problem?
s
Γik = Γik : λikI (9k) The third-level problem includes exact AC power flow equations to
s obtain final dispatch decisions. Since the integer variables obtained in
θikphase = θikphase : λikII (9l) the first-level problem and the worst-case realization of uncertain
s parameters obtained in the second-level problem are fixed in this pro-
QkC = QkC : λikIII (9m)
blem, the third-level problem is an NLP problem which can be solved
The objective function (9a) represents the operation cost of both using the CONOPT solver in GAMS [19]. To enhance the computational
thermal and wind units as well as the cost of unserved load. Constraint efficiency of the third level, the obtained dispatch variable values at the
(9b) limits the active power output of each thermal unit. Constraints second level (using linearized AC power flow equations) are used as
(9c)–(9j) function similarly as (4e)–(4l). However, the third-level model initial values in the third-level problem. Moreover, since the third-level
uses the quadratic cost function instead of the piece-wise linear one and problem is nonlinear, dual cuts are sent from the third level to the first
non-linear AC power flow equations instead of linear ones in order to level aiming at maintaining the first level’s linearity.
increase the modeling accuracy. To include the impact of contingencies The equations solved at each level, the problem type of each level
in the third-level model, constraints (9f), (9h), (9i), and (9j) should be (i.e., MILP and NLP) as well as the associated solver used to solve the
added for the post-contingent states of the considered contingencies. problem are given in Table 1.
Constraints (9k)–(9m) fix the integer variables on the values obtained
from solving the first-level problem where λikI ,λikII, and λikIII denote the 4. Simulation results
dual variables pertaining to (9k), (9l) and (9m), respectively. When the
nonlinear ACOPF problem of the third level is solved, UB ACOPF = ΨIII is 4.1. Data
updated. Provided that |UB ACOPF −LB ACOPF | ⩽ ε , the optimal operating
state of the proposed AR-ACOPF problem is obtained and the proposed To illustrate the proposed AR-ACOPF solution approach, the IEEE
s
solution strategy terminates; otherwise, the values of ΨIII ,λikIs,λikIIs and
300-bus and Polish 2746 bus test systems are used. The required data
λikIIIs are sent to the first-level problem to generate the corresponding
was obtained from [1]. Table 2 briefly shows the characteristics of the
dual cut (3q).
test systems. Furthermore, the IEEE 300-bus test system is expanded
Note that due to considering slack variables (i.e., unserved loads) in
adding ten 200-MW wind farms located at nodes 30, 60, 90, 120, 150,
the equality constraints of each level (i.e., in (4j), (4k), (9h) and (9i)),
180, 210, 240, 260, and 300 and the Polish 2746-bus test system is
the problem will not be infeasible and as more cuts are added, more
expanded using thirty 200-MW wind farms located at
accurate solutions are obtained. Since these slack variables are added to
buses = 91 × n ,∀ n = 1,…,30 . These wind units are the same as those
the objective function with a high penalty cost, in the optimal solutions
located in the Northern flat lands of Texas [20] and the wind power
they generally become zero.
marginal cost is considered to be 5 $/MWh. Load shed cost is con-
In the following we separately explain how to solve each level
sidered to be 2000 $/MWh [21]. The proposed solution approach is
problem of the tri-level min-max-min optimization problem including
implemented within GAMS using CPLEX and CONOPT [19] on a laptop
AC power flow equations, which was formulated above.
computer with an Intel(R) Core(TM) i7 2.50 GHz CPU, and 8 GB of
How to solve the first level of the tri-level min–max-min optimization
memory.
problem?
The first-level problem aims at obtaining the integer variable values
pertaining to transformer tap settings, phase shifter settings, and shunt 4.2. Obtained results for the IEEE 300-bus test system
capacitor/reactor settings. It also includes the primal cuts added from
the second level and dual cuts added from the third level. Since both the The results obtained from the proposed AR-ACOPF approach for the
primal and dual cuts are linear, this first level includes an MILP problem IEEE 300-bus test system are illustrated in Fig. 2, considering different
which can easily be solved using the CPLEX solver in GAMS [19]. The values for 
Pdi , 
Pwi (including 5%, 10%, 15% and 20% of their nominal
obtained integer values are then fixed in the second and third level estimates) and μ = 0 , μ = 50 , μ = 100 , μ = 200 , and μ = 310 . The
problems. budget of uncertainty μ can vary between 0 and 310 (300 load buses
How to solve the second level of the tri-level min–max-min optimization + 10 wind units).
problem? As Fig. 2 illustrates, total costs for a specific deviation from forecast
The second level includes a max-min problem (second-level problem values (e.g., 
Pdi = 5%Pdi and  Pwi = 5%Pwi ) increase as μ increases. The
plus a linearized version of the third-level one) to obtain the worst-case
realizations of uncertain parameters. The integer variable values ob- Table 1
Equations, problem type, and solver of each level problem.
tained in the first level are fixed within the second level and a linear-
ization approach has been applied to the AC power flow equations of Level Equations Problem type Solver
the second level due to following reasons:
First (3a)–(3r) MILP CPLEX
Second (5a)–(5i), (6a)–(6c), (8a)–(8i) MILP CPLEX
(1) A linearization approach enables us to use duality theory to recast
Third (9a)–(9m) NLP CONOPT
the max-min problem into a single maximization one to make it

138
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

Table 2
Characteristics of the test systems.

Test Units Branches Tap changing Phase Capacitor/ Reference

system transformers shifters reactor of data
banks

IEEE 69 411 107 0 14 [1]

300-
bus
Polish 520 3514 172 2 1 [1]
274-
6-
bus

Fig. 4. Total costs for Polish 2746-bus test system for different deviations from nominal
estimates and different values of μ .

system is used. For this test case, the budget of uncertainty μ can vary
between 0 and 2776 (2746 load buses + 30 wind units). The results
obtained for the Polish 2746-bus test system are illustrated in Fig. 4,
considering different values for  Pdi , 
Pwi (including 5%, 10%, 15% and
20% of their nominal estimates) and μ = 0 , μ = 500 , μ = 1000 ,
μ = 2000 , and μ = 2776.
The same variation patterns of Figs. 2 and 3 are observed in Fig. 4.
In fact, as the worst-case realizations become worse, either in terms of μ
or deviations from nominal estimates, the total cost increases to achieve
Fig. 2. Total costs of IEEE 300-bus test system for different deviations from nominal
increasingly robust solutions. By considering contingencies, a similar
estimates and different values of μ .
result pattern of that in Fig. 3 is obtained for this test system, but for the
sake of conciseness these results are not shown here.
reason is that by increasing μ a higher number of uncertain parameters
can adopt their worst-case realizations resulting in a more robust yet
more expensive operating state. Furthermore, for a specific μ (other 4.4. Analyzing the effect of considering the AC power flow equations
than μ = 0 ), total cost increases as deviation from forecast values rises
from 5% to 20%. The reason is that the uncertain parameters can ex- To model the network more accurately, the proposed approach
perience a higher deviation from their forecasted values. In other considerers nonlinear AC power flow equations at the third-level pro-
words, the worst-case realization becomes worse. However, when μ = 0 blem. To evaluate the effect of different network models, three cases are
none of the uncertain parameters can deviate from their forecasted considered: the third-level problem incorporates DC power flow equa-
values (yielding the deterministic case) and thus increasing the devia- tions (Case A), linear AC power flow equations (Case B) and nonlinear
tion ranges Pdi and Pwi has no effect on the total cost. AC power flow equations (Case C is the proposed model). The obtained
To evaluate the effectiveness of the proposed AR-ACOPF approach costs and computation times for the IEEE 300-bus and Polish 2746-bus
in the presence of contingencies, we consider five single-contingencies, test systems with different values for μ at a deviation of 15% (i.e.
including the outage of transmission lines 109–110, 82–96, 69–77, 
Pdi = 15%Pdi and  Pwi = 15%Pwi ) are reported in Tables 3 and 4, re-
19–20, and 56–58 and report the obtained results in Fig. 3. The same spectively. From Tables 3 and 4, we observe that as increasingly rea-
variation patterns of Fig. 2 are observed in Fig. 3. However, the total listic network models are considered, both cost and computation
costs obtained for both deterministic ( μ = 0 ) and robust ( μ > 0 ) cases burden increase, since tighter constraints, increasingly limiting the so-
in Fig. 3 are higher than those reported in Fig. 2 due to contingencies. lution space, are imposed to the optimization problem. However, it can
be observed that the highest computation time of Table 4, which cor-
4.3. Obtained results for the Polish 2746-bus test system responds to the Polish 2746-bus test system with nonlinear AC power
flow equations (i.e., Case C) and the highest value of μ (i.e., 2776), is
To show the effectiveness of the proposed adaptive robust AR- only 271.02s (i.e. about 4.5 minutes), which is a reasonable
ACOPF approach on a larger test system, the Polish 2746-bust test
Table 3
Total cost obtained for Case A, Case B and Case C.

Test μ Total Cost (Thousand $) Cost Cost

system increment of increment of
Case A Case B Case C Case C with Case C with
respect to respect to
Case A (%) Case B (%)

IEEE 0 696.318 704.328 717.413 2.94 1.82

30- 310 857.328 869.149 881.239 2.71 1.37
0-
bus
Polish 0 1468.352 1493.547 1528.866 3.96 2.31
27- 2776 1783.586 1823.763 1867.599 4.50 2.35
46-
bus
Fig. 3. Total costs of IEEE 300-bus test system considering five single-contingencies.

139
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

Table 4 Table 5
Total computation time obtained for Case A, Case B and Case C. Total cost of out-of-sample analysis for Case A, Case B and Case C.

Test system μ Total time (s) 

(·) Cases Total out-of-sample cost (Thousand $)

Case A Case B Case C μ=0 μ = 50 μ = 100 μ = 200 μ = 310

IEEE 300-bus 0 4.02 6.23 45.05 5% A 39570.981 39242.146 38708.179 38260.679 38265.709
310 3.99 6.16 42.36 B 33683.224 32238.468 31497.849 30992.489 30973.685
Polish 2746-bus 0 10.54 19.02 268.48 C 28750.995 25123.445 23523.04 22684.655 22682.735
2776 10.07 20.98 271.02
10% A 39570.981 38472.816 37889.100 37178.679 37170.906
B 33683.224 31978.589 30947.991 29469.469 29453.341
C 28750.995 25904.965 24023.455 23058.68 23057.120
computation time for a large-scale OPF problem.
15% A 39570.981 38026.891 37624.322 37019.638 37005.957
B 33683.224 32117.903 31023.687 29987.389 29975.823
4.5. Evaluating the robustness of solutions using an ex-post out-of-sample C 28750.995 26687.74 24747.250 23768.584 23767.385
analysis 20% A 39570.981 37253.869 36957.359 36280.984 36263.385
B 33683.224 32879.733 32034.785 31087.392 31069.695
The effectiveness of the proposed AR-ACOPF approach is illustrated C 28750.995 27471.765 25705.080 24830.400 24829.580
using an ex-post out-of-sample analysis. One set of out-of-sample sce-
narios including wind power realizations and load realizations has been
obtained from the Electric Reliability Council of Texas (ERCOT) [22]. deviation ranges of uncertain parameters and the budget of uncertainty.
We have used 8760 hourly historical data of year 2014 for the out-of- The most appropriate values for these settings can be determined based
sample analysis. This analysis can be carried out as follows: on the lowest total cost obtained in the out-of-sample analysis. Note
that if the robustness level increases, the obtained total cost increases as
′ ′
s
Ψ OS = min ∑ (An Pn2 + Bn Pn + Cn ) + ∑ Cwi·(Pwis −P issp′ ) shown in Fig. 2 due to an increasingly robust solution. It is reasonable
n ∈ ω NG i ∈ ωw that if the model is increasingly robust and immunized against various
′ uncertainty realizations, the robustness cost increases and the objective
+ ∑ (CnU Pr ns
U −C D Pr D ) +
′ n ns′
∑ (C ShP isSh′ ) s′ ∈ ωs
n ∈ ω NG i ∈ωI (10a) function value increases accordingly. On the other hand, as a higher
robustness level is imposed, the model becomes robust against more
′ ′
0 ⩽ P issp′ ⩽ Pwis , i ∈ ωw, s′ ∈ ωs (10b) severe realizations of the uncertain parameters, which eliminates the
need for compensating power flow mismatches through load shedding
′
U ⩽ R U ,max
0 ⩽ Pr ns n ∈ ω NG , s′ ∈ ωs in these scenarios. This is the robustness worth. The results reported in
′ n (10c)
Table 5 illustrate the tradeoff between robustness cost and robustness
′ worth. For instance, as the robustness level increases from 5% to 10% in
0 ⩽ Pr nsD′ ⩽ RnD,max n ∈ ω NG , s′ ∈ ωs (10d)
case B, more uncertain parameters' realizations are included in the
s′ ′ bounded intervals and thus the proposed model is immunized against
0 ⩽ P isSh′ ⩽ Pdi i ∈ ω I , s′ ∈ ωs (10e) their extreme values. Note that the robustness worth becomes greater
than the robustness cost leading to a better result. However, by further
(9c), (9f)–(9j) (10f) increasing the robustness level of case B from a 10% bounded interval
In out-of-sample formulation (10a)–(10f), the OPF decision vari- to a 15% bounded interval, the model becomes over conservative,
ables Pn , Γik , θikphase , and QkC are fixed on the values obtained from cases which means that the robustness cost becomes greater than the ro-
A, B, or C. Also the uncertain parameters are fixed on the values of the bustness worth. The reason is that the most severe realizations have
s′ ′ already been covered in the 10% bounded interval and by further in-
out-of-sample scenarios (i.e., Pdi , Pwis ). As a result, the first term of the
objective function (10a) is a constant just added to obtain the total cost. creasing the robustness level we are immunizing the model against less
The second, third, and fourth terms are the cost of wind power gen- severe realizations. Thus, after-the-fact total cost of case B increases in
eration, up/down deployed reserves, and load shed, respectively. Wind the 15% bounded interval compared to 10% bounded interval as shown
power spillage, up/down spinning reserves, and load shed are limited in Table 5. In other words, μ = 310 and the deviation of 10% identify
through (10b)–(10e). For the results of the out-of-sample analysis re- the case for which the tradeoff between the robustness cost and the
ported in Table 5, the cost coefficients of reserve deployment are set robustness worth reaches the optimum equilibrium for case B. Similar
equal to the energy cost, which is in line with the other research works out-of-sample results have been obtained for the Polish 2746-bus test
that consider reserve cost, such as [6] and [10]. However, these cost system, but for the sake of conciseness these results are not reported
coefficients are just a part of the input data of the out-of-sample ana- here.
lysis and this analysis can easily work with different values for cost
coefficients of reserve deployment. 4.6. Analyzing the effect of piecewise linear cost function
Total cost obtained from the out-of-sample analysis of IEEE 300-bus
test system for Cases A to C in different conditions are reported in To analyze the effect of considering a piecewise linear cost function
Table 5. It can be observed that the proposed model (i.e., Case C) for all instead of linear one in the second level of the proposed AR-ACOPF, we
values of μ and for all values of 
Pdi , 
Pwi obtains the lowest operational introduce a new Case C′. This case is similar to Case C but considers a
cost due to its more accurate modeling of nodal power balance con- linear approximation of the cost function in the second level rather than
straints, which leads to lower required deployed reserves and load the piecewise linear one. Note that both cases C and C′ incorporate
shedding in different out-of-sample scenarios. Furthermore, for this test quadratic cost function and nonlinear AC power flow equations in the
case, the most appropriate settings of the uncertainty set for Case A, third level. The obtained costs and the computation time with μ = 50
Case B, and Case C are μ = 310 and (·) = 20%, 10%, and 5%, respectively, and μ = 100 at a deviation of 15% are reported in Table 6.
indicated by using bold fonts in Table 5. In addition to evaluating the From Table 6 it can be concluded that considering piecewise linear
long-term performance, other important advantage of the proposed out- approximation results in slightly higher total cost. The reason is that
of-sample analysis is determining the optimal settings of the un- using piecewise linear approximation increases the modeling accuracy
certainty set of each AR-ACOPF approach. These settings are the resulting in a worse “worst case realization” which makes the final

140
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

Table 6
Total cost and computation time for Case C and Case C′.

Test system μ Total cost (Thousand $) Out-of-sample cost (M$) Computation time (s)

Case C′ Case C Case C′ Case C Case C′ Case C

IEEE 300-bus 50 829.147 829.230 26.864 26.687 41.87 43.23

100 859.503 859.825 24.914 24.747 42.21 43.56

solutions more robust and flexible at the expense of a higher total cost. also show that considering the exact non-linear AC power flow equa-
A lower out-of-sample cost of Case C with respect to Case C′ illustrates tions may increase the objective function cost and computation burden
this aspect, i.e., a more robust behavior of Case C with respect to var- of the AR-ACOPF model, but improves its long-run performance eval-
ious realizations of the uncertain parameters. uated through an out-of-sample analysis. It is also pointed out that
Unlike previous OPF research works that consider piecewise linear considering more accurate objective function through piecewise linear
approximation of the cost function to solve the OPF problem in a more approximation in the second-level problem obtains comparatively
tractable way, in our work the piecewise linear approximation has been worse “worst case realizations” leading to increasingly robust operating
used to obtain more accurate worst case realizations (and not solving state at a higher cost. This higher robustness has been illustrated
the OPF problem) as the OPF is actually solved in the third level con- through out-of-sample analysis.
sidering exact nonlinear cost functions. Moreover, the power system operator using this model can consider
the uncertainty sources in a power system efficiently due to following
5. Conclusion reasons:

This paper presents a novel AR-ACOPF model formulated as a tri- (1) The written software package appears as a black box to the op-
level min-max-min optimization problem. Uncertainty sources include erators and they do not get involved in the modelling details.
system demand and wind power generation. The nonlinear generation (2) The reasonable computation time of the proposed OPF tool facil-
cost functions of the thermal units (leading to a more accurate model- itates its application in practice. The total computation time of the
ling of the production cost), and nonlinear AC power flow equations proposed solution approach for the IEEE 300-bus and Polish 2746-
(leading to a more accurate network modelling) are represented in bus test systems are about 45 s and 270 s, respectively, which are
detail. reasonable computation times for these large-scale ACOPF pro-
We observe that the robustness of the proposed model can be con- blems.
trolled thorough the budget of uncertainty and the uncertainty devia- (3) The required data of the proposed AR-ACOPF is analogous to the
tions. We point out that considering a higher budget of uncertainty or data required by commonly used methods in the industry, such as
considering a higher deviation range for the uncertain parameters leads deterministic models, which facilitates its utilization in industry.
to a more flexible operating state, but at a higher operation cost. We

Appendix A

Linearizing power flow equations is a frequently used approximation in power systems. Therefore, different linearization approximations have
been presented in the literature, e.g., in [23–25]. The second level of the proposed tri-level approach incorporates a max-min problem which is not
solvable via available software packages. Using a particular instance of the linear approximation family enables us to dualize the second-level
problem and recast it as a single maximization problem. In this appendix, we propose a simple yet efficient linearization approach to implement the
underlying idea behind the proposed AR-ACOPF approach. However, any other linear approximation approach can be used instead. Moreover, since
the final solutions are obtained from the third-level problem, where exact nonlinear AC power flow equations are considered, the obtained solutions
are sufficiently accurate.
Our proposed linearization approach for nonlinear AC power flow equations, which are given in (9h) and (9i), is introduced in this appendix. The
nonlinear parts of (9h) and (9i) are NP1 and NP2 , given in (A1) and (A2), respectively:

NP1 = ∑ Vk Vi (Gki cos(θk−θi ) + Bkisin(θk−θi ))

i ∈ωI (A1)

NP 2 = ∑ Vk Vi (Gkisin(θk−θi )−Bki cos(θk−θi ))

i ∈ωI (A2)
Since in a real-word power system, the bus voltages are usually close to 1 p.u., we can write the voltage of bus k as Vk = 1 + ΔVk , which ΔVk is a small
value (e.g., −0.05 ⩽ ΔVk ⩽ 0.05). As a result, the nonlinear term Vk Vi can be written as
Vk Vi = 1 + ΔVk + ΔVi + ΔVk ΔVi ≅ 1 + ΔVk + ΔVi (A3)
neglecting the product ΔVk ΔVi which becomes a very small value. Additionally, we use the first order approximation of Taylor Series for the
nonlinear sin(·) and cos(·) functions. Using this approximation and (A3), the nonlinear parts NP1 and NP2 can be written as (A4) and (A5):

NP1 = ∑ (1 + ΔVk + ΔVi )(Gki + Bki (θk−θi ))

i ∈ωI (A4)

NP 2 = ∑ (1 + ΔVk + ΔVi )(Gki (θk−θi )−Bki )

i ∈ωI (A5)
However, (A4) and (A5) are still nonlinear due to the product of ΔVk (or ΔVi ) and Δθki = θk−θi . However, difference between voltage angles of two
connected buses is often small (e.g., |Δθki | ⩽ π /12 ) to avoid branch overloads and decreasing angle stabilities' margins. Thus, the product of ΔVk (or
ΔVi ) and Δθki becomes very small. By neglecting these products, NP1 and NP2 become as:

141
A. Attarha et al. Electrical Power and Energy Systems 96 (2018) 132–142

NP1 = ∑ (Bki (θk−θi ) + Gki (1 + ΔVk + ΔVi ))

i ∈ωI (A6)

NP 2 = ∑ (Gki (θk−θi )−(1 + ΔVk + ΔVi ) Bki )

i ∈ωI (A7)
By substituting NP1 and NP2 of (A6) and (A7) in nonlinear AC power flow equations (9h) and (9i), linearized AC power flow equations (4j) and (4k)
are obtained. Linearized AC power flow equations (4j) and (4k) are more accurate than DC power flow equations, since DC power flow equations
only include the term ∑i ∈ ω I Bki (θk−θi ) of (A6). Additionally, (A7) and voltage and reactive power are not considered in the DC power flow equations.
Our linearization method, presented in (A1)–(A7), uses the first order approximation of the Taylor Series for the nonlinear sin(·) and cos(·)
functions. Moreover, the bilinear terms are approximated by linear ones in our method by neglecting small second-order terms. In [23], the same
approximation for the nonlinear sin(·) and cos(·) functions has been used. However, we expand the bilinear terms Vk Vi , while the bilinear terms
Vi (Vi −Vj ) are expanded in [23]. In the linear approximation method presented in [24,25], the first and second order approximations of Taylor series
have been used for the nonlinear sin(·) and cos(·) functions, respectively. Moreover, in the linear approximation method of [24,25], the product terms
are expanded in terms of Vi2 and θi2 , which are subsequently substituted by new auxiliary variables to obtain a linear approximation.

References [12] Bertsimas D, Litvinov E, Sun XA, Zhao J, Zheng T. Adaptive robust optimization for
the security constrained unit commitment problem. IEEE Trans Power Syst
2013;28(1):52–63.
[1] Attarha A, Amjady N. Solution of security constrained optimal power flow for large- [13] Ding T, Bie Z, Bai L, Li F. Adjustable robust optimal power flow with the price of
scale power systems by convex transformation techniques and Taylor series. IET robustness for large-scale power systems. IET Gener Transm Distrib
Gener Transm Distrib 2016;10(4):889–96. 2016;10(1):164–74.
[2] Abdi H, Derafshi Beigvand S, La Scala M. A review of optimal power flow studies [14] Amjady N, Dehghan S, Attarha A, Conejo AJ. Adaptive robust network-constrained
applied to smart grids and microgrids. Renew Sustain Energy Rev AC unit commitment. IEEE Trans Power Syst 2017;32(1):672–83.
2017;71(1):742–66. [15] Dehghan S, Amjady N, Conejo AJ. Reliability-constrained robust power system
[3] Hazra J, Sinha AK. A multi-objective optimal power flow using particle swarm expansion planning. IEEE Trans Power Syst 2016;31:2383–92.
optimization. Euro Trans Electr Power 2011;29(1):1028–45. [16] Jiang R, Wang J, Guan Y. Robust unit commitment with wind power and pumped
[4] Amjady N, Ansari M. Non-convex security constrained optimal power flow by a new storage hydro. IEEE Trans Power Syst 2012;27(2):800–10.
solution method composed of Benders decomposition and special ordered sets. Int [17] Baringo L, Conejo AJ. Correlated wind-power production and electric load scenarios
Trans Electr Energy Syst 2013;24(6):842–57. for investment decisions. Appl Energy 2013;101:475–82.
[5] Roy PK, Paul C. Optimal power flow using krill herd algorithm. Int Trans Electr [18] Zeng B, Zhao L. Solving two-stage robust optimization problems using a column-
Energy Syst 2015;25(8):1397–419. and-constraint generation method. Oper Res Lett 2013;41(5):457–61.
[6] Sharifzadeh H, Amjady N. Stochastic security-constrained optimal power flow in- [19] Generalized Algebraic Modeling Systems (GAMS). [Online], Available: < http://
corporating preventive and corrective actions. Int Trans Electr Energy Syst www.gams.com > .
2016;26(11):2337–52. [20] Alliance for Sustainable Energy, Department of Energy (DOE), National Renewable
[7] Hamon C, Perninge M, Söder L. A stochastic optimal power flow problem with Energy Laboratory (NREL), SAM version 2014.1.14.
stability constraints—Part I: approximating the stability boundary. IEEE Trans [21] Conejo AJ, Carrión M, Morales JM. Decision making under uncertainty in electricity
Power Syst 2013;28(2):1839–48. markets. Springer; 2010.
[8] Hamon C, Perninge M, Söder L. A stochastic optimal power flow problem with [22] http://www.ercot.com.
stability constraints—Part II: approximating the stability boundary. IEEE Trans [23] Yang J, Zhang N, Kang C, Xia Q. A state-independent linear power flow model with
Power Syst 2013;28(2):1849–57. accurate estimation of voltage magnitude. IEEE Trans Power Syst
[9] Liang J, Molina DD, Venayagamoorthy GK, Harley RG. Two-level dynamic sto- 2017;32:3607–17.
chastic optimal power flow control for power systems with intermittent renewable [24] Yang Z, Zhong H, Xia Q, Kang C. A novel network model for optimal power flow
generation. IEEE Trans Power Syst 2013;28(3):2670–8. with reactive power and network losses. Electr Power Syst Res 2017;144:63–71.
[10] Summers T, Warrington J, Morari M, Lygeros J. Stochastic optimal power flow [25] Yang Z, Zhong H, Bose A, Zheng T, Xia Q, Kang C. A linearized OPF model with
based on convex approximations of chance constraints. In: Power systems compu- reactive power and voltage magnitude: a pathway to improve the MW-Only DC
tation conference (PSCC); 2014. OPF. IEEE Trans Power Syst 2017. http://dx.doi.org/10.1109/TPWRS.2017.
[11] Yong T, Lasseter RH. Stochastic optimal power flow: formulation and solution. In: 2718551.
Power engineering society summer meeting, 2000, vol. 1. IEEE; 2000. p. 237–42.

142