IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO.

6, NOVEMBER 2014 2775

A Novel Affine Arithmetic Method to Solve Optimal

Power Flow Problems With Uncertainties
Mehrdad Pirnia, Student Member, IEEE, Claudio A. Cañizares, Fellow, IEEE,
Kankar Bhattacharya, Senior Member, IEEE, and Alfredo Vaccaro, Senior Member, IEEE

Abstract—An affine arithmetic (AA) method is proposed in this 2003, and the additional wind capacity has not contributed to
paper to solve the optimal power flow (OPF) problem with un- the peak load carrying capability, as wind blows more during
certain generation sources. In the AA-based OPF problem, all the off-peak periods [4].
state and control variables are treated in affine form, comprising
OPF has been used for determining adequate generation
a center value and the corresponding noise magnitudes, to repre-
sent forecast, model error, and other sources of uncertainty without reserves in power systems dispatch [5]. Methods such as the
the need to assume a probability density function (pdf). The pro- Monte Carlo simulation (MCS) and interval arithmetic (IA)
posed AA-based OPF problem is used to determine the operating have been proposed in the literature to model and analyze
margins of the thermal generators in systems with uncertain wind uncertainties arising from load and generation, in OPF prob-
and solar generation dispatch. The AA-based approach is bench- lems. Although the MCS method is popular for its simple
marked against Monte Carlo simulation (MCS) intervals in order implementation, it is computationally expensive; on the other
to determine its effectiveness. The proposed technique is tested and
hand, analytical methods suffer from lack of accuracy and
demonstrated on the IEEE 30-bus system and also a real 1211-bus
European system. mathematical complexity [6]. Therefore, other methods have
been suggested in the literature as alternatives to analytical and
Index Terms—Affine arithmetic, generation uncertainty, in-
numerical approaches.
terval analysis, optimal power flow.
Probabilistic approaches are suggested to solve the OPF prob-
lems considering uncertainties in power system operations. One
I. INTRODUCTION of the first attempts in solving the probabilistic OPF (P-OPF)
problem is reported in [7], where the multivariate Gram-Char-

T HE increased focus on renewable generation has brought

forth several concerns pertaining to planning and opera-
tion of modern power systems, because of their inherent char-
lier method is employed to model the probability density func-
tion (pdf) of uncertain variables. In [8], the error between the
forecasted and the actual demand is assumed to have a Gaussian
acteristic of intermittency. In order to fulfill the requirements of probability function. The model is then solved using MCS, and
the evolving smart power grid, the intermittency need be taken the statistical characteristics (e.g., mean and variance) of the
into account in traditional unit commitment (UC), economic system variables, i.e., active power generation and system losses
load dispatch (ELD), optimal power flow (OPF) models, as well are calculated, so that the dispatcher is able to allocate enough
as in long term planning models of transmission and generation spinning reserve capacity for a given time interval. A different
systems. approach based on a sensitivity analysis technique is proposed
Obtaining the margins of operations for thermal generators in in [9], wherein operating constraint violations and their proba-
the presence of uncertain generation (e.g., wind and solar) and bilities are determined for the whole planning horizon, and then
load [1] helps significantly in determining the required reserve an iterative approach is used to adjust the control variables while
capacities, so that the system operates reliably and economi- satisfying their limits.
cally. System operators provide such ancillary services to main- In order to improve the computational efficiency of OPF so-
tain system reliability [2], [3]. The Electric Reliability Council lution methods, analytical methods have been proposed. For in-
of Texas (ERCOT) has faced resource adequacy issues, since stance, in [10] the Cumulant method is suggested to simplify the
less thermal capacity has been added as compared to wind, since convolution of independent random variables, and then Gram-
Charlier/Edgeworth Expansion theory is used to reconstruct the
pdfs from the cumulants. Furthermore, in [11] the Cumulant
Manuscript received August 19, 2013; revised December 18, 2013 and March method is used to solve the P-OPF problem using the loga-
04, 2014; accepted April 02, 2014. Date of publication May 02, 2014; date
of current version October 16, 2014. This work was supported by National rithmic barrier Interior Point method and the results for variable
Sciences and Engineering Research Council (NSERC) of Canada. Paper no. load are compared using Gaussian distribution and Gamma dis-
TPWRS-00982-2013. tribution. Since some of these approaches require derivatives of
M. Pirnia, C. A. Cañizares, and K. Bhattacharya are with the Department of
nonlinear functions, the functions are assumed to be differen-
Electrical and Computer Engineering, University of Waterloo, Waterloo, ON
N2L 3G1, Canada (e-mail: [email protected]; [email protected]; tiable. Another method to solve the P-OPF problem based on
[email protected]). the two-point estimate method is proposed in [12] which allevi-
A. Vaccaro is with the Department of Engineering, University of Sannio, ates the calculation of derivatives and thus improves the com-
82100 Benevento, Italy (e-mail: [email protected]).
Color versions of one or more of the figures in this paper are available online
putational efficiency of the solution process. Finally, in [13],
at http://ieeexplore.ieee.org. a Chance Constrained Programming (CPP) method is used to
Digital Object Identifier 10.1109/TPWRS.2014.2316114 solve an OPF problem with uncertain demand variables; in order

0885-8950 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
2776 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 6, NOVEMBER 2014

to enhance the execution time, the OPF problem is linearized The IA method is the simplest self-validated range analysis
and a back-mapping approach is implemented to compute the technique, providing conservative bounds for interval values of
probabilities of satisfying the inequality constraints. , with an upper bound and a lower bound . Basic IA op-
An IA method is used in [14] to determine strict bounds to erations between two intervals and assume two conditions:
the solution of the power flow problem for uncertain system 1) none of the intervals are empty, and 2) the upper and lower
parameters, where the interval linear power flow equations are bounds of each interval are finite real numbers. These assump-
solved using iterative methods or employing explicit inverse of tions prevent certain operations such as and
matrices to obtain the hull of the solution set. The IA method [18]. One of the main disadvantages of IA is error ex-
lacks accuracy because of expansive intervals, with increasing plosion, resulting in very conservative final bounds, which are
sources of uncertainties and system size. Self-validated compu- too wide to be useful. Error explosion usually occurs in long
tation approaches, based on affine arithmetic (AA) are proposed chains of computation, where the input of one operation is the
in [15], [16] and [17] for power flow analysis, treating uncer- output of another and is typical in power system analysis prob-
tain variables in affine form. The AA-based method is shown to lems. The main reason for the wider bounds in IA calculations
provide better bounds than IA, since it considers the correlation is the independency problem, in which the correlation between
between the input variables. IA values is neglected. The AA method, on the other hand is a
The present paper proposes a novel application of the AA range analysis technique, which not only handles external un-
method to the OPF problem to compute intervals to capture the certainties such as forecast errors, but also internal errors such
internal and external uncertainties, and subsequently determine as arithmetic roundoff, series truncation and function approxi-
the spinning reserve requirement for a system with variable mation [18]. Although the AA method is computationally more
generation resources without the need for pdfs. The AA-based expensive than the IA method, it provides narrower intervals,
method has significant advantages over existing methods to thus justifying the extra cost.
study the uncertainties in OPF problems, which can be summa- An affine representation of a value (e.g., variable or pa-
rized as follows: rameter) is represented in the following form:
• The method does not rely on the pdf approximation of the
uncertain variables, such as demand and generation. (1)
• In the AA approach, the correlation among the variables
is considered and, therefore, the output intervals are not as Each noise variable , which lies between and 1, represents
conservative as in IA. an independent source of uncertainty, and each coefficient
• The proposed method takes into account the internal errors models the magnitude of that uncertainty. Each affine form can
caused by truncations and approximations. be converted to an interval form by adding or subtracting the
• The method is accurate, as it yields results close to the summation of the absolute values of all noise magnitudes to or
“exact” intervals. from the central value in order to find the upper bound
• The proposed approach is efficient, as it does not require and lower bound , as follows:
many iterations to converge, and hence it is much faster
than other existing methods such as MCS.
(2)
The proposed method is tested and validated using the IEEE
30-bus system and a real 1211-bus European system, demon-
strating the efficiency and accuracy of the technique in real- where is called the total deviation of the affine form .
istic applications, using MCS as the reference for comparison
purposes. A. Affine Operations [18]
The rest of the paper is organized as follows: Section II
presents a brief background on the AA method, and Section III Each affine operation computes an affine form , which
proposes an AA-based OPF solution methodology. In is consistent with the affine input values as follows:
Section IV, the proposed method is tested and validated
on the IEEE 30-bus system and a real 1211-bus European (3)
system. Finally, in Section V the main conclusions and contri-
butions of the paper are highlighted. Thus, consider and :

(4)
II. BACKGROUND
(5)
Numerical calculations are prone to errors in obtaining the
exact values of mathematical computations, because of the pres-
ence of uncertainties. Most of the uncertainty analysis tech- Then, the following elementary affine operations can be defined:
niques, such as the MCS method, try to capture external un-
certainties affecting the input data, and neglect the impact of
internal errors, caused by approximation and truncation. To re-
(6)
solve this issue, “self-validated computation” (SVC) or “auto-
matic result verification” such as the AA methods are proposed, (7)
which inherently keep track of the internal errors [18]. (8)
PIRNIA et al.: A NOVEL AFFINE ARITHMETIC METHOD TO SOLVE OPTIMAL POWER FLOW PROBLEMS WITH UNCERTAINTIES 2777

where is a constant. In (4)–(5), observe that each noise vari- represent the value of a variable under deterministic assump-
ables shows a dependency between the two values and . tions, i.e., when uncertainties are neglected, and noise mag-
Therefore, in (6), contributes to the same uncertainty in the re- nitudes represent deviations of the OPF variables due to the
sulting affine form , with the noise magnitudes corresponding uncertainties.
to the coefficients of in the affine values and .
A. AA-Based Mathematical Model
B. Non-Affine Operations The proposed AA-based OPF model uses a cost minimization
Non-affine operations require an affine approximation and an objective function, with all the uncertain variables, such as real
extra term, called approximation error, to represent internal er- and reactive power generation and , real and reactive
rors. For instance, a non-affine multiplication operation between power demand and , bus voltage magnitude , real
the two affine values and can be written as and imaginary components of bus voltages and , real and
imaginary components of bus currents and , and line
(9) currents in affine form. The following equations correspond
to the rectangular form of the AA-based OPF problem:
The result of this non-affine function is an affine value con- (14)
taining the information provided by and , and the approx-
imation error represented by , where the lower bound of
(15)
can be obtained as follows:
(16)
(10) (17)
(18)
The simplest and most conservative affine approximation (19)
(e.g., trivial affine approximation) can be calculated as follows: (20)

(11) (21)

where is the set of all buses, is the set of all gen-

Therefore, after substituting (11) in (9), the product function can erator buses, is the set of all thermal generates, and is
be represented as the set of all lines. Furthermore, and are affine
real and reactive power mismatch functions, and , ,
, , , , and are minimum and
maximum limits for real power, reactive power, line currents,
(12) and bus voltage magnitudes, respectively, at bus . This model
This affine approximation is computationally efficient, but not cannot be solved directly, as it needs affine representations of
the most accurate because the error of this approximation is at the uncertain variables. The objective function (14) is the cost
most four times the error reported by the most accurate method, interval for the affine function , which is dependent on
i.e., the Chebyshev approximation. Furthermore, is assumed the affine real power generation of thermal generators. The ob-
to be 1 throughout this paper, i.e., jective is neither to minimize the cost of the worst/best case nor
the expected value, but the cost interval.
(13) All the state and control variables in (14)–(21) are in affine
form, comprising a center value and the corresponding noise
magnitudes. The center values are obtained by solving a deter-
Even though this simplification may result in wider intervals, ministic OPF, in which the median of the given intervals for
it reduces the number of noise symbols for each non-affine upper and lower bounds of real and reactive power demand
operation, and since there are many non-affine operations in
and , and and are considered deterministic demands
OPF models, this would significantly reduce the number of
as follows:
noise symbols in realistic applications. In order to overcome
the large intervals, and provide more accurate bounds the
contraction method described in [19] is used in this paper, as (22)
demonstrated in Section III.
(23)
III. AA-BASED OPTIMAL POWER FLOW
An AA-based method is proposed in this paper to solve OPF where is the set of all buses with uncertain demand. In this
problems with uncertainties. Thus, all the variables associated paper, uncertain wind and solar generation sources are treated
with the proposed AA-based OPF problem (e.g., the bus voltage as interval negative loads with a constant power factor, and are
magnitude and the voltage angle, or real and imaginary compo- hence represented using (22) and (23) in the deterministic OPF.
nents of bus voltages in rectangular form) can be stated in affine To obtain the noise magnitudes of real and imaginary com-
form with a center value and noise magnitudes. Center values ponents of bus voltages, a sensitivity analysis is carried out,
2778 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 6, NOVEMBER 2014

where the generation and demand are perturbed by a small mag- where and are the center values for real and imaginary
nitude (e.g., ) at each node. A deterministic OPF is solved components of bus voltages at bus , respectively, and , ,
for each of these variations, thus solving as many OPFs as the , and are defined in (24)–(27).
number of uncertain inputs in the system, which corresponds to The affine forms of real and imaginary components of bus
the number of renewable generators being studied. Therefore, voltages, and can be used to calculate, from (13), the square
the noise magnitudes can be obtained as follows [15]: of bus voltage magnitude, using the following relationship:

(24) (30)

Thus, has the following form:

(25)

(26)

(27) (31)

where and are partial deviations of real and imaginary Here and are truncation errors due to non-affine oper-
components of bus voltages due to changes in real power injec- ations. Furthermore, knowing the affine forms of the real and
tion, and and are partial deviations of real and imaginary imaginary components of the bus voltage magnitude, the real
components of bus voltages at bus due to changes in reactive and reactive power can be calculated using the following affine
power injection at bus . The parameters and are the operations:
new real and imaginary components of the bus voltages when
the real and reactive power injections are perturbed; and (32)
are the initial values of real and imaginary components of bus
voltages obtained from the deterministic model; and and
are the amount of perturbation in real and reactive power where , and and are the real and imaginary
injections at bus . components of the -bus matrix, respectively. Therefore, the
Note that since the power flow equations are nonlinear, the linear affine forms of real and imaginary bus currents and
suggested calculations in (24)–(27) may result in the underes- can be calculated as follows:
timation of the solution and, therefore, not include the “exact”
solutions. In order to avoid underestimation of the solution, an
“affine extension” technique suggested in [15] and [20] is used
here to increase each noise magnitude using an independent (33)
magnification coefficient. Also, it is assumed that an OPF so-
lution for the center value can be obtained as the starting point;
hence, the derivatives (24)–(27), which can be obtained by in-
verting Jacobians of the center-value OPF solution, exist. In the (34)
case that a center-value OPF solution cannot be obtained, the
proposed technique cannot be applied; this may be resolved by where and have the following general forms after affine
redefining the intervals. It should be noted that, since the clas- operations:
sical OPF problem is highly non-convex, using NLP solution
methods such as the IP, do not guarantee a global optimal so- (35)
lution. To prevent widely different solutions during the pertur-
bation method used in the sensitivity analysis, at each run, the
initial values for each variable are set at the values obtained
from solving the center-value OPF. Since these perturbations
are small, this procedure guarantees in practice obtaining local (36)
maxima close to the center-value OPF solutions.
The affine forms of the real and imaginary components of the Here, and are the center values for real and imaginary
bus voltage magnitude and , which are linear functions of components of current magnitudes. and are partial de-
noise variables and representing the uncertainties of viations of the real component of current at bus for deviation
in real and reactive power injection at a bus , respectively; and
active power and reactive power injections at bus , can then be
presented as follows: and are partial deviations of imaginary component
of current at bus for deviation in real and reactive power injec-
(28) tion at bus , respectively. Note that the real and imaginary com-
ponents of the current share the same sources of uncertainties,
i.e., real and reactive power injections and . Further-
(29)
more, using affine and non-affine operations and , , , and
PIRNIA et al.: A NOVEL AFFINE ARITHMETIC METHOD TO SOLVE OPTIMAL POWER FLOW PROBLEMS WITH UNCERTAINTIES 2779

, the real and reactive power mismatch and (40) for real and reactive powers can be represented in interval
in (15) and (16), respectively, can be calculated as follows: forms and , where

(49)
(37)
(50)

(51)
(38)
(52)
where and represent the affine real and reactive power in-
jections, and have the following affine forms, with center value
and associated partial deviations: Here, and are the fol-
lowing functions of noise variables and , with the most
conservative values when they are equal to 1, and represent the
total amount of deviation from the center value:
(39)

(53)
(40)

where and are the center values of affine real and reac-
tive power injections; and are the partial deviations of (54)
real and reactive power injections due to changes in real power
injections at a bus , respectively; and are the partial These intervals can be used to determine the operational range
deviations of real and reactive power injections due to changes of generation for dispatchable generators (e.g., thermal, hydro),
in reactive power injections at bus , respectively; and and when there are sources of uncertainty such as renewable gener-
are real and reactive power injection truncation errors, based ation in the system.
on (13). The above affine forms can be calculated as follows:
B. Contraction Method
(41)
The intervals obtained in (49)–(52) are dependent on the
(42) value of the noise variables and , which are between
and 1. The most conservative intervals are obtained by
(43) fixing these variables at their maximum values; however, this
results in large intervals. Hence, based on the method presented
in [19], [21], and [22], a contraction method is used to reduce
(44) the bounds. In order to reasonably contract the intervals, while
respecting the physical characteristics of the system (e.g.,
(45) voltage and generation limits), the discussed next LP is solved
to obtain the minimum values of noise variables; the suggested
optimization problem is linear since all the affine variables are
(46) constructed using linear affine operations (6)–(13). Thus, by
replacing in the AA-based OPF model (14)–(21), all uncertain
variables with their AA forms defined in (28)–(40), one obtains
the following representation of the OPF model:

(55)
(47)
(56)

(57)

(58)
(48)
(59)
This formulation, as previously mentioned, is the most conser-
vative but computationally efficient for calculating the magni-
tude of the internal errors and . The affine forms (39) and (60)
2780 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 6, NOVEMBER 2014

(61)
(62)
(63)
The above linear noise contraction model can be easily solved
using available LP solvers, such as CPLEX [23]. Then the ob-
tained values of and are used in (49)–(52).
In the above LP problem, the objective function is the
affine linear expansion of (14), as follows:

(64)
Note that in the above equation, due to an affine product, a new
error magnitude is introduced. The line currents and
in (60) are calculated as follows:

(65)

(66)
Fig. 1 depicts the procedure used to calculate the intervals
for the affine variables and hence arrive at the solution intervals
to the OPF with uncertainties. Note that the demand intervals at
each bus are the input of the model. These intervals consider the
uncertainties associated with both demand and renewable gen-
eration by considering the latter as negative loads. Then affine
variables and are constructed using both a deterministic
OPF model to obtain center values, and a sensitivity analysis
based on several perturbed OPFs to obtain the noise magnitudes. Fig. 1. Proposed AA-based OPF.
Affine operations are then applied on these affine variables in
order to calculate other affine variables, such as and . To
minimize the size of the intervals, a contraction method is used preconditioning, and hence it could not be used for the types of
to minimize the noise magnitudes associated with each of these studies presented in the current paper.
variables. In the 30-bus system, one of the generators is replaced with a
wind turbine to study the effects of uncertainty arising from the
IV. RESULTS AND DISCUSSIONS
renewable sources in the system. The 1211-bus real European
In this section, the proposed AA-based OPF is tested using the system has 160 generators (thermal, hydro, wind, and solar),
IEEE 30-bus system [24] and a real European 1211-bus system, 58 are solar and 8 are wind generators, for a total renewable
and compared with the MCS solution. The IA method has not capacity of 11 678 MW; considering that the total generation
been used as a mean for comparison in this paper, as the bounds capacity of the system, including imports, is 183 GW, this cor-
provided by the IA approach are too conservative or simply responds to 6.4% of the generation capacity. The total system
cannot be obtained and therefore it is not a feasible method. demand is 153 GW, including exports. The transmission system
The IA wider bounds are simply because of not considering comprises 1567 lines and 122 fixed transformers. The intermit-
the correlation among the variables which leads to error explo- tency in wind and solar generation is assumed to be compen-
sion. In fact, one of the authors has studied the application of sated by thermal generation in both of the test systems, and
the IA method to the OPF, and observed that the IA-based OPF thus the proposed AA-based method is used to determine the
problem diverges for realistic systems with a reasonable number power output intervals of these generators, hence calculating the
of uncertainties due to too wide intervals, which resulted in sin- generation reserve needed to reliably and optimally supply the
gularities of the Jacobian of the Lagrangian function, even after demand.
PIRNIA et al.: A NOVEL AFFINE ARITHMETIC METHOD TO SOLVE OPTIMAL POWER FLOW PROBLEMS WITH UNCERTAINTIES 2781

Fig. 2. Bus voltage magnitude intervals. Fig. 3. Bus voltage magnitude intervals, selected buses.

The proposed model is simulated in the General Alge-

braic Modeling System (GAMS) [25]. The solvers used, i.e.,
COINIPOPT for the OPF solution [26], and CPLEX for the
LP contraction problem, have their parameters set at their
respective default, off-the-shelf settings, so as not to bias their
“standard” performance. Major settings such as tolerance level
or maximum number of iterations of the solver are by default
the same for all solvers (e.g., feasibility tolerance is ).
The implemented MCS method converges after 3000 itera-
tions for both the IEEE 30-bus test systems, and the 1211-bus
system, assuming that the uncertain parameters have uniform
distribution within the bounds defined for the assumed variable
generation. Fig. 4. Real power generation intervals.
A. 30-Bus System
In the IEEE 30-bus test system, a wind turbine is assumed to
be located at Bus 2. It is considered that this generator power
output varies in a range of its forecasted center value,
, since this is treated as a negative injection
with unity power factor.
All the center values are obtained by executing the nonlinear
OPF model with the deterministic data, and the noise mag-
nitudes are obtained from the sensitivity analysis technique
described in Section III. The intervals obtained by the MCS
method are used as the benchmark to check the validity of the
obtained AA-based intervals. Figs. 2 and 3 compare the bus Fig. 5. Real power generation interval at Bus 8.
voltage magnitude bounds using both AA (VupAA and VloAA)
and MCS (VupMC and VloMC) methods. Note that there is no
significant difference between the two approaches. The bounds B. 1211-Bus System
obtained by the MCS method lie slightly inside the AA-based The AA method is also applied to a real European 1211-bus
ones, because the MCS method provides the “exact” intervals, system. Figs. 8 and 9 show the real and reactive power devi-
whereas the AA-based approach provides more conservative ations of the thermal generators, and Fig. 10 shows the devia-
margins, as expected. Observe that the difference between the tion of bus voltage magnitudes at all buses with respect to the
AA-based and MCS-based methods is very small. center value, in percentage, obtained from both the AA-based
The real power intervals, shown in Figs. 4 and 5, depict the approach and MCS. In this case, a deviation active power
interval output for thermal generators at Bus 1 and Bus 8, when injection for wind and solar with unity power factor is assumed.
the wind turbine output varies in the given interval at Bus 2; all The deviations are obtained as follows:
other generators are synchronous condensers and hence have
zero real power output. Figs. 6 and 7 demonstrate the recative
power generation intervals. As shown, the resulted AA-based (67)
margins of reactive power generation closely match the ones ob-
tained from the MCS method. Note that the difference between Observe that the errors in this case are significantly large for
the upper and lower bounds of the illustrated intervals in Fig. 6 some generators than those obtained for the smaller IEEE
is very small. 30-bus system, as expected, since the number of uncertainties
2782 IEEE TRANSACTIONS ON POWER SYSTEMS, VOL. 29, NO. 6, NOVEMBER 2014

Fig. 6. Reactive power generation intervals.

Fig. 9. Percentage of reactive power deviation for thermal generators.

Fig. 7. Reactive power generation interval at Bus 11.

Fig. 10. Percentage of deviation for bus voltage magnitudes.

Fig. 8. Percentage of real power deviation for thermal generators. TABLE I

THERMAL CAPACITIES, USING MCS AND AA-BASED APPROACHES

is much larger, i.e., 1 in the small system versus 66 in the real

method in the maximum total thermal generation required to
compensate for the wind and solar power generation uncertain-
ties. Note that the 1.02% total error in Table I is only 1231 MW
in the real European system.
The AA method is faster than the MCS approach, as the MCS
method needs 3000 iterations for convergence, while the AA
method requires as many iterations as the number of uncertain such as variable wind and solar power generation. The inter-
variables, i.e., one iteration for the IEEE 30-bus test system vals obtained from the proposed technique can be used to ap-
and 66 iterations for the real 1211-bus system. Furthermore, proximate the margins of operation for dispatchable generators
the additional OPFs are relatively small perturbations of the needed to properly account for system uncertainties. In order
center-value OPF, and hence convergence can be quickly at- to test the efficiency and accuracy of the AA-based OPF, the
tained by using the center-value OPF solution as starting point. method was tested on the IEEE 30-bus test system and a real
1211-bus European system, and benchmarked against the MCS
V. CONCLUSIONS method. The AA-based approach was shown to efficiently yield
A novel AA-based model has been proposed to solve the intervals close to the MCS bounds. This method can be used to
OPF problem with intervals to represent system uncertainty, also study the impact of demand variations in power systems,
PIRNIA et al.: A NOVEL AFFINE ARITHMETIC METHOD TO SOLVE OPTIMAL POWER FLOW PROBLEMS WITH UNCERTAINTIES 2783

and other similar applications where intervals could be used to [20] D. Grabowski, M. Olbrich, and E. Barke, “Analog circuit simulation
represent uncertainties, without the need to assume pdfs. using range arithmetics,” in Proc. Design Automation Conf. 2008 Asia
and South Pacific, Seoul, Korea, 2008, pp. 762–767.
[21] H. Sherali and W. P. Adams, A Reformulation-Ninearization Tech-
ACKNOWLEDGMENT nique for Solving Discrete and Continuous Nonconvex Problems.
New York, NY, USA: Springer, 1999.
The authors would like to thank Prof. G. Verbic for sharing [22] J. Ninin, P. Hansen, and F. Messine, “A reliable affine relaxation
method for global optimization,” Groupe D’études et de Recherche en
his studies on the IA-based OPF.
Analyse des Décisions, 2010.
[23] Cplex, 10.0 User’s Manual, 2006.
[24] Univ. Washington, UWashington Test System Archive, 1961 [On-
REFERENCES line]. Available: http://www.ee.washington.edu/research/pstca/pf30/
[1] PJM, PJM Reserve Requirement Study, Oct. 2012 [Online]. pg_tca30bus.htm
Available: http://www.pjm.com/~/media/committees-groups/sub- [25] R. E. Rosenthal, GAMS—A User’s Guide. Washington, DC, USA:
committees/raas/20110929/20110929-2011-pjm-reserve-require- GAMS Development, 2013.
ment-study.ashx [26] M. J. Saltzman, COIN-OR: An Open-Source Library for Optimiza-
[2] K. W. Cheung, P. Shamsollahi, D. Sun, J. Milligan, and M. Potishnak, tion. Boston, MA, USA: Kluwer, 2002.
“Energy and ancillary service dispatch for the interim ISO New Eng-
land,” IEEE Trans. Power Syst., vol. 15, no. 3, pp. 968–974, Aug. 2000. Mehrdad Pirnia (S’14) received the B.Sc degree in industrial engineering from
[3] K. W. Cheung, “Ancillary service market design and implementation Bu-AliSina University, Iran, in 2001 and the M.A.Sc degree in management
in North America: From theory to practice,” in Proc. Electric Utility sciences and Ph.D. degree in electrical engineering from the University of Wa-
Deregulation and Restructuring and Power Technologies, Redmond, terloo, Waterloo, ON, Canada, in 2009 and 2014, respectively.
WA, USA, 2008. He is currently an instructor with the Department of Management Science
[4] R. Baldick, “Wind and energy markets: A case study of Texas,” IEEE of the University of Waterloo. His research interests are in electricity market
Syst. J., vol. 6, no. 1, pp. 27–35, 2012. modeling, renewable energy integration, and large scale optimization.
[5] A. Gómez-Expósito, A. Conejo, and C. A. Canizares, Electric Energy
Systems: Analysis and Operation. Boca Raton, FL, USA: CRC, 2008.
[6] P. Zhang and S. T. Lee, “Probabilistic load flow computation using the Claudio A. Cañizares (S’86–M’91–SM’00–F’07) received the Electrical En-
method of combined cumulants and Gram-Charlier expansion,” IEEE gineer degree from the Escuela Politecnica Nacional (EPN), Quito, Ecuador, in
Trans. Power Syst., vol. 19, no. 1, pp. 676–682, Feb. 2004. 1984 and the M.Sc. and Ph.D. degrees in electrical engineering from the Univer-
[7] G. L. Viviani and G. T. Heydt, “Stochastic optimal energy dispatch,” sity of Wisconsin-Madison, Madison, WI, USA, in 1988 and 1991, respectively.
IEEE Trans. Power App. Syst., vol. PAS-100, no. 7, pp. 3221–3228, He held different teaching and administrative positions at EPN from 1983
1981. to 1993. He has been with the Electrical and Computer Engineering Depart-
[8] M. E. El-Hawary and G. A. N. Mbamalu, “A comparison of proba- ment at the University of Waterloo, Waterloo, ON, Canada, since 1993, where
bilistic perturbation and deterministic based optimal power flow solu- he has held various academic and administrative positions and is currently a full
tions,” IEEE Trans. Power Syst., vol. 6, no. 3, pp. 1099–1105, Aug. Professor, the Hydro One Endowed Chair, and an Associate Director of the Wa-
1991. terloo Institute for Sustainable Energy (WISE). His research activities concen-
[9] T. S. Karakatsanis and N. D. Hatziargyriou, “Probabilistic constrained trate in the study of stability, modeling, simulation, control and computational
load flow based on sensitivity analysis,” IEEE Trans. Power Syst., vol. issues in power systems within the context of competitive electricity markets
9, no. 4, pp. 1853–1860, Nov. 1994. and smart grids.
[10] P. Zhang and S. T. Lee, “Probabilistic load flow computation using the Dr. Cañizares is a Fellow of the Royal Society of Canada and the Canadian
method of combined cumulants and Gram–Charlier expansion,” IEEE Academy of Engineering, has been the recipient of various IEEE-PES Working
Trans. Power Syst., vol. 19, no. 1, pp. 676–682, Feb. 2004. Group awards, and holds and has held several leadership positions in IEEE-PES
[11] A. Schellenberg, W. Rosehart, and J. Aguado, “Cumulant-based prob- technical committees and subcommittees.
abilistic optimal power flow (P-OPF) with Gaussian and gamma dis-
tributions,” IEEE Trans. Power Syst., vol. 20, no. 2, pp. 773–781, May
2005. Kankar Bhattacharya (M’95–SM’01) received the Ph.D. degree in electrical
[12] G. Verbic and C. A. Canizares, “Probabilistic optimal power flow in engineering from the Indian Institute of Technology, New Delhi, in 1993.
electricity markets based on a two-point estimate method,” IEEE Trans. He was on the faculty of Indira Gandhi Institute of Development Research,
Power Syst., vol. 21, no. 4, pp. 1883–1893, Nov. 2006. Mumbai, India (1993–1998), and Chalmers University of Technology, Göte-
[13] H. Zhang and P. Li, “Chance constrained programming for optimal borg, Sweden (1998–2002). He has been with the Department of Electrical and
power flow under uncertainty,” IEEE Trans. Power Syst., vol. 26, no. Computer Engineering at the University of Waterloo, Waterloo, ON, Canada,
4, pp. 2417–2424, Nov. 2011. since 2003, and is currently a full Professor and Associate Director of the Wa-
[14] Z. Wang and F. L. Alvarado, “Interval arithmetic in power flow anal- terloo Institute for Sustainable Energy (WISE). His research interests are in
ysis,” IEEE Trans. Power Syst., vol. 7, no. 3, pp. 1341–1349, Aug. power system operations and economics, electricity markets, deregulation and
1992. retail competition issues, distributed generation economics and planning, global
[15] A. Vaccaro, C. A. Canizares, and D. Villacci, “An affine arithmetic- energy models, scenario analysis, long-term planning, and energy-environment
based methodology for reliable power flow analysis in the presence of linkages and sustainable development.
data uncertainty,” IEEE Trans. Power Syst., vol. 25, no. 2, pp. 624–632,
May 2010.
[16] M. Pirnia, C. A. Canizares, K. Bhattacharya, and A. Vaccaro, “An Alfredo Vaccaro (M’01–SM’09) received the M.Sc. degree with honors in elec-
affine arithmetic method to solve the stochastic power flow problem tronic engineering in 1998 from the University of Salerno, Salerno, Italy.
based on a mixed complementarity formulation,” in Proc. 2012 IEEE From 1999 to 2002, he was an Assistant Researcher at the University of
Power and Energy Society General Meeting, San Diego, CA, USA, Salerno, Department of Electrical and Electronic Engineering. Since March
2012, pp. 1–7. 2002, he has been an Assistant Professor in electric power systems at the
[17] A. Vaccaro, C. Canizares, and K. Bhattacharya, “A range arithmetic- Department of Engineering of the University of Sannio, Benevento, Italy. His
based optimization model for power flow analysis under interval un- special fields of interest include soft computing and interval-based method
certainty,” IEEEE Trans. Power Syst., vol. 28, no. 2, pp. 1179–1186, applied to power system analysis and advanced control architectures for
May 2013. diagnostic and protection of distribution networks.
[18] L. H. De Figueiredo and J. Stolfi, “Self-validated numerical methods Prof. Vaccaro is an Associate Editor and member of the Editorial Boards
and applications,” in Brazilian Mathematics Colloquium Monograph, of IET Proc. Renewable Power Generation, the International Journal of Elec-
Rio de Janiero, Brazil, 1997. trical and Power Engineering, the International Journal of Reliability and Safety,
[19] J. Ninin and F. Messine, “A mixed affine reformulation method for the International Journal on Power System Optimization, and the International
global optimization,” in Proc. TOGO, 2010, pp. 1–4. Journal of Soft Computing.