Robust Transmission Expansion Planning with

Uncertain Generations and Loads using Full

Probabilistic Information
Cristina Roldán, and Raquel Garcı́a-Bertrand Roberto Mı́nguez
Department of Electrical Engineering ESCISSION Company
Universidad de Castilla- La Mancha Spain
Ciudad Real, Spain [email protected]
{Cristina.Roldan, Raquel.Garcia}@uclm.es

Abstract—Recent advances in Transmission Network Expan- ΩD Set of indexes of loads.

sion Planning (TNEP) have demonstrated that two-stage adaptive D Uncertainty set.
robust optimization (ARO) renders the expansion planning prob-
lem tractable for real systems and, at the same time, constitutes a B. Constants
relevant approach to deal with uncertain demand and generation
capacity in the TNEP problem. However, the use of cardinality α Confidence level.
and polyhedral uncertainty sets does not consider the correlation β Reliability index.
of uncertain parameters and makes the interpretation of final δ Scale parameter of the Weibull distribution.
results difficult from a probabilistic view-point. The recent use  Convergence tolerance for the level III loop.
of level II structural-reliability and decomposition techniques
allows the incorporation of ellipsoidal uncertainty sets, which IL Convergence tolerance for the inner loop.
permits the use of expected values and the variance-covariance OL Convergence tolerance for the outer loop.
matrix of uncertain parameters (First-Order Second-Moment, κ Shape parameter of the Weibull distribution.
FOSM). However, it also constitutes an approximation. This µln,D Mean values vector of the logarithm of load dD .
paper presents an iterative method which, using FOSM and µD Mean value of load dD D
appropriately modifying the reliability index associated with l l within the set Ω .
the desired objective function quantile, transforms the model σ ln,D Standard deviation values vector of the logarithm
into a level III structural-reliability method that allows the of load dD .
consideration of the complete probabilistic structure of the σlD Standard deviation value of load dD l within the
uncertain parameters. Numerical results from a realistic case set ΩD .
study are presented and discussed. ω Correlated wind speed vector.
Index Terms—Correlated uncertainty, ellipsoidal uncertainty ωs Sampled wind speed vector.
set, nested decomposition, structural reliability, transmission Π Investment budget.
network expansion planning, two-stage robust optimization. ρlh Element (l, h) of the load correlation matrix.
ρ0lh Element (l, h) of the load correlation matrix in
N OMENCLATURE the logarithmic space.
This section lists the main notation used throughout the Σ Variance-covariance matrix.
paper. Symbol “ ˆ ” is used to represent the optimal value of a Σln,D Variance-covariance matrix of the logarithm of
specific variable. Additional symbols with superscripts “(k)” load dD .
and “(m)” are used to indicate the value of a specific variable ΣW Variance-covariance matrix of the transformed
at iterations k and m of the outer loop, respectively. Similarly, values of n.
superscript “(ν)” and “(ι)” are used to denote results obtained ς Scaling factor.
at iteration ν and ι of the inner loop and the level III loop, b Vector of operating cost coefficients.
respectively. CL Vector of investment cost coefficients.
d¯ Vector of mean or forecast values.
A. Sets Lln,D Matrix obtained from Cholesky factorization of
Ω(·) Feasibility set of x. ΣLn,D .
LW Matrix obtained from Cholesky factorization of
ΣW .
This work has been partially funded by the Ministry of Science, Innova- PD Vector of demand scenarios.
tion, and Universities of Spain under Projects RTI2018-096108-A-I00 and
RTI2018-098703-B-I00 (FEDER, UE), and in part by the Universidad de PGmax Vector of wind power scenarios.
Castilla-La Mancha under Grant 2019-GRIN-26950. n Vector of correlated standard normal variables.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 nnt Vector of correlated standard normal trans- and demand consumptions within a given geographical area
formed variables. may be significant [2]–[4].
Over the last years, two main frameworks have been
C. Decision Variables used to tackle the uncertainty characterizing TNEP, stochastic
γ Approximation of the worst-case operating cost. programming [5]–[8] and robust optimization [9] and [10].
d Vector of uncertain generation capacities and While stochastic programming requires accurate probabilistic
demands. information and relies on a computationally expensive dis-
dD Vector of uncertain demands. cretization of uncertainty based on scenarios, robust optimiza-
dG Vector of uncertain generation capacities. tion models the uncertainty by decision variables within a
v Vector of binary variables representing expan- pre-specified uncertainty set. Moreover, instead of optimiz-
sion decisions. ing the expected value of the objective function over all
x Vector of second-stage variables associated with scenarios under consideration as in stochastic programming,
the system operation under uncertainty such as robust optimization provides a solution which is immune to
production and consumption levels, load shed- all plausible uncertainty realizations within the uncertainty
ding, and network-related variables. set, thereby being optimal for the worst-case realization of
xm Vector of operation-related variables under the uncertain parameters. Taking this into consideration, recent
worst case identified at iteration m. advances on TNEP problems [11]–[15] proved that the use
of Adaptative Robust Optimization (ARO) frameworks [16]
D. Functions and [17], as opposed to stochastic programming approaches,
Φ(·) Cumulative distribution function of the standard renders the TNEP problem considering uncertainties compu-
normal distribution. tationally tractable for real systems.
cO,min (·) Minimum operating cost. ARO materializes in a three-level formulation: i) the first
cO,min
M C (·) Minimum operating cost from Monte Carlo sim- level minimizes the cost of expansion (also minimizes the
ulation. maximum regret), ii) the second level selects the worst re-
cO,wc (·) Worst-case operating cost. alization of the uncertain parameters maximizing the system’s
FW (·) Weibull cumulative distribution function. operation costs within the uncertainty set, and iii) the third
level selects the optimal decision variables to minimize oper-
E. Others ating costs for given values of first and second level variables.
η Vector of dual variables associated with the The main difference among contributions using ARO for
constraints fixing the values of d. TNEP is the solution approach and how the uncertainty set
C MP Approximation of the optimal total cost provided is defined. Jabr [11] presents a two-stage robust optimization
by the master problem. approach for TNEP under uncertainty of renewable generation
C SP Approximation of the optimal total cost provided and load. The approach is based on the application of Benders
by the subproblem. decomposition, which involved the iterative solution of a
d˜ Vector of wind power and demand scenarios master problem and a max-min subproblem associated with
from Nataf model. the second stage. The original max-min subproblem was cast
zD Standard normal random vector for demand sce- as a single-level maximization problem relying on duality
narios calculation. theory. To deal with uncertain parameters, Jabr proposed a
zW Standard normal random vector for wind power cardinality uncertainty set where binary variables and uncer-
scenarios calculation. tain budgets were introduced. Chen et al. [12] and Ruiz and
Conejo [13] addressed the same problem modeled in [11] but
I. I NTRODUCTION using different polyhedral uncertainty sets. Unlike [11], in
Transmission network expansion planning (TNEP) analyzes that case the max-min subproblem was transformed into an
the issue of how to expand or reinforce the transmission equivalent mixed-integer linear program using Karush-Kuhn-
network in order to meet the future demand with the available Tucker optimality conditions. Mı́nguez and Garcı́a-Bertrand
generation capacity [1]. This problem is challenging due to the [14] improved computational performance by taking different
great amount of uncertainty associated mainly with consump- features from [11], [12] and [13]. In particular, [14] replaced
tions and available generation capacity. Uncertainty related the lower-level problem with a dual one, and solved the re-
to demand has been the first issue raised in conventional sulting bi-level problem using a primal cutting plane algorithm
TNEP models. Within the current context where the growing within a decomposition scheme.
installation of renewable-based generation assets comes into Previous authors rely on cardinality and polyhedral un-
play, networks planners also face unprecedented levels of certainty sets, [11]–[14]. However, the inherent hypothesis
uncertainty related to available generation capacity. Moreover, that uncertain parameters follow a symmetric distribution
due to the strong relation of renewable energies to meteoro- within the selected intervals does not match reality, specially
logical phenomena and the existence of similar consumption for power production associated with wind generation. Due
patters, the spatial correlations of renewable energy sources to this limitation, Roldán et al. [15] propose a two-stage

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
adaptative robust approach for TNEP where correlated un- The remainder of the paper is structure as follows. Section
certainty sources are explicitly accounted for. The approach II describes the adaptive robust formulation of the TNEP
relies on an ellipsoidal uncertainty set based on the first and problem and presents the definition of the uncertainty set
second moments of the probability distribution of the uncertain solution approach to solve the TNEP problem. Section III
parameters, i.e., mean or forecast values and the variance- details the new characterization of uncertain parameters and
covariance matrix. However, in terms of the definition of the the probabilistic information required. Section IV is devoted to
require quantile to be optimized, this method also constitutes the proposed level III method. Numerical results for a realistic
an approximation and does not allow to use the knowledge case study are given in Section V. Finally, relevant conclusions
about probability distributions associated with wind speeds are drawn in Section VI.
and/or correlations.
II. A DAPTIVE TNEP COMPACT FORMULATION
The aim of this work is to extend the methodology proposed
by [15] and incorporate the complete probabilistic informa- This section presents the ellipsoidal uncertainty set suitable
tion associated not only with wind generation but also with for the correlation structure of uncertainty sources, the robust
demands. The use of an ellipsoidal uncertainty set allows to counterpart for transmission network expansion planning un-
establish an analogy between the second stage problem and der uncertainty and the solution approach to solve the TNEP
an instance of mathematical programming arising in structural problem. Note that for clarity in the exposition, we adopt the
reliability First-Order-Second-Moment (FOSM) [15]. Based same nomenclature used by [15].
on this analogy, the resulting trilevel program is solved by the A. Uncertainty characterization
combined use of a column-and-constraint generation algorithm Analogously to [15] and [18], we propose characterizing
and a decomposition technique applied in structural reliability. the uncertainty of generation capacities dG and demands dD
Therefore, this work presents an iterative method which, using through an ellipsoidal uncertainty set D. Let us assume that
FOSM and modifying appropiately the reliability index associ- uncertain parameter vector d = (dG , dD )T has nominal or
ated with the desired objective function quantile, transform the expected values d¯ and variance-covariance matrix Σ, respec-
model into a level III structural-reliability method that allows tively. Based on [18], the ellipsoidal uncertainty set can be
the consideration of the complete probabilistic structure of written using the Mahalanobis distance as follows:
the uncertain parameters involved and the calculation of the n
T o
−1
D (β) = d − d¯ (Σ) d − d¯ ≤ β 2 .


operating costs probability distribution. (1)
The FOSM method provides an approximation of the
In contrast with cardinality and polyhedral uncertainty sets
confidence level (probability of obtaining an operating cost
used in previous works [11]–[14], the ellipsoidal uncertainty
not exceeding the optimal value of the objective function in
set (1) allows to consider the correlation structure of uncertain
the middle-level problem) due to the simplificacion adopted
parameters through their variance-covariance matrix Σ, and
in the probability calculation and the use of only the first
besides, one unique nonnegative parameter β controls the level
two moments of the probability distribution of the uncertain
of robustness.
parameters. Thus, by incorporating the complete probabilistic
structure of uncertain parameters we make sure that once the B. Compact problem formulation
optimal solution has been achieved, the objective function truly The two-stage adaptive robust TNEP problem can be ex-
corresponds to the optimal required quantile. pressed, in compact form, as the following trilevel mathemat-
The major contributions of this paper are as follows: ical programming problem:
1) The proposed method represents an extension of the T
novel two-stage adaptive robust optimization model for Minimize C L v + ςcO,wc (v) (2)
v
the transmission network expansion planning problem subject to:
under correlated uncertainty developed in [15]. T
2) The described iterative method uses Level III structural- CL v ≤ Π (3)
reliability and decomposition techniques based on First- v ∈ {0, 1} (4)
Order Second-Moment methods, which allows the con- O,wc
where c (v) is provided by:
sideration of sophisticated joint density functions.
3) The Nataf model is used to define the real joint proba- cO,wc (v) = Maximize cO,min (v, d) (5)
d
bility distribution model for wind generations and loads.
subject to:
4) By incorporating the complete probabilistic structure
T −1
d − d¯ (Σ) d − d¯ ≤ β 2


of uncertain parameters and modifying the reliability (6)
index, the exact quantile of the minimum operating cost O,min
and where c (v, d) is obtained from:
associated with a previously defined confidence level is
O,min
calculated. c (v, d) = Minimize bT x (7)
x
5) The use of Monte Carlo simulation allows to obtain the
subject to:
density function associated with costs, which gives rise
to an easy interpretation of the robust solution. x ∈ Ω(v, d). (8)

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 The first stage problem (2)–(4), which corresponds to the • Set the total cost associated with the outer loop
upper-level problem, identifies the least-cost expansion plan v, C MP = 0.
while the second-stage problem (5)–(8), corresponding to the 2) Initialization of the inner loop. Set the iteration counter
two lowermost optimization levels, identifies the worst-case of the inner loop ν to 1, select initial values for d(ν)
value of each uncertain parameter in order to get the worst- and set the operating cost associated with the inner loop
case operating cost cO,wc (v). C SP = ∞.
The objective function of the upper-level problem (2) lies 3) Solution of the subproblem for lower-level variables.
in the minimization of the total cost, which comprises the Solve the subproblem for lower-level variables and the
investment cost and the worst-case operating cost. The scaling given expansion plan v (k) and given values of the
factor ς is used to make investment and worst-case operating random parameters d(ν) :
costs comparable quantities. Constraint (3) sets an upper bound
on the investment cost and constraint (4) keeps the integrality
of v. The objective of the middle-level problem (5) is the cO,min (v (k) , d(ν) ) = Minimize bT x (9)
d,x
maximization of the minimum operating cost. Constraint (6)
corresponds to the ellipsoidal uncertainty set described in subject to:
Subsection II-A. Finally, the lower-level problem (7) mini- x ∈ Ω(v (k) , d) (10)
mize the operating cost with respect to operating decisions. d=d (ν)
: η. (11)
Constraint (8) represents operational constraints such as nodal
power balances, line flows through existing lines, lines flows
through candidate lines, flow and voltage angle limits, etc. For This step provides η (ν) and cO,min (v (k) , d(ν) ). Update
a detailed explanation of each level, we recommend [15]. the optimal value of the total cost minimized in (2)
T
C SP = C L v (k) + ςcO,min (v (k) , d(ν) ).
C. Solution approach for TNEP problem 4) Update of the iteration counter of the inner loop. In-
crease the iteration counter ν ← ν + 1.
Once the uncertainty set is properly defined and the problem
5) Solution of the subproblem for middle-level variables.
formulation is presented, the methodological proposal for
Solve the subproblem for middle-level variables and for
solving problem (2)–(8) is based on performing two loops.
given d(ν−1) and η (ν−1) :
The outer loop is based on iteratively solving a single-
level master problem and a max-min subproblem through
a column-and-constraint generation algorithm. The master T
Maximize cO,min (v (k) , d(ν−1) ) + η (ν−1) (d − d(ν−1) )
problem constitutes a relaxation for problem (2)–(8) where d
a set of valid operating constraints are iteratively added, while (12)
the subproblem is a max-min problem comprising the second subject to:
stage problem (5)–(8). The inner loop is another iterative
T −1
d − d¯ (Σ) d − d¯ ≤ β 2 .


(13)
procedure which consists on the resolution of the max-min
subproblem at each outer loop iteration. Moreover, the use
of the ellipsoidal uncertainty set described in Subsection II-A The application of the decomposition method presented
allows to establish an analogy between the max-min subprob- in [19] relies on the availability of an analytical expres-
lem and a mathematical programming problem associated with sion for middle-level variables, d. Therefore, problem
First-Order Second-Moment (FOSM) approach from structutal (12)–(13) can be solved through the following equivalent
reliability, for which a decomposition based method was devel- analytical expression:
oped in [19]. Therefore, the inner loop is solved through that
decomposition method, which involves the iterative solution Ση (ν−1)
of two optimization problems: 1) the subproblem for middle- d = d¯ + β q . (14)
T
level variables, wherein lower-level variables x are fixed in the η (ν−1) Ση (ν−1)
max-min subproblem; and 2) the subproblem for lower-level
variables, wherein middle-level variables d are fixed in the
It must be highlight that in case expression (14) leads
max-min subproblem. Note that the inner loop only guarantees
to negative values of d, which is not physically possible
the attainment of a suboptimal solution, which corresponds
in this problem, d must be recalculated directly solving
to a local optimum in case no degeneracy exists, i.e, if the
the problem (12)–(13) and enforcing d ≥ 0.
gradient of the minimum operating cost is unique [15].
This step provides d(ν) .
The iterative scheme is summarized in the following algo-
6) Inner-loop convergence checking. If |d(ν) − d(ν−1) | ≤
rithm:
IL , then set d(k) ← d(ν) and go to step 7; otherwise,
1) Initialization of the outer loop. go to step 3.
• Set the iteration counter of the outer loop k to 1. 7) Update of the iteration counter of the outer loop. In-
(k) crease the iteration counter k ← k + 1.
• Set the initial expansion plan v = 0.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 8) Master problem solution. Solve the master problem for Then, the sampled wind speed vector ωs can be calculated
given d(m) , m = 1, . . . , k − 1: using the following inverse transformation:
κ
T
Minimize C L v + ςγ (15) Φ(nnt ) = FW (ωs ) = 1 − e−(ωs /δ) . (22)
α,v,xm
Finally, the sampled wind speed vector can be transform into
subject to:
wind power production scenarios PGmax by using the power
Constraints (3)–(4) (16) curve of a turbine .
γ ≥ bT xm ; m = 1, . . . , k − 1 (17) Regarding loads, we also use the Nataf transformation.
xm ∈ Ω(v, d (m)
); m = 1, . . . , k − 1 (18) However, with the aim of ensure the positiveness of loads, their
distributions are considered to be log-normally distributed:
α ≥ 0. (19)
ln(dD ) = µln,D + Lln,D z D , (23)
(k)
This step provides values of the decision variables v ln,D
where L is the Cholesky decomposition of the variance-
and γ. Update the optimal value of the total cost, namely
T covariance matrix related to Σln,D , z D represents a standard
C MP = C L v (k) + ςγ.
normal random vector and µln,D corresponds to the mean
9) Outer-loop convergence checking. If (C SP −
values vector of the logarithm of loads dD , whose components
C MP )/C SP ≤ OL , the algorithm stops; otherwise, go
can be obtained as follows:
to step 2.
µD
µln,D
l = ln r l D
 D 2 ; ∀l ∈ Ω , (24)
III. I MPROVING UNCERTAINTY MODELLING σl
1 + µD
l
So far, the uncertainty has been characterized by using only
first and second-order moment information of the random where µD
l and σlD
are, respectively, the expected value and
parameters involved, which is equivalent to assuming that standard deviation of load dD D
l within the set Ω of loads.
uncertain parameters are normally distributed. Although it The components of the variance-covariance matrix Σln,D
allows to incorporate the correlation structure of uncertain are:
σ D ρ σh D
 
parameters, in terms of the definition of the required quantile ln 1 + lµDlh µD
to be optimized, it also constitutes an approximation of the Σln,D
lh = σlln,D ρ0lh σhln,D , ρ0lh = l h
;
real or more precise uncertainty region obtained from data σlln,D σhln,D
(see [15]). To solve this shortcoming, it is required to define ∀(l, h) ∈ ΩD , (25)
the appropriate joint probability distribution model for wind q
2
generations and loads. where σlln,D = ln(1 + σlD /µD l ), ρlh is the element (l, h)
It is widely accepted that wind speeds follow a Weibull dis- of the load correlation matrix and ρ0lh is the corresponding
tribution. For that reason, regarding wind speeds, we propose element in the logarithmic space.
to use the Nataf model [20], [21] to get wind power scenarios Demand scenarios PD are easily obtained by using the
D
because it allows working easily with marginal distributions inverse of the logarithm, PD = eln(d ) .
and also with correlations among wind sites. Then, assuming Note that for a given reliability index β and once the
a correlated wind speed vector ω and its Weibull cumulative optimal solution of problem (2)–(8) is calculated, it is easy to
distribution function FW (ω) with scale and shape parameters perform an out-of-sample assessment relying on Monte Carlo
δ and κ respectively, the Nataf model allows to transform simulation by using the corresponding wind power PGmax and
that correlated speed vector to a correlated standard normal demand PD scenarios obtained through the Nataf model.
variable vector n as follows:
IV. P ROPOSED METHOD
−(ω/δ)κ
FW (ω) = 1 − e = Φ(n) (20) Based on the analogy between the max-min subproblem
and the FOSM approach from structural reliability analysis
where Φ(·) is the cumulative density function of the standard [15], it is possible to estimate the probability of obtaining
normal distribution. Note that it is straightforward to obtain the an operating cost not exceeding the worst-case operating cost
correlated standard normal vector n by the inverse cumulative cO,wc (v) determined by the second stage problem, which is to
distribution function Φ−1 (·). say that the second-stage problem represents an approximation
Once the correlated standard normal vector n is calculated, of the quantile of the minimum operating cost associated with
we can get the associated variance-covariance matrix ΣW . a confidence level α.
Applying the Cholesky factorization to the variance-covariance This estimation is exact if the random parameters are nor-
matrix ΣW = LW (LW )T and simulating a vector of inde- mally distributed, but an approximation for those cases where
pendent standard normal random variables z W , the correlated they do not. For that reason, because of the FOSM method
standard normal transformed vector nnt is obtain as follows: only uses the first two moments of the probability distribu-
tion of the uncertain parameters (mean value and variance-
nnt = LW z W , (21) covariance matrix), the real probability of non-exceedancing

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
the optimal solution of the operational costs
ĉO,wc (v) of second-order moment information of the uncertain pa-
problem (2)–(8), Prob ĉO,min (v, d) ≤ ĉO,wc (v) , might differ rameters in order to obtain the optimal expansion plan
from the confidence level in which we are basing, depending v̂ (ι) and the worst-case operating cost ĉO,wc (v), (see
on the real joint probability distribution of the random param- Subsection II-C).
eters involved. 4) Level III evaluation of the quantile probability. Perform
Moreover, a further benefit of the above analogy is the a Monte Carlo simulation using the wind power and
possibility to gain knowledge about the quality of the solution demand scenarios given by Nataf model for the given
of the max-min subproblem by an out-of-sample assessment expansion plan v̂ (ι) and the worst-case operating cost

(ι)
relying on Monte Carlo simulation. Note that for a given ĉO,wc (v) . This step allows the calculation of a
reliability index β and once the optimal solution of problem better approximation of the probability of obtaining an
(2)–(8) is calculated and the actual realizations of power operating cost not exceeding ĉO,wc (v) given by (27).
capacities and loads are known, it is possible to calculate the O,min(ι) (ι)

5) Convergence checking. If |Prob cM C ≤ ĉO,wc −
minimum operational cost for each scenario by solving the α| ≤  the algorithm stops; otherwise, go to Step 6.
following optimization problem: 6) Update β. If ι = 1, the new value of β is equal to:
cO,min
MC = Minimize bT x, (26) √ SP (ι) − α
˜
x∈Ω(v̂,d) β (ι+1) = β (ι) − 2π 2 , (28)
e(−β (ι) /2)
where v̂ corresponds to the optimal expansion plan obtained (ι) (ι)

by solving problem (2)–(8) and d˜ represents wind power and where SP (ι) = Prob cO,min
MC ≤ ĉO,wc . Expression
demand scenarios from the Nataf transformation. (28) constitutes a Newton approximation with an addi-
If we sample ns scenarios, it is possible to determine the tional approximation of quantile probability derivative.
true probability of non exceeding the worst-case operating cost On the other hand, if ι > 1, the updating formula related
ĉO,wc by summing the number of times the optimal value of to the secant method is used:
problem (26) is lower than or equal to ĉO,wc (v) divided by SP (ι) − α
the number of sampled scenarios, i.e: β (ι+1) = β (ι) − . (29)
SP (ι) − SP (ι−1)
ns
!
1 if cO,min O,wc Increase the iteration counter ι ← ι + 1 and continue in
M C,i ≤ ĉ
X
0 if cO,min O,wc Step 3.
i=1 M C,i > ĉ
Prob (cO,min
MC ≤ ĉ O,wc
) = . V. N UMERICAL RESULTS
ns
(27) Results from a realistic case study based on a modified
Note that if ns → ∞, then the resulting probability con- version of the extended IEEE 118-bus system [22] are pre-
verges to exact value. Hereinafter, this probability is referred sented in this section in order to show the performance of
to as simulated probability. the proposed method. The case study comprises 118 buses,
Therefore, since the calculation of the simulated probability 54 conventional generators, 10 wind farms, 91 loads, and 180
in (27) associated with ĉO,wc depends on the reliability index existing transmission lines. Additionally, 72 existing lines can
β used within the FOSM approach, we aim at modifying this be duplicated to build additional lines with an investment
reliability index β so that the simulated probability matches budget equal to $20-million. All demand consumptions and
the desired one α. all generation capacities of wind farms are uncertain. System
In summary, for a given confidence level, for instance α = data and a figure depicting the 118-bus system can be found
0.9, we aim at obtaining the reliability index β required within in [23].
FOSM solution technique described in Subsection II-C, so that In order to present a realistic case study, we assume that the
the simulated probability obtained through (27) equals α. 10 wind farms considered correspond to real off-shore wind
It is all summed up in the following iterative procedure: farms in locations where we have at our disposal wind speed
data. To get wind power scenarios from available speed data
1) Input
we consider the use of a 5 MW turbine [24], whose power
• Select the confidence level α.
curve is defined by the following sigmoid function:
• Select the number of samples ns .
• Select the tolerance of the process . 
0 if ω < 3
• Select the required data for uncertain parameters

3
+t2 ω 2 −t3 ω+t4

nt t1 ω if 3 ≤ ω < 11.3

¯ variance-covariance matrix Σ and
(mean values d, PGmax (ω) = 1000 (30)
˜
Nataf model real scenarios d). 
 5nt if 11.3 ≤ ω < 25
0 if ω ≥ 25

2) Initialization. Set the iteration counter to ι = 1 and
select the first approximation of the reliability index β, where t1 , t2 , t3 and t4 are equal to 2.9524523, 11.48234,
β (1) = φ−1 (α). 77.0958543 and 164.5205987, respectively. Moreover nt rep-
3) Solution of problem (2)–(8) using FOSM approach at resents the number of turbines of each wind farm (nt : 6, 3,
iteration ι. Solve problem (2)–(8) using only first and 4, 8, 10, 3, 10, 4, 5 and 12).

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
 TABLE I final solution has been achieved.
E XTENDED IEEE 118- BUS SYSTEM –M EAN VALUES , STANDARD As can be observed, unlike FOSM (level II) decomposition
DEVIATIONS AND CORRELATION MATRIX OF MAXIMUM WIND FARM
POWER PRODUCTIONS method [15], the proposed iterative procedure based on level
Correlation matrix
III structural-reliability methods allows to determine the true
µG σG simulated probability.
Bus 23 24 26 27 31 100 82 96 73 75
(MW) (MW)
It must be stressed that this procedure makes possible to
23 498.116 380.815 1 0.2839 0.1804 0.0571 0.0315 0.0335 0.0033 0.0018 -0.0015 -0.0046
24 427.504 186.380 0.28389 1 0.0375 0.1390 0.0925 0.1152 0.0114 -0.0015 0.0075 -0.0159 optimize the required objective function quantile considering
26 457.241 171.859 0.1804 0.0375 1 0.0206 0.0101 -0.0020 0.0164 0.0076 -0.0127 -0.0084
27 569.653 413.698 0.0571 0.1390 0.0206 1 0.3342 0.0838 0.0030 0.0065 0.0007 0.0067 the probability distribution model that better fits data, without
31 1350.36 399.970 0.0315 0.0925 0.0101 0.3342 1 0.0833 0.0107 0.0052 -0.0043 -0.0043 restrictions on which model is chosen. Nevertheless, we have
100 401.954 153.185 0.0335 0.1152 -0.0021 0.0838 0.0833 1 0.0002 0.0192 0.0215 0.0112
82 475.670 274.519 0.0033 0.0114 0.0164 0.0030 0.0107 0.0002 1 0.0226 0.0295 0.0134 proposed an specific model that allows the consideration of
96 159.405 89.769 0.0018 -0.0015 0.0076 0.0065 0.0052 0.0192 0.0226 1 -0.0201 -0.0038
73 113.972 85.581 -0.0015 0.0075 -0.0127 0.0007 -0.0043 0.0215 0.0295 -0.0201 1 0.2411 marginal distributions and correlations among variables.
75 360.125 146.934 -0.0046 -0.0159 -0.0084 0.0067 -0.0043 0.0112 0.0134 -0.0038 0.2411 1

VI. C ONCLUSION
TABLE II This paper extends the use of structural-reliability and
I TERATIVE S OLUTION P ROCESS OF THE P ROPOSED A LGORITHM FOR THE
EXTENDED IEEE 118- BUS SYSTEM
decomposition techniques to solve the transmission network
expansion planning problem. It basically transforms Level II
Scale worst-case Investment
ι operating costs costs β (ι) SP (ι) Error
Time [15] into a Level III, which allows the consideration of not
(s) only the first two moments of the probability distribution of
(M$) (M$)
1 1774.1298 12.0740 1.2816 0.8411 0.0589 9207 the uncertain parameters involved, but also their complete
2 1829.5206 27.4420 1.6172 0.8851 0.0149 10446 probability structure.
3 1848.5538 31.0770 1.7308 0.8989 0.0011 11521 The proposed method allows to minimize a given quantile
4 1850.3159 31.0770 1.7379 0.8997 0.0003 11459 of uncertain operational and investment costs. Besides, the use
5 1850.9767 31.0770 1.7433 0.8999 0.0001 8172 of Monte Carlo simulation method allows to obtain the density
(ι) (ι)

Error = |Prob cO,min
MC ≤ ĉO,wc − α| function associated with cost, which provides an interpretation
of the robust solution easy to understand.
An example with a network of 118 buses have been solved.
Mean or forecast values, standard deviation values and However, the sustainability of the proposed method for real-
correlation coefficients of maximum power productions of life large-scale systems is subject to the development of tech-
wind generating units are given in Table I. In addition, a niques to accelerate the method or to reduce the computational
correlation coefficient equal to 0.5 is used for demands. time.
The desired confidence level is equal to α = 0.9 and a From numerical results, the following conclusions are
probability tolerance of  = 1.5 · 10−3 has been assumed. drawn:
The calculation of the true probability of non exceeding the 1) The consideration of only the first two moments of
worst-case operating cost ĉO,wc is done by performing a the probability distribution of the uncertain parameters
Monte Carlo simulation of 10000 samples. Simulations have yields a lower confidence level with respect to the
been implemented and solved on a Dell PowerEdge R920X64 desired one.
with four Intel R
Xeon R
E7-4820 processors at 2.00 GHz 2) Modifying appropiately the reliability index β allows
and 768 GB of RAM using CPLEX 12.6 under GAMS 24.2 to obtain a simulated probability almost equal to the
[25]. Convergence tolerances IL and OL , and the optimality required one.
tolerance for the branch-and-cut algorithm of CPLEX were set 3) Incorporating the complete probability structure of un-
at 10−8 . Results provided in Table II are obtained. Moreover, certain parameters is relevant in terms of investment
computational times for each iteration are given in the last decisions.
column of Table II.
In summary, this paper proposed a novel iterative method
The solution associated with iteration 1 corresponds to
to extend the use of robust optimization for solving the
FOSM (level II) decomposition method. Note that in iter-
transmission network expansion planning problem for any
ation 1, the true probability of not exceeding the worst-
type of probability distribution related to uncertain parameters.
case operating cost ĉO,wc is lower than the desired one, i.e.
Although this paper only consider the uncertainty associated
0.8411 < 0.90. Thus, the reliability index is increased for the
to wind generation and demands, the proposed method can be
next iteration using expression (28), i.e. 1.2816 → 1.6172.
extended to other sources of uncertainty, e.g. elastic load or
However, this growth in the reliability index is not enough,
another types of renewable energy, in a straightforward and
and the simulated probability of exceedance is still lower than
simple way.
the required value, i.e. 0.8851 < 0.90, thus the reliability index
is updated again using the secant formula (29). Note that in the R EFERENCES
last iteration (ι = 5), the simulated probability is almost equal
[1] S. Lumbreras and A. Ramos, “The new challenges to transmission
to the required one, and the difference between the obtained expansion planning. Survey of recent practice and literature review,”
value and the desired one equals the required tolerance. The Electr. Pow. Syst. Res., vol. 134, pp. 19–29, May 2016.

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020
[2] J. M. Morales, L. Baringo, A. J. Conejo, and R. Mı́nguez, “Probabilistic IEEE Trans. Power Syst., vol. 29, no. 6, pp. 3069–3077, Nov. 2014.
power flow with correlated wind sources,” IET Gener. Transm. Distrib., [13] C. Ruiz and A. J. Conejo, “Robust transmission expansion planning,”
vol. 4, no. 5, pp. 641–651, May 2010. Eur. J. Oper. Res., vol. 242, no. 2, pp. 390–401, Apr. 2015.
[3] L. Xie, P. M. S. Carvalho, L. A. F. M. Ferreira, J. Liu, B. H. Krogh, [14] R. Mı́nguez and R. Garcı́a-Bertrand, “Robust transmission network
N. Popli, and M. D. Ilić, “Wind integration in power systems: Operational expansion planning in energy systems: Improving computational perfor-
challenges and possible solutions,” Proc. IEEE, vol. 99, no. 1, pp. 214– mance,” Eur. J. Oper. Res., vol. 248, no. 1, pp. 21–32, Jan. 2016.
232, Jan. 2011. [15] C. Roldán, R. Mı́nguez, R. Garcı́a-Bertrand and J. M. Arroyo, “Robust
[4] W. P. Bell, P. Wild, J. Foster, and M. Hewson, “Wind speed and electricity transmission network expansion planning under correlated uncertainty,”
demand correlation analysis in the Australian National Electricity Market: IEEE Trans. Power Syst., vol. 34, no. 3, pp. 2071–2082, May 2019.
Determining wind turbine generators’ ability to meet electricity demand [16] A. Thiele, T. Terry, and M. Epelman, “Robust linear optimization with
without energy storage,” Econ. Anal. Policy, vol. 48, pp. 182–191, Dec. recourse,” Univerity of Michigan, IOE Technical Report TR09-01, 2010.
2015. [17] D. Bertsimas, E. Litvinov, A. Sun, J. Zhao, and T. Zheng, “Adaptative
[5] M. Carrión, J. M. Arroyo, and N. Alguacil, “Vulnerability-constrained robust optimization for the security constrined unit commitment ,” IEEE
transmission expansion planning: A stochastic programming approach,” Trans. Power Syst., vol. 28, no. 1, pp. 52–63, Feb. 2013.
IEEE Trans. Power Syst., vol. 22, no. 4, pp. 1436–1445, Nov. 2007. [18] A. Ben-Tal and A. Nemirovski, “Robust convex optimization,” Math.
[6] J. A. López, K. Ponnambalam, and V. H. Quintana, “Generation and Oper. Res., vol. 23, no. 4, pp. 769–805, Nov. 1998.
transmission expansion under risk using stochastic programming,” IEEE [19] R. Mı́nguez, A. J. Conejo, and R. Garcı́a-Bertrand, “Reliability and de-
Trans. Power Syst., vol. 22, no. 3, pp. 1369–1378, Aug. 2007. composition techniques to solve certain class of stochastic programming
[7] L. P. Garcés, A. J. Conejo, R. Garcı́a-Bertrand, and R. Romero, “A problems,” Reliab. Eng. Syst. Saf., vol. 96, no. 2, pp. 314–323, Feb. 2011.
bi-level approach to transmission expansion planning within a market [20] A. Nataf, “Détermination des distribution de probabilités dont les marges
environment,” IEEE Trans. Power Syst., vol. 24, no. 3, pp. 1513–1522, sont données,” Comptes Rendus de l’Academie des Sciences, vol. 225, pp.
Aug. 2009. 42–43, 1962.
[8] J. H. Roh, M. Shahidehpour, and L. Wu, “Market-based generation and [21] J. M. Morales, R. Mı́nguez, and A. J. Conejo, “A methodology to
transmission planning with uncertainties,” IEEE Trans. Power Syst., vol. generate statiscally dependent wind speed scenarios,” Applied Energy,
24, no. 3, pp. 1587–1598, Aug. 2009. vol. 87, no. 3, pp. 843–855, March 2010.
[9] P. Wu, H. Cheng, and J. Xing, “The interval minimum load cutting [22] I. Peña, C. B. Martinez-Anido, and B.-M. Hodge, “An extended IEEE
problem in the process of transmission network expansion planning 118-bus test system with high renewable penetration,” IEEE Trans. Power
considering uncertainty in demand,” IEEE Trans. Power Syst., vol. 23, Syst., vol. 33, no. 1, pp. 281–289, Jan. 2018.
no. 3, pp. 1497–1506, Aug. 2008. [23] C. Roldán, R. Mı́nguez, and R. Garcı́a-Bertrand, “Robust transmission
[10] H. Yu, Y. Chung, and K. P. Wong, “Robust transmission network expansion planning with uncertain generations and loads using Full
expansion planning method with taguchi’s orthogonal array testing,” IEEE Probabilistic Information,” [Online]. Available: https://drive.google.com/
Trans. Power Syst., vol. 26, no. 3, pp. 1573–1580, Aug. 2011. open?id=1i1gKhTnAv4zrRT38PgXmVyMjH7sao9YX
[11] R. A. Jabr, “Robust transmission network expansion planning with [24] J. Jonkman, S. Butterfield, W. Musial, and G. Scott, “Definition of a 5
uncertain renewable generation and loads,” IEEE Trans. Power Syst., vol. MW reference wind turbine for off-shore system development,” National
28, no. 4, pp. 4558–4567, Nov. 2013. Renewable Energy Laboratory, Technical Report, Feb. 2009.
[12] B. Chen, J. Wang, L. Wang, Y. He, and Z. Wang, “Robust optimization [25] The GAMS Development Corporation website, 2018. [Online]. Avail-
for transmission expansion planning: Minimax cost vs. minimax regret,” able: http://www.gams.com

21st Power Systems Computation Conference Porto, Portugal — June 29 – July 3, 2020
PSCC 2020