Received 16 May 2022, accepted 22 June 2022, date of publication 30 June 2022, date of current version 8 July 2022.

Digital Object Identifier 10.1109/ACCESS.2022.3187403

Evaluation of Numerical Methods for TSCOPF in a

Large Interconnected System
MOHAMMADAMIN AGHAHASSANI , EDGARDO D. CASTRONUOVO , (Senior Member, IEEE),
PABLO LEDESMA , AND FRANCISCO ARREDONDO
Department of Electrical Engineering, Universidad Carlos III de Madrid, 28911 Leganés, Spain
Corresponding author: Mohammadamin Aghahassani ([email protected])
This work was supported by the Spanish Agencia Estatal de Investigación under Project PID2019-104449RB-I00 and Project
AEI/10.13039/501100011033.

ABSTRACT Transient stability-constrained optimal power flow (TSCOPF) models comprehensively

analyze the security and economic operation of power systems. However, they require a high computational
effort and can suffer from convergence problems when applied to large systems. This study analyzes
the performance of eleven numerical integration algorithms applied to ordinary differential equations that
represent power system dynamics in a TSCOPF model. The analyzed algorithms cover a range of explicit
and implicit methods, including the recently published semi-explicit and semi-implicit Adams-Bashforth-
Moulton formulas, together with several initialization techniques. The integration methods are applied to a
model of the Iberian Peninsula power system, and their performance is discussed in terms of convergence,
accuracy, and computational effort. The results show that most implicit methods converge to the solution,
even for large time steps. In particular, the Adams-Moulton method of order two and Simpson’s rule, both
initialized with RK4, outperform the trapezoidal rule, which is the default method in TSCOPF models.

INDEX TERMS Numerical methods, optimal power flow, power system stability, transient stability,
TSCOPF.

NOMENCLATURE
z Generation cost (in eh ).
Pm,i Mechanical input power of the
Ng , Nb Number of synchronous generators and
synchronous generator at bus i
buses, respectively.
(in p.u.).
ci Cost coefficient of the synchronous gener-
e Vi,0 , αi,0 Steady state voltage magnitude and angle
ator at bus i (in MWh ).
at bus i (in p.u. and rad, respectively).
Pg,i,0 , Qg,i,0 Steady-state active and reactive power
Vi,t , αi,t Voltage magnitude and angle at
generated by the synchronous generator at
bus i and time t (in p.u. and rad,
bus i (in p.u.).
respectively).
Pg,i,t , Qg,i,t Active and reactive power generated by the
Ibr Steady state current of branch br (in p.u.).
synchronous generator at bus i and time t
Yi,j , θi,j Magnitude and angle of the element (i, j)
(in p.u.).
in the bus admittance matrix Y (in p.u. and
Pd,i,t , Qd,i,t Active and reactive load at bus i and time t
rad, respectively).
(in p.u.).
Pe,i,t , Qe,i,t Electric output active and reactive power in YL,br,j , θL,br,j Magnitude and angle of the element (br, j)
the rotor of the synchronous generator at in the line admittance matrix YL (in p.u. and
bus i and time t (in p.u.). rad, respectively).
δi,t Rotor angle of synchronous generator i at
The associate editor coordinating the review of this manuscript and time t in synchronously rotating reference
approving it for publication was Siqi Bu . frame (in rad).

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

70562 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 10, 2022
M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

δCOI ,t The rotor angle corresponds to the center algebraic equations using a numerical integration algorithm.
of inertia (COI) at time t (in rad). They are then added to the power flow equations. The
ωs Synchronous rotor speed (in rad resulting optimization problem is then solved using an NLP
s ).
1ωi,0 Steady-state rotor speed deviation of the solver. This technique is robust and can handle unstable
synchronous generator at bus i (in p.u.). cases; however, it becomes computationally demanding when
1ωi,t Rotor speed deviation of the synchronous applied to large power systems.
generator at bus i and time t (in p.u.).
0
xd,i Direct axis transient reactance of the syn- A. LITERATURE REVIEW
chronous generator at bus i (in p.u.). Although there are different families of numerical methods
EI0 ,i Internal voltage of the synchronous gener- for integrating differential equations [14]–[22], the trape-
ator at bus i (in p.u.). zoidal rule is practically the only method used in TSCOPF
Hi Inertia constant of the synchronous gener- studies based on simultaneous discretization [2], [8], [9],
ator at bus i (in s). [11], [14], [23], [24], because it is easy to program and
Di Damping factor of the synchronous gener- numerically stable. To the knowledge of the authors, there
ator at bus i (in p.u.). are no previous studies on the performance of numerical
V min , V max Lower and upper limits of the bus voltage methods in TSCOPF models, except for [7] and [25], which
magnitude (in p.u.). investigated a limited number of methods in a relatively small
max
Ibr Limit of the current in branch br (in p.u.). system. This study explores a broader range of integration
Pmax Limit of the active power generated by the methods to reduce the computation time in large-scale
g,i
synchronous generator at bus i (in p.u.). TSCOPF models.
Qmin min Numerical methods can be classified as explicit or implicit.
g,i , Qg,i Lower and upper limits of the reactive
power generated by the synchronous gen- The explicit algorithms use known quantities from past steps
erator at bus i (in p.u.). to directly obtain the solution to the current step, whereas
δ max Limit of the rotor angle (in rad). the implicit methods require current and prior steps for
1t Integration time step (in s). the calculation. Most commercial power system simulators
use explicit numerical methods because they are easier to
implement. PowerWorld, for instance, uses second-order
I. INTRODUCTION Runge-Kutta [26], whereas PSS/E and PSLF use second-
Transient stability-constrained optimal power order Adams-Bashforth [27]. The implicit methods are
flow (TSCOPF) is an optimization problem that offers numerically more stable and better suited to stiff systems
a comprehensive approach to power system optimization than the explicit methods, but this comes at the cost
by simultaneously addressing economic and operational of solving nonlinear equations at each step, resulting in
objectives while considering both static and dynamic stability higher computational time. In general, there should be a
constraints. TSCOPF has recently received increasing trade-off between numerical stability and computational cost.
attention [1], [2] because system operators are forced to Predictor-corrector methods combine some advantages of
operate power systems close to their operational limits owing implicit and explicit methods. A predictor-corrector method
to the high penetration of non-programmable renewable uses two sets of equations, one for prediction and the other
generation, liberalized electricity markets, and environmental for correction. In [21] and [22], the conventional predictor-
restrictions. corrector Adams-Bashforth-Moulton (ABM) method is mod-
The size and complexity of interconnected power sys- ified to construct semi-explicit and semi-implicit ABMs,
tems make large-scale TSCOPF problems challenging, with improving the computation accuracy.
several authors following different approaches [3], [4]. One Numerical methods can also be classified based on the
approach, followed in [5], formulates TSCOPF as a nonlinear number of steps required for each iteration. Single-step
programming problem (NLP) in which, the dynamic behavior methods are based solely on the previous step; for example,
of the power system is obtained using PSS/E, an external the Runge-Kutta fourth-order (RK4) method has an accuracy
power system simulator independent of the optimization of order four but at the cost of significant computation time
problem. In this sequential technique, the generation dispatch because four function evaluations are required at each time
changes at each iteration based on simulation results and step [28], [29]. As the computation time is a critical issue in
constraint violations. This method has the problem of falling TSCOPF models, it is preferable to use multistep methods
into numerical instabilities and convergence problems when that require only one function evaluation at each step and
simulating transiently unstable cases. improve the accuracy by using data from several previous
A more common approach followed in this work is to steps. The number of steps in this study is limited to two
use a simultaneous discretization algorithm [6] that directly because, while increasing the number of steps improves
incorporates the dynamic simulation in the optimization accuracy, it also threatens numerical stability [30].
problem. In [2], [7]–[13], differential equations describing Moreover, two-step methods require two successive start-
power system dynamics are transformed into corresponding ing points for the initialization. The initial condition provides

VOLUME 10, 2022 70563

 M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

Nb
the first starting point, and the second starting point can X Hi δi,t
be calculated using another one-step method, such as Euler, δCOI ,t = (10)
Hi
i=1
trapezoidal rule, or RK4. Therefore, it is important to consider
Nb
the error incurred when approximating this extra initial point X
Ibr = YL,br,j Vj,0 [cos(θL,br,j + αj,0 )
because it may affect the final solution.
j=1
+ jsin(θL,br,j + αj,0 )] (11)
B. CONTRIBUTION
The main contribution of this work is a comprehensive V min ≤ Vi,0 ≤ V max (12)
max
analysis of numerical methods in an optimization model |Ibr | ≤ Ibr (13)
based on simultaneous discretization. The model is solved 0 ≤ Pg,i,0 ≤ Pmax
g,i (14)
using a standard interior-point algorithm, which is the main
Qmin max
g,i ≤ Qg,i,0 ≤ Qg,i (15)
approach in TSCOPF. The results show that alternative
methods can outperform the trapezoidal rule in TSCOPF −δ max ≤ δi,t − δCOI ,t ≤ δ max (16)
applications. The performance of each method is tested on Table 1 summarizes the physical meaning of each equation.
a 216-buses, 75-generators model of the Iberian Peninsula Wind power plants are modeled as a fixed generation with
system, including Portugal and Spain, in which six severe a power factor of one and a cost of zero. The transient
contingencies are evaluated. These methods are discussed stability limit (16) is determined as the maximum rotor angle
in terms of accuracy, computational time, and numerical deviation from the center of inertia (COI), as is customary
stability. Additionally, the effect of initialization on the in TSCOPF studies [11], [13]. Equations (1)-(16) represent a
accuracy of two-step methods is analyzed. The result is a set typical TSCOPF model based on simultaneous discretization
of suitable methods that can serve as better alternatives to the that provides the optimal dispatch with steady-state and
trapezoidal rule in future TSCOPF studies. transient stability limits after a severe fault. When the solution
The remainder of this paper is organized as follows. provided by the OPF becomes unstable, the stability limit
Section II develops the optimization model. Section III imposed by (16) modifies the OPF generation dispatch,
describes several numerical methods applied to the TSCOPF thereby increasing the cost of obtaining a secure operation.
model. Section IV introduces a case study. Section V presents In the TSCOPF model, differential equations (7) and (8)
simulation results. Finally, conclusions are presented in are discretized using a numerical algorithm and solved
Section VI. simultaneously with (1)-(6) and (9)-(16) using an NLP solver.
However, in conventional transient stability simulations, (7)
II. OPTIMIZATION MODEL and (8) are discretized and solved sequentially. TSCOPF
The optimization problem consists of an objective function provides an optimal dispatch together with the simulation
minimizing the generation cost (1) and a set of nonlinear of the fault included in the model; this simulation coincides
equality and inequality constraints (2)-(16): with a conventional transient stability simulation that takes
the optimal dispatch as the initial point.
Ng Fig. 1 shows the procedure used to build and solve
X the TSCOPF model. A Python program reads all relevant
z = min ci Pg,i,0 (1)
data, builds the pre-fault, fault, and post-fault admittance
i=1
subjected to matrices for each contingency, and writes the TSCOPF
Nb
model containing (1)-(16) in the GAMS modeling language.
X In this work, at the block marked with an arrow in Fig. 1,
Pg,i,t − Pd,i,t = Vi,t Vj,t Yi,j cos(αi,t − αj,t − θi,j )
different numerical integration methods are used to discretize
j=1
differential equations (7) and (8). For example, applying the
(2)
Nb
trapezoidal rule, which is the default method in TSCOPF,
X to (7) and (8) yields:
Qg,i,t − Qd,i,t = Vi,t Vj,t Yi,j sin(αi,t − αj,t − θi,j )
ωs 1t
j=1 δi,n − δi,n−1 = (1ωi,n + 1ωi,n−1 ), (17)
(3) 2
Di 1t Di 1t
  
0
Pe,i,t xd,i = EI0 ,i Vi,t sin(δi,t − αi,t ) (4) 1+ 1ωi,n − 1 − 1ωi,n−1
4Hi 4Hi
0
Qe,i,t xd,i = (EI0 ,i )2 − EI0 ,i Vi,t cos(δi,t − αi,t ) (5) 1t
= (Pa,i,n + Pa,i,n−1 ). (18)
Pa,i,t = Pm,i − Pe,i,t (6) 4H
dδi,t The complete TSCOPF model is then solved in GAMS
= ωs 1ωi,t (7)
dt using the interior-point optimizer IPOPT, which is an NLP
d1ωi,t 1 solver suited for large-scale nonlinear optimization problems.
= (Pa,i,t − Di 1ωi,t ) (8)
dt 2Hi The model is managed directly by IPOPT, and there is no need
1ωi,0 = 0 (9) for a power system solver outside the GAMS solver.

70564 VOLUME 10, 2022

M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

TABLE 1. Optimization model constraints. system dynamics can be expressed as:

dx
= f(x, t) with x(t0 ) = x0 , (19)
dt
where x is the vector of state variables, x0 is the vector of
initial conditions at t = t0 , and f(x, t) is a known function.

A. TWO-STEP METHODS
Most numerical methods described in this section can be
defined using the following generic expression:
α0 xn + α1 xn−1 + α2 xn−2
= 1t(β0 fn + β1 fn−1 + β2 fn−2 ), (20)
where xn is the vector of numerical approximations of exact
solutions x(tn ) at time tn = t0 + n1t; αi and βi are the
coefficients representing different numerical methods; and
fn = f(x(tn ), tn ). The application of (20) to (7) and (8) results
in:
α0 δi,n + α1 δi,n−1 + α2 δi,n−2
= 1tωs (β0 1ωi,n + β1 1ωi,n−1 + β2 1ωi,n−2 ), (21)
and
α0 1ωi,n + α1 1ωi,n−1 + α2 1ωi,n−2
1t
= [β0 (Pa,i,n − Di 1ωi,n )
2Hi
+ β1 (Pa,i,n−1 − Di 1ωi,n−1 )
+ β2 (Pa,i,n−2 − Di 1ωi,n−2 )]. (22)
Table 2 lists the names and main characteristics of some of
the analyzed methods. A discussion about the accuracy and
numerical stability of these methods can be found in [31]. One
method does not have a unique name and is thus referred to
as method A. The specific equation of each method in the
TSCOPF model can be found by substituting parameters αi
and βi in (21) and (22). For example, the Adams-Moulton
method of order two can be formulated as:

5 8
δi,n − δi,n−1 = ωs 1t 1ωi,n + 1ωi,n−1
12 12

1
− 1ωi,n−2 , (23)
12
and
1t

5
1ωi,n − 1ωi,n−1 = (Pa,i,n − Di 1ωi,n )
2Hi 12
8
FIGURE 1. Flowchart showing the implementation of the TSCOPF model. + (Pa,i,n−1 − Di 1ωi,n−1 )
12 
1
− (Pa,i,n−2 − Di 1ωi,n−2 ) . (24)
III. NUMERICAL METHODS 12
A set of representative numerical methods for the solution of The trapezoidal rule is the only one-step method in Table 2;
ordinary differential equations (ODEs) (7) and (8) is selected the rest are two-step algorithms. One-step numerical methods
to evaluate their performance in the TSCOPF model. These (if α2 = β2 = 0 in (20)) approximate xn using xn−1 , the
methods correspond to some of the most commonly used two- value obtained in the previous step. The starting point x0 is
step algorithms [31] and newly constructed semi-explicit and calculated from initial conditions, typically solving a power
semi-implicit methods. The ODEs representing the power flow. Two-step methods, on the other hand, require xn−1 and

VOLUME 10, 2022 70565

 M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

TABLE 2. Investigated numerical methods [31].

1t

5 p p
xn−2 in the previous two steps. Although the value of x0 for 1ωi,n − 1ωi,n−1 = (P − Di 1ωi,n )
the initial step is calculated from initial conditions, finding 2Hi 12 a,i,n
the value of x1 in a two-step method requires another one-step 8
+ (Pa,i,n−1 − Di 1ωi,n−1 )
method such as Euler, trapezoidal rule or RK4. 12 
Two of the methods in Table 2 are explicit, which means 1
− (Pa,i,n−2 − Di 1ωi,n−2 )
that β0 = 0 and xn can be found explicitly from previous time 12
steps. Explicit two-step methods are not absolutely stable (A- (25)
stable); however, Adams-Bashforh is a zero-stable explicit
method with the maximum possible order [31]. The rest of the The SEABM method modifies the last equation by
p
methods in Table 2 are implicit methods, which means that substituting Pa,i,n with Pa,i,n , and the SIABM method
p
xn cannot be directly expressed in terms of known quantities additionally substitutes 1ωi,n with 1ωi,n . A discussion on
from the previous steps because β0 6 = 0. the numerical properties of ABM, SEABM, and SIABM can
The methods with the same order of accuracy can produce be found in [22] and [32].
different levels of error, depending on the error constant.
Adams-Moulton with order three gets closer to the exact C. MERGED TRAPEZOIDAL RULE
solution than Adams-Bashforth with order two but at the In addition to the methods listed in Table 2 and the
cost of solving a nonlinear equation at each step. However, predictor-corrector methods, a method called the merged
the Adams-Moulton is not the highest-order possible implicit trapezoidal rule (MTR), is analyzed. The MTR is obtained
method with zero-stability. Simpson’s rule with order four by merging (17) and (18) and eliminating the speed, which
reaches the highest order because of its symmetrical structure. yields:
Finally, the trapezoidal rule has the highest possible order Di 1t Di 1t
among A-stable implicit methods, with the lowest error (1 + )δi,n − 2δi,n−1 + (1 − )δi,n−2
4Hi 4Hi
constant. Since the absolute stability is difficult to obtain, the ωs 1t 2
condition can be lowered to A0 -stability [31]. = (Pa,i,n + 2Pa,i,n−1 + Pa,i,n−2 ). (26)
8Hi
B. PREDICTOR-CORRECTOR METHODS where the rotor angle deviation δi,n is expressed as a function
Predictor-corrector methods proceed in two stages. The first of the two previous steps. As 1t is squared, MTR does not
stage typically uses an explicit method to extrapolate the follow the general form of the two-step methods expressed
value at the next point, and the second stage uses an implicit by (20).
method to refine the initial approximation.
Three predictor-corrector methods are tested in this IV. CASE STUDY
work: the Adams-Bashforth-Moulton (ABM) method and The numerical methods in Table 2, along with ABM,
its two variants, the semi-implicit and semi-explicit Adams- SEABM, SIABM, and MTR, are implemented in the
Bashforth-Moulton methods (SIABM and SEABM, respec- TSCOPF model described in Section II and applied to a case
tively). The ABM method applied to (7) and (8) yields: representing the Iberian Peninsula transmission system. The
 
p 3 1 case is based on the Bialek European model [33], which can
δi,n − δi,n−1 = ωs 1t 1ωi,n−1 − 1ωi,n−2 ,
2 2 be downloaded from the link provided in [34].
1t 3 Fig. 2 shows the single-line diagram of the studied system,

p
1ωi,n − 1ωi,n−1 = (Pa,i,n−1 − Di 1ωi,n−1 ) which includes Spain and Portugal and contains 216 buses,
2Hi 2
 368 transmission lines, 75 generating units, and 143 loads
1
− (Pa,i,n−2 − Di 1ωi,n−2 ) , consuming 50.2 GW. The power systems of the rest of
2 continental Europe and Morocco are each reduced to a single

5 p 8 equivalent bus connected to a synchronous generator/load.
δi,n − δi,n−1 = ωs 1t 1ωi,n + 1ωi,n−1
12 12 In this study, the Iberian Peninsula exports 700 MW to France

1 and 497 MW to Morocco before the fault. The voltage limits
− 1ωi,n−2 ,
12 in (12) are set at a minimum of 0.9 p.u. and a maximum of

70566 VOLUME 10, 2022

M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

FIGURE 3. Synchronous generator rotor angles. Fault 1.

FIGURE 2. Single-line diagram of the Iberian Peninsula transmission
system model.

a 3.60 GHz processor and 24.0 GB RAM. The resulting

1.05 p.u. The maximum rotor angle deviation with respect
nonlinear optimization model contains 141,868 variables
to the COI in (16) is 110◦ . The complete data, including the
and 167,181 equality and inequality constraints using the
parameters of the power system and the static and dynamic
trapezoidal rule and a 10 ms time step.
constraints, can be found as supplementary material to this
document.
The original case was modified to accommodate a total V. RESULTS
capacity of 9380 MW of wind power. Before the fault, the This section compares the performance of the seven methods
wind power production was 2345 MW, which amounts to listed in Table 2, ABM, SEABM, SIABM, and MTR,
25 % of the total wind capacity and 5.07 % of the total when applied to differential equations (7) and (8) in the
generation. It has also been modified to account for the TSCOPF model. The trapezoidal rule, which is the numerical
reactive power consumption in the loads. Shunt capacitors method used by default in TSCOPF based on simultaneous
were added to stabilize the voltages and reduce the reactive discretization, is highlighted in black in the figures to serve
power flow in the lines. Six different faults are analyzed. as a term of comparison. The results for Fault 1 in Fig. 2 are
The first four faults are modeled as a short-circuit in a discussed in detail in Subsections V-A to V-C. The remaining
transmission line near a substation and cleared after 200 ms faults are summarized in Subsection V-D.
by disconnecting the affected line. Short-circuits are the most Fig. 3 shows an example of the solution of the TSCOPF
challenging faults to analyze in transient stability studies. model in the case of Fault 1. The dispatch provided by
However, other types of faults, such as a generator outage TSCOPF ensures transient stability by restricting the angle
and the separation of the Iberian Peninsula from the rest of deviation of each synchronous generator with respect to the
continental Europe, are also explored in this study. The fault COI. Therefore, the total generation cost of the dispatch
locations are marked in Fig. 2. Each fault represents a relevant provided by TSCOPF is higher than the cost obtained with an
incident on the power system of the Iberian Peninsula. OPF and this difference can be seen as the cost of ensuring
• Fault 1 is near a large nuclear power plant (2000 MW) transient stability.
operating at full load. Fig. 4 shows the increase in the TSCOPF total production
• Fault 2 is close to one of the tie-lines connecting the cost when different numerical methods and time steps
Iberian Peninsula to the rest of continental Europe. are applied to the optimization model. Unless otherwise
• Fault 3 is next to the tie-line connecting the Iberian specified, the trapezoidal rule is used to initialize two-step
Peninsula and Morocco. methods. The solutions tend to converge at approximately
• Fault 4 is close to a critical substation called La Mudarra, 2.4% as the time step decreases. Most of the results are
which connects several lines carrying power from power reliable with a time step of 10 ms, which is a common value
plants to the northwest to loads in the center, where the in transient stability studies.
capital city Madrid is located. On the other hand, as the time step increases, the transient
• Fault 5 is the loss of the largest operating power plant stability cost decreases, indicating that the solution is not
(2580 MW) in Portugal. only less accurate but also less secure. This is because some
• Fault 6 is the trip of the three tie-lines connecting the economic power plants that are more prone to causing a loss
Iberian Peninsula to the rest of continental Europe. of synchronism tend to increase their production. As a result,
The model is written in GAMS [35] and solved using the large time steps might result in operating points that can be
IPOPT solver [36]–[38] on a 64-bit personal computer with transiently unstable during real operations.

VOLUME 10, 2022 70567

 M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

FIGURE 4. Increase in the TSCOPF generation cost compared to the OPF. FIGURE 5. Rotor angles of three synchronous generators near the faulted
Fault 1. bus. Fault 1, using method AB(2) with time step 20 ms.

A. TWO-STEP IMPLICIT METHODS

Fig. 4 shows that TSCOPF with the BDF(2) method fails to
converge for small time steps.
The trapezoidal rule, Simpson’s rule, AM(2), ABM,
SEABM, SIABM, and, to a lesser extent, method A produce
good results in terms of accuracy and convergence, and are
deemed suitable for TSCOPF studies. MTR, introduced in
Section III-C, provides similar results to the trapezoidal rule
for 20 ms and larger time steps, but it does not converge for
smaller time steps.

B. TWO-STEP EXPLICIT METHODS

The relatively poor properties of the explicit methods
compared with the implicit ones in terms of numerical FIGURE 6. Rotor angles of three synchronous generators near the faulted
stability are shown in Figs. 4, 5, and 6. Fig. 4 shows that bus. Fault 1, using method AB(2) with time step 25 ms.
explicit methods AB(2) and midpoint rule converge with
only small time steps. To illustrate this point, Figs. 5 and 6
show the rotor angles of the three generators near the faulted and variables by 7.9 % and 9.25 %, respectively. ABM takes
bus provided by the TSCOPF solution using method AB(2). substantially longer on the CPU than the other methods
As the contingency is the same, the solution should be very because two extra equations are solved at each step. The
similar. However, as the time step increases from 20 to 25 ms, SEABM and SIABM methods can save computation time
the rotor angles of the two generators exhibit increasing compared with the ABM by using already computed values
oscillations. This is problematic because the solution in Fig. 6 instead of predicting all variables. Therefore, SEABM and
can be considered valid if the results are not further inspected. SIABM have fewer equations in their prediction stages than
It is noteworthy that these types of false solutions can be the ABM. However, the CPU times of SEABM and SIABM
detected and removed by introducing additional constraints in in TSCOPF studies are still much higher than that of AM(2),
the TSCOPF model. On the other hand, when the time step is indicating that they are not yet appropriate substitutes for
sufficiently small, and AB(2) and the midpoint rule converge, AM(2). It should be noted that all the two-step methods in
they provide more conservative results than the other methods Fig. 7 are initialized using the trapezoidal rule.
because their error constants are positive,as shown in Table 2.
Fig. 7 shows the CPU time required to converge to the C. EFFECT OF THE INITIALIZATION METHOD ON
solution. All the methods in this figure require more CPU TWO-STEPS METHODS
time as the time step decreases because the number of Different initialization methods can be used to determine the
equations and variables increases in the TSCOPF model. value of x1 in (20) in two-step algorithms. This section applies
It can be observed that the trapezoidal rule provides the best three different initialization methods – Euler’s method,
performance in terms of CPU time. It can also be observed trapezoidal rule, and RK4 – to the most promising methods,
that MTR does not save much CPU time compared to the according to the previous sections. Fig. 8 shows the increase
trapezoidal rule, despite reducing the number of equations in TSCOPF cost as a percentage of the OPF in each case.

70568 VOLUME 10, 2022

M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

FIGURE 7. The CPU time for numerical methods to obtain the convergent FIGURE 9. CPU time vs absolute value of the convergence error for
solution. Fault 1. different methods and time steps; Fault 1. The methods in the shaded
area show the best performance.

TABLE 3. Contingencies.

FIGURE 8. Increase in the TSCOPF generation cost compared to the OPF.

Fault 1, using different initialization methods.

The RK4 initialization offers the most accurate results,

whereas Euler initialization provides the least accurate
results. With the RK4 initialization, the two-step methods
AM(2), Simpson’s rule and method A produce better results
than the trapezoidal rule. The highest accuracy for time steps
between 8 ms and 40 ms is given by AM(2), starting with
RK4.
In addition to the sensitivity of the numerical methods
to the time step, the total CPU time is another important
FIGURE 10. Synchronous generator rotor angles. Fault 3.
factor to consider. Fig. 9 displays the CPU time vs. the
absolute value of the convergence error for methods with
different initializations. Each point in the figure corresponds time steps of 25 and 20 ms, and Simpson’s rule with RK4
to a different time step, from 8 ms to 25 ms. To calculate initialization and a time step of 20 ms.
the absolute error, the generation cost of each method s is
compared with the generation cost of a reference solution sref , D. OTHER CONTINGENCIES
obtained with AM(2), RK4 initialization and a time step of Similar results are obtained when different numerical meth-
8 ms: ods are applied to other contingencies. Table 3 shows the
s − sref increase in the TSCOPF total production cost, CPU time,
absolute error = × 100% (27) and the number of iterations obtained by the Adams-Moulton
s
method using the RK4 initialization and a time step of 25 ms.
The most efficient methods are those in the shaded area on Fault 2 is close to the easternmost tie-line between Spain
the bottom left of Fig. 9, with a small error and a short CPU and France. This fault results in a cost increment of 2.51 %,
time. These methods are AM(2) with RK4 initialization and because part of the generation is re-dispatched to ensure that

VOLUME 10, 2022 70569

 M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

TABLE 4. Summary of the investigated numerical methods.

the Iberian Peninsula system does not lose synchronism with than 20 ms, and their application to the TSCOPF model can
the rest of continental Europe. As shown in Table 3, the hide numerical instabilities. On the other hand, the implicit
solution to this fault requires significant computational effort methods, except for the backward differentiation formula
in terms of the CPU time and iterations. of order 2, converge for time steps as large as 50 ms.
Fault 3 is close to the tie-line between Spain and Morocco. Consequently, the implicit methods with relatively large time
When the stability limit represented by (16) constrains the steps can mitigate the computational burden of TSCOPF
angle of the bus representing Morocco, there is no feasible problems. Furthermore, it is shown that the newly formulated
solution. As a result, Morocco’s aggregated bus is removed semi-explicit and semi-implicit predictor-corrector methods
from (10) and (16), providing a solution in which the transient are not considerably more efficient than AM(2) in TSCOPF
stability constraint is not violated, but the Moroccan system studies. Table 4 summarizes the performance of the numerical
should be disconnected from the Iberian grid, as shown in methods throughout the study and their main advantages
Fig. 10. There is no cost increase due to Fault 3 in the case of and limitations. Finally, applying the TSCOPF model to
the disconnection of Morocco. the Iberian Peninsula system helped to identify faults
Fault 4, regarded as a typical security issue in the Iberian that significantly impact the total generation cost. Three
Peninsula, does not violate the stability limits and, therefore, significant faults are found in the northeastern area of the
does not result in a re-dispatch of power production. analyzed system: one is a short-circuit near a large generator,
Therefore, this fault is not a critical contingency in the studied the other is a short-circuit near a tie-line connection with
scenario. France and the rest of continental Europe, and the third is a
Faults 5 and 6 are the only studied incidents that are not total disconnection of the Iberian Peninsula from the rest of
short-circuits. Fault 5 is the loss of the largest operating continental Europe.
power plant (2580 MW) in Portugal, leading to a 5.06 % cost More studies on other real power systems could be useful
increase owing to the resulting power imbalance. Fault 6 is to confirm these results, but there are no indications that
the disconnection of the Iberian Peninsula from the rest of the results will be different. In the future, new developments
continental Europe, which increases the cost by 2.82 %. in mathematical integration methods must be evaluated for
Finally, the algorithms have been tested for different application in TSCOPF studies.
convergence starting points in voltages, angles, and active and Conflict of Interest: The authors declare that they have no
reactive powers. The results are identical when different con- conflict of interest.
vergence starting points are applied to the same model, and
REFERENCES
the CPU time and the number of iterations are very similar.
[1] S. Xia, Z. Ding, M. Shahidehpour, K. W. Chan, S. Bu, and G. Li,
VI. CONCLUSION ‘‘Transient stability-constrained optimal power flow calculation with
extremely unstable conditions using energy sensitivity method,’’ IEEE
The main result of this study is that two of the stud- Trans. Power Syst., vol. 36, no. 1, pp. 355–365, Jan. 2021.
ied methods outperform the trapezoidal rule, which has [2] F. Arredondo, P. Ledesma, E. D. Castronuovo, and M. Aghahassani,
been, until now, the default method in TSCOPF models ‘‘Stability improvement of a transmission grid with high share of renewable
energy using TSCOPF and inertia emulation,’’ IEEE Trans. Power Syst.,
based on simultaneous discretization. These methods are vol. 37, no. 4, pp. 3230–3237, Jul. 2020.
Adams-Moulton of order two and Simpson’s rule, both with [3] S. Abhyankar, G. Geng, M. Anitescu, X. Wang, and V. Dinavahi, ‘‘Solution
RK4 in the initialization stage. The improvement they provide techniques for transient stability-constrained optimal power flow—Part I,’’
IET Gener., Transmiss. Distrib., vol. 11, no. 12, pp. 3177–3185, 2017.
in terms of computing time and accuracy is best seen in Fig. 9. [4] G. Geng, S. Abhyankar, X. Wang, and V. Dinavahi, ‘‘Solution techniques
The results also show that most methods converge and for transient stability-constrained optimal power flow—Part II,’’ IET
provide reasonably good results when a time step of 10 ms is Gener., Transmiss. Distrib., vol. 11, no. 12, pp. 3186–3193, Aug. 2017.
[5] M. Schlegel, K. Stockmann, T. Binder, and W. Marquardt, ‘‘Dynamic
used, which is typical in transient stability studies. However, optimization using adaptive control vector parameterization,’’ Comput.
the explicit methods fail to converge for time steps greater Chem. Eng., vol. 29, no. 8, pp. 1731–1751, Jul. 2005.

70570 VOLUME 10, 2022

M. Aghahassani et al.: Evaluation of Numerical Methods for TSCOPF

[6] L. T. Biegler, A. M. Cervantes, and A. Wächter, ‘‘Advances in simultaneous [30] G. Wanner and E. Hairer, Solving Ordinary Differential Equations II.
strategies for dynamic process optimization,’’ Chem. Eng. Sci., vol. 57, Berlin, Germany: Springer, 1996, vol. 375.
no. 4, pp. 575–593, Feb. 2002. [31] D. F. Griffiths and D. J. Higham, Numerical Methods for Ordinary
[7] F. Arredondo, E. Castronuovo, P. Ledesma, and Z. Leonowicz, ‘‘Analysis Differential Equations: Initial Value Problems. London, U.K.: Springer,
of numerical methods to include dynamic constraints in an optimal power 2010.
flow model,’’ Energies, vol. 12, no. 5, p. 885, Mar. 2019. [32] F. E. Cellier and E. Kofman, Continuous System Simulation. New York,
[8] G. Geng, V. Ajjarapu, and Q. Jiang, ‘‘A hybrid dynamic optimization NY, USA: Springer, 2006.
approach for stability constrained optimal power flow,’’ IEEE Trans. Power [33] Q. Zhou and J. W. Bialek, ‘‘Approximate model of European intercon-
Syst., vol. 29, no. 5, pp. 2138–2149, Sep. 2014. nected system as a benchmark system to study effects of cross-border
[9] Q. Jiang and Z. Huang, ‘‘An enhanced numerical discretization method for trades,’’ IEEE Trans. Power Syst., vol. 20, no. 2, pp. 782–788, May 2005.
transient stability constrained optimal power flow,’’ IEEE Trans. Power [34] N. Hutcheon and J. W. Bialek, ‘‘Updated and validated power flow model
Syst., vol. 25, no. 4, pp. 1790–1797, Nov. 2010. of the main continental European transmission network,’’ in Proc. IEEE
[10] Q. Jiang and G. Geng, ‘‘A reduced-space interior point method for transient Grenoble Conf., Jun. 2013, pp. 1–5.
stability constrained optimal power flow,’’ IEEE Trans. Power Syst., [35] GAMS—Cutting Edge Modeling. Accessed: Sep. 10, 2021. [Online].
vol. 25, no. 3, pp. 1232–1240, Aug. 2010. Available: https://www.gams.com/
[11] D. Gan, R. J. Thomas, and R. D. Zimmerman, ‘‘Stability-constrained [36] A. Wächter and L. T. Biegler, ‘‘On the implementation of an interior-point
optimal power flow,’’ IEEE Trans. Power Syst., vol. 15, no. 2, pp. 535–540, filter line-search algorithm for large-scale nonlinear programming,’’ Math.
May 2000. Program., vol. 106, no. 1, pp. 25–57, May 2006.
[12] A. Tiwari and V. Ajjarapu, ‘‘Optimal allocation of dynamic VAR support [37] IPOPT documentation. Accessed: Sep. 10, 2021. [Online]. Available:
using mixed integer dynamic optimization,’’ IEEE Trans. Power Syst., https://coin-or.github.io/Ipopt/
vol. 26, no. 1, pp. 305–314, Feb. 2011. [38] L. S. Vargas, V. H. Quintana, and A. Vannelli, ‘‘A tutorial description
[13] Y. Yuan, J. Kubokawa, and H. Sasaki, ‘‘A solution of optimal power flow of an interior point method and its applications to security-constrained
with multicontingency transient stability constraints,’’ IEEE Trans. Power economic dispatch,’’ IEEE Trans. Power Syst., vol. 8, no. 3, pp. 1315–1324,
Syst., vol. 18, no. 3, pp. 1094–1102, Aug. 2003. Aug. 1993.
[14] H. W. Dommel and N. Sato, ‘‘Fast transient stability solutions,’’ IEEE
Trans. Power App. Syst., vol. PAS-91, no. 4, pp. 1643–1650, Jul. 1972. MOHAMMADAMIN AGHAHASSANI received
[15] J. R. Marti and J. Lin, ‘‘Suppression of numerical oscillations in the EMTP the B.Sc. degree in electrical engineering from
power systems,’’ IEEE Trans. Power Syst., vol. 4, no. 5, pp. 739–747, the Amirkabir University of Technology, Iran,
May 1989. in 2013, and the M.Sc. degree in electrical
[16] J. Y. Astic, A. Bihain, and M. Jerosolimski, ‘‘The mixed Adams-BDF
engineering from the Politecnico de Milano, Italy,
variable step size algorithm to simulate transient and long term phenomena
in 2017. He is currently pursuing the Ph.D. degree
in power systems,’’ IEEE Trans. Power Syst., vol. 9, no. 2, pp. 929–935,
May 1994. with the Department of Electrical Engineering,
[17] J. Sanchez-Gasca, R. D’aquila, W. Price, and J. Paserba, ‘‘Variable Universidad Carlos III de Madrid, Spain. He has
time step, implicit integration for extended-term power system dynamic been an Academic Visitor at University College
simulation,’’ in Proc. Power Ind. Comput. Appl. Conf., May 1995, Dublin, Ireland. His research interests include
pp. 183–189. power system optimization and transient stability analysis.
[18] T. Noda, K. Takenaka, and T. Inoue, ‘‘Numerical integration by the 2-stage
diagonally implicit Runge–Kutta method for electromagnetic transient EDGARDO D. CASTRONUOVO (Senior
simulations,’’ IEEE Trans. Power Del., vol. 24, no. 1, pp. 390–399, Jan. Member, IEEE) received the B.S. degree in elec-
2009. trical engineering from the National University
[19] D. Fabozzi and T. Van Cutsem, ‘‘Simplified time-domain simulation of of La Plata, Argentina, in 1995, the M.Sc. and
detailed long-term dynamic models,’’ in Proc. IEEE Power Energy Soc. Ph.D. degrees from the Federal University of Santa
Gen. Meeting, Jul. 2009, pp. 1–8. Catarina, Brazil, in 1997 and 2001, respectively,
[20] G. Tzounas, I. Dassios, and F. Milano, ‘‘Small-signal stability analysis of and the Ph.D. degree from INESC TEC, Portugal,
numerical integration methods,’’ IEEE Trans. Power Syst., early access, in 2005. He worked with the CEPEL, Brazil,
Jan. 27, 2022, doi: 10.1109/TPWRS.2022.3146345.
and INESC TEC. He is currently a Professor
[21] A. Tutueva, T. Karimov, and D. Butusov, ‘‘Semi-implicit and semi-explicit
with the Department of Electrical Engineering,
Adams-Bashforth-Moulton methods,’’ Mathematics, vol. 8, no. 5, p. 780,
May 2020. University Carlos III of Madrid, Spain, on leave from CentraleSupèlec,
[22] A. Tutueva and D. Butusov, ‘‘Stability analysis and optimization of France. His research interests include are in optimization methods applied
semi-explicit predictor–corrector methods,’’ Mathematics, vol. 9, no. 19, to power system problems, renewable generation, and storage. He is the
p. 2463, Oct. 2021. Vice-President of the IEEE Spain Section.
[23] F. Milano, ‘‘An open source power system analysis toolbox,’’ IEEE Trans.
Power Syst., vol. 20, no. 3, pp. 1199–1206, Aug. 2005. PABLO LEDESMA received the Ph.D. degree from the University Carlos
[24] R. Zarate-Minano, T. Van Cutsem, F. Milano, and A. J. Conejo, ‘‘Securing III of Madrid, in 2001. He is currently an Associate Professor with the
transient stability using time-domain simulations within an optimal power University Carlos III of Madrid. He has worked with the Spanish TSO
flow,’’ IEEE Trans. Power Syst., vol. 25, no. 1, pp. 243–253, Feb. 2010. on several projects in large-scale integration of renewable energy. He is
[25] M. Aghahassani, E. D. Castronuovo, P. Ledesma, and F. Arredondo, an Academic Visitor with Chalmers University, Sweden, and Strathclyde
‘‘Application of linear multi-step methods to a transient stability con- University, U.K. His research interests include are transient stability and
strained optimal power flow model,’’ in Proc. IEEE Int. Conf. Environ. dynamic modeling of power systems.
Electr. Eng. IEEE Ind. Commercial Power Syst. Eur. (EEEIC/I& CPS
Europe), Jun. 2020, pp. 1–6. FRANCISCO ARREDONDO received the Ph.D.
[26] Powerworld. Accessed: Sep. 10, 2021. [Online]. Available:
degree in electrical engineering from the Univer-
https://www.powerworld.com/
sity Carlos III of Madrid, in 2019. He is currently
[27] M. Y. Borodulin, ‘‘Effect of numerical integration on critical time
evaluation in power system stability studies,’’ in Proc. IEEE Power Energy
an Assistant Professor with the University Carlos
Soc. Gen. Meeting, Jul. 2013, pp. 1–5. III of Madrid. He has been an Academic Visitor
[28] K. Prabhashankar and W. Janischewsyj, ‘‘Digital simulation of multima- with the Wrocław University of Technology,
chine power systems for stability studies,’’ IEEE Trans. Power App. Syst., Poland. His research interests include the power
vol. PAS-87, no. 1, pp. 73–80, Jan. 1968. system stability, renewable generation, energy
[29] K. Stanton and S. Talukdar, ‘‘New integration algorithms for transient storage systems, and mathematical optimization
stability studies,’’ IEEE Trans. Power App. Syst., vol. PAS-89, no. 5, methods.
pp. 985–991, May 1970.

VOLUME 10, 2022 70571