Electric Power Systems Research 174 (2019) 105859

Contents lists available at ScienceDirect

Electric Power Systems Research

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

An efficient numerical method for computing eigenvalue sensitivity with T

respect to operational parameters of large-scale power systems☆
Yongjie Li , Wenlei Li, Xuhua Shi, Xiaxiao Chen
⁎

Ningbo University, No. 818, Fenghua Rd., Jiangbei District, Ningbo City, Zhejiang 315211, China

ARTICLE INFO ABSTRACT

Keywords: Eigenvalue sensitivity with operational parameters provides necessary information for tuning power system
Eigenvalue sensitivity operating status and improving small signal stability. Presently, with the inter-connection of regional power
Large-scale power systems grids, power systems expand to be of large scale, which raises demanding requirements for computational ef-
Operational parameter ficiency of eigenvalue sensitivity analysis. For addressing this issue, a modified numerical method is proposed for
Parallel computing
computing eigenvalue sensitivity with respect to power system operational parameters, including active and
Numerical method
reactive power outputs of generation units, load demands and bus voltages. Well computational efficiency is
obtained by the proposed method, which is mainly attributed to the reuse of LU factorization in solving per-
turbed power flow equations, the numerical finite difference for evaluating state matrix sensitivity, and utili-
zation of parallel computing techniques. Numerical experiments on different scale power systems from realistic
world are performed to validate its effectiveness and efficiency. Parallelization of the proposed method is im-
plemented for further reducing overall time cost. Moreover, the proposed method is integratable to functional
modules of existing power system analysis software, which makes its implementation convenient.

1. Introduction determine operating point of power systems along with the stability
property. The relation between operational parameters and eigenvalues
Modern power systems are continuously expanding due to the inter- is of high non-linearity and discontinuity, and brings difficulties for
connections of regional grids. High dimension of power system dynamic evaluating the operational eigenvalue sensitivity.
model brings many challenges for stability analysis. For securely op- Analytical methods were proposed for computing operational ei-
erating power systems, small signal stability should be guaranteed, genvalue sensitivity in [9–12]. Two issues of applying analytical
which is able to be enhanced by re-dispatching power generations methods should be addressed. One is that, the analytical expression is
based on eigenvalue sensitivity. Additionally, in recent decades, mas- tedious to derive and lack of flexibility for integrating various compo-
sive integration of renewable energies makes power system more time- nents. The other is that, for large-scale systems, the high problem di-
varying than ever, which raises rigorous demand on computational mension might make computational time cost intolerant. Alternatively,
efficiency of analysis methods for eigenvalue and its sensitivity. in [5], a numerical perturbation method is used to evaluate operational
There are two types of eigenvalue sensitivity. One is with respect to eigenvalue sensitivity for re-dispatching power generation, which is
parameters of controller loops, e.g. coefficients of power system stabi- also very time consuming for large-scale systems. In [7], eigenvalue
lizer [1–3], FACTs device [4], etc. This type of eigenvalue sensitivity is sensitivity is evaluated by central difference based on sequentially
very useful for tuning controllers to obtain better stabilization perfor- solving power flow, linearizing dynamical model and computing tar-
mance. The other type of eigenvalue sensitivity is with respect to geted eigenvalues. Another important issue mentioned in [7] is that,
parameters that determine system operating status [5–8], i.e. opera- magnitude of perturbation should be properly set. Too small pertur-
tional parameters. The operational parameters generally include power bation may cause convergence problem in solving perturbed power
output of generation unit, load demand and bus voltage, etc. In the flow equations, while too big one may produce unprecise results. In
aspect of computation, operational eigenvalue sensitivity is more other words, computational accuracy should be carefully considered for
complex and difficult to be evaluated, as operational parameters numerical methods. To some extent, the drawbacks of the above

☆
The work is supported by the Natural Science Foundation of Zhejiang Province (LY18F03004) and the Natural Science Foundation of China (51677164,
61803214).
⁎
Corresponding author.
E-mail address: [email protected] (Y. Li).

https://doi.org/10.1016/j.epsr.2019.04.037
Received 10 November 2018; Received in revised form 12 March 2019; Accepted 27 April 2019
Available online 10 May 2019
0378-7796/ © 2019 Elsevier B.V. All rights reserved.
Y. Li, et al. Electric Power Systems Research 174 (2019) 105859

Nomenclature QG vector of reactive power outputs of all generation units

Ixk x-axis component of injected current at bus k
Nomenclature used in this paper is given as follows. Iyk y-axis component of injected current at bus k
Gij real part of the ith row jth column element of admittance
ΩNk set of buses connected with bus k including itself matrix
ΩGk set of generation units at bus k Bij imaginary part of the ith row jth column element of ad-
ΩLk set of load demands at bus k mittance matrix
Vxk x-axis component of voltage at bus k x vector of differential variables
Vyk y-axis component of voltage at bus k y vector of electric network voltages in rectangular co-
Vx vector of all x-axis components of bus voltages ordinates
Vy vector of all y-axis components of bus voltages S vector of considered operational parameters
PGi active power output of the ith generation unit Sk the kth operation parameter in S
QGi reactive power output of the i-th generation unit  generalized state matrix
PG vector of active power outputs of all generation units

methods cause that functional module of eigenvalue sensitivity analysis dimension of A.

has not been integrated into present power system analysis software,
ˆ v = Ev ,
A ˆ Tu =
A T Eu (3)
including PSS/E [13], PSAT [14], etc.
For enhancing both computational efficiency and accuracy of op-
ˆ = A B ,
A E= I 0
erational eigenvalue sensitivity analysis of large-scale power systems, C D 0 0 (4)
we propose a modified numerical method in this paper. The remaining
contents are organized as follows. Section 2 introduces the mathema-
tical model of eigenvalue sensitivity problem. The general procedures 2.2. Eigenvalue sensitivity
of analytical method and conventional numerical method are also
provided. Section 3 provides the framework of the proposed method, For the eigenvalue sensitivity analysis problem in this paper, con-
and techniques for improving computational efficiency. In Section 4, sidered operational parameters include active power PG and reactive
effectiveness and computational efficiency of the proposed method are power QG of generation units. [PG, QG]T is collected as operational
validated by comparative experiments on four test systems from rea- parameter vector S, which denotes an operating point of power system.
listic world, which have 39, 2383, 9241 and 13,659 buses, respectively. At a given operating point S, network voltages are able obtained by
Parallel performance of the proposed method is also tested. Section 5 solving power flow Eqs. (5)–(8), where Ixk and Iyk are generation cur-
gives the conclusions. rents injected into electric network. For brevity, power flow Eqs. (5)
and (6) are simplified as (9) in the following contents.
2. Formulation for eigenvalue sensitivity PGi PLi = Vxk Ixk + Vyk Iyk , k N
i Gk i Lk (5)
2.1. Small signal stability model of power system
QGi QLi = Vxk Iyk + Vyk Ixk , k N
Electro-mechanical dynamics of power system is considered for i Gk i Lk (6)
small signal stability analysis. The system dynamics are able to be de-
Ixk = Vxj Gkj Vyj Bkj , k N
scribed by a set of differential and algebraic equations (DAE) as (1), j (7)
Nk
where x is the vector of differential variables including generator rotor
angle, rotor speed, signals from controllers, etc., and y is the vector of Iyk = Vxj Bkj + Vyj Gkj, k N
electric network voltages in rectangular coordinates, which denotes j Nk (8)
[Vx, Vy]T. The dynamical models of generator, excitation system and
turbine governor are provided in Appendices A–C, respectively, where
S = h (y ) (9)
the differential equations and several set of parameters are given. More For an equilibrium point of system (1), x0 is determined by y0 via f
details on the electro-mechanical model of power systems used in this (x0, y0) = 0. Matrix pencil (Aˆ , E) is then uniquely determined by (x0,
paper refer to [14,15]. y0). Hence, eigenvalue of (A ˆ , E) is actually a function of operational
parameter S. Sensitivity of eigenvalue λ with respect to any operational
x = f (x , y )
parameter Sk ∈ S is able to be evaluated by (10), where sensitivity of
0 = g (x , y ) (1) generalized state matrix  is required [9–11].
If (x0, y0) satisfies that f(x0, y0) = 0 and g(x0, y0) = 0, (x0, y0) is an ˆ
DA
d uT dS v
equilibrium point for DAE system (1). Small signal stability of system = T k
(1) is investigated within the neighborhood of an equilibrium point. dSk u Ev (10)
The stability property is the same with its linearized model, which is
given by (2) where A, B, C and D are the Jacobian matrix of f and g at 2.3. Sensitivity of generalized state matrix
(x0, y0).
x =A x+B y For conventional analytical methods [9,10], sensitivity of  with
respect to S is computed based on differential chain rule, which is
0 =C x +D y (2)
briefly summarized as follows. Firstly, sensitivities of  with respect to
For the stability analysis of linear system (2), a generalized eigen- network voltage y are evaluated by (11). The required partial differ-
value problem is formulated by (3), where λ, v and u are the eigen- ential of injected current (Ixi, Iyi) on y are the linear coefficients of (7)
value, corresponding generalized right eigenvector and left eigenvector and (8). Then, sensitivity of y with respect to S is approximated by
of matrix pencil (Aˆ , E) , respectively. Generalized state matrix  and E linearizing (9), which is given as (12). Based on (11) and (12), sensi-
are provided by (4), where I is an unit diagonal matrix with the same tivity of  with respect to S is able to be evaluated by (13), where

2
Y. Li, et al. Electric Power Systems Research 174 (2019) 105859

superscript −1 is the matrix inverse operation, subscript (i, k) stands for

the ith row kth column element of a matrix and nB is the number of
electric network buses.

dAˆ Aˆ Aˆ Ixi Aˆ Iyi

= + +
dVxk Vxk i
Ixi Vxk i
Iyi Vxk
Nk Nk

dAˆ Aˆ Aˆ Ixi Aˆ Iyi

= + +
dVyk Vyk Ixi Vyk Iyi Vyk Fig. 1. Reutilization of LU factorization of initial Jacobian matrix.
i Nk i Nk (11)

y=[ y h]
1 S (12)
performed without re-formulating and re-factorizing Jacobian matrix
dAˆ dAˆ dVxi dAˆ dVyi dAˆ over again. Fig. 1 demonstrates how to reuse LU factorization for dif-
1
= + [ y h](i, k ) ferent operational parameters. Only one iteration is required to make
dSk dVxi dSk dVyi dSk dVxi
i N i N
(14) converge in most cases with slightly perturbed ΔSk, e.g. 10−6 in
dAˆ 1 the numerical experiments.
+ [ y h](nB + i, k )
dVyi (13)
−1 3.2. Sparse matrix computations
As [∇ yh] is dense matrix, evaluation of (13) has 2nB sparse
matrix additive operations. Assume that sparse matrix dA ˆ /dVxi or
dAˆ /dVyi has nz non-zero elements in average, the computational com- Eigenvalues and the corresponding eigenvectors of linear system (2)
are required at the first stage of evaluating eigenvalue sensitivity For
plexity of (13) is about O(2nBnz). Additionally, evaluation of dAˆ /dVxi or
dAˆ /dVyi has computational complexity about O(nGnz), where nG is the large-scale systems. As only a part of eigenvalues of weak or negative
number of generation units. The overall complexity of calculating state damping ratio are of interest, i.e. critical eigenvalues, partial eigenvalue
matrix sensitivity by the reported analytical method is O(2(nG + nB)nz). method, e.g. Implicitly Restart Arnoldi [16], Krylov-Schur [17], Two-
For perturbation based numerical method, it is not necessary to sided Jacobian Davidson [18], are suitable for computing these eigen-
values. For generalized state matrix  , its submatrix A is of block-di-
derive and evaluate expression (11) and (13). By perturbing operational
agonal structure shown by (16). Submatrix B and C are also block-wise
parameter Sk with a small ΔSk, power flow equation (14) is solved for
spare matrix. D shares the same structure with admittance matrix of
getting the new equilibrium point, where S0 is the initial operating
electric network. Hence  and E are very sparse, sparsity oriented
point and ek is the kth unit vector. Generalized state matrix  is re-
computational routines for matrix operations in applying partial ei-
evaluated at perturbed operating point determined by S0 + ΔSk, where
ΔSk = ekΔSk. The sensitivity of  is able to be evaluated by finite dif- genvalue methods are able to be utilized in solving eigenvalue problem
ference (15). Only one subtraction operation of two sparse matrix is (3).
required. The computational complexity of (15) is about O A = diag(A1, A2, …, AnG) (16)
(nGnz + 2nB), which is lower than O(2(nG + nB)nz) of analytical
method. It is theoretically inferred that calculating sensitivity of  by Filling the elements of  S0+ Sk and its subtraction with  S0 for
perturbation based numerical approach is efficient than the reported evaluating state matrix sensitivity by (15) is able to be performed in
analytical methods. sparse format, which reduces both memory usage and computational
time. For evaluating eigenvalue sensitivity by (10), sparsity of the
S0 + ek Sk = h (y0 + y ) (14) matrix  and E is also able to be utilized.
ˆ
dA ˆS +
A ˆS
A
0 Sk 0
3.3. Application of parallel computing
dSk Sk (15)
In the case of massive operation parameters for large-scale power Hierarchical parallelization is able to be implemented for the pro-
system, the primary issue of evaluating (15) is computational efficiency posed numerical method. Computations including solving (14), evalu-
of solving the perturbed power flow equation (14). For utilizing Newton ating (15) and (10) for different operational parameters are in-
method to solve (14), Jacobian matrix ∇yh of (9) is evaluated and LU- dependent, except that the LU factorization of J0 are shared.
factorized to solve correction equations in every iteration. For large- Computations are also able to be decoupled by different critical ei-
scale power systems, the number of operational parameters and the genvalues. The applicable parallel framework is provided in Fig. 2. Well
dimension of power flow equation are huge, which might bring un- load balance with using multiple processing units is expected to be
acceptable computational burden. For improving computational effi- achieved under this framework, as computation burden for each pair of
ciency of eigenvalue sensitivity analysis of large-scale power systems, a operational parameter and eigenvalue stays almost the same.
modified numerical method is proposed, featured by acceleration in
solving the perturbed power flow equations, utilization of sparsity of 3.4. Complete procedures
the state matrix and parallel computing techniques.
The complete procedures of the proposed numerical method are
3. Improved numerical method for eigenvalue sensitivity analysis provided by Fig. 3. Note that, in step (1), λ denotes the critical eigen-
value, which is obtained with partial eigenvalue methods. Step (2)–Step
3.1. Reuse of LU factorization in solving multiple perturbed power flow (5) are able to be performed under the parallel framework of Fig. 2.

For solving (14) with Newton method, the most computationally 4. Numerical experiments
expensive step is solving linear correction equations for every iteration,
which includes LU factorization of Jacobian matrix and following for- 4.1. Description of test systems
ward/backward substitutions. For small perturbation ΔSk, the solution
y0 + Δy is close to the initial y0. It is viable that the Jacobian matrix J0 Four test systems from open-source software MATPOWER [19] are
at operating point S0 and its LU factors are able to be re-used to solve used to validate the accuracy and efficiency of the proposed method.
perturbed power flow equation. Hence, for perturbation on different The test systems come from real-world power systems, which are New
operational parameters, only forward/backward substitutions are England 39-bus system, Polish 2383-bus system, European 9241-bus

3
Y. Li, et al. Electric Power Systems Research 174 (2019) 105859

Fig. 2. Parallel framework for the proposed numerical method.

Table 2
Critical eigenvalue information.
System No. Dimension of  α β Damping ratio ξ

1 248 0.0041 7.0280 −0.0587%

2 10,325 −0.2228 7.9675 2.7956%
3 37,267 −0.0235 7.3245 0.3202%
4 80,514 −0.0056 7.9034 0.0704%

α and β are the real and imaginary part of critical eigenvalue, respectively.

and CNM, respectively. The results are plotted in Fig. 4, where reactive
Fig. 3. Complete procedures of the proposed numerical method.
power of generation unit 4 serves as the slack variable in solving (14)
and its sensitivity is set to be zero mandatorily. It is observed that
system and European 13,659-bus system, respectively. Electro-me- sensitivities obtained by the three methods are in well consistence.
chanical dynamics are considered for generation units, which includes For test system 2–4, the evaluated damping ratio sensitivity are
generator, excitation and turbine governor. The differential models of provided by Figs. 5–7. AM is not performed for system 4, due to un-
the above components and model parameters are provided in Appen- acceptable computational cost on memory and time. The results of the
dices. Load demand is modeled as constant impedance. The profiles of three methods are also in well consistence, which demonstrates that the
test systems are provided in Table 1. Critical eigenvalues with the proposed method is effective with well accuracy.
weakest damping ratio is considered. Its real part, imaginary part and For quantitatively assessing computational accuracy of the proposed
corresponding damping ratio are provided in Table 2. MNM, the maximal relative errors (19) of eigenvalue sensitivity are
In numerical experiments, investigated operational parameters are calculated based on the results of CNM and summarized in Table 3. It is
active and reactive power outputs of generation units. Damping ratio observed that satisfying accuracy of MNM is obtained.
sensitivity is used to show the computational results, which is obtained
d d d
by (17) based on eigenvalue sensitivity. Generation units with the lar- e = max | MNM CNM
|/| CNM
|
Sk S dSk dSk dSk (19)
gest participation factor in the critical oscillation mode are selected in
eigenvalue analysis, and the quantity of selected generation units of the
four test systems are 10, 50, 100 and 200, respectively. 4.3. Computational efficiency
d 2 d
d
= + For comparing computational efficiency, the three methods are used
dSk 3 2 + 2 dSk 3 2 + 2 dSk (17) to evaluate sensitivity of critical eigenvalue of the 4 test systems. The
For testing computational accuracy and efficiency of the proposed number of considered operational parameters is 20, 100, 200, and 400,
modified numerical method (MNM for brevity), analytical method respectively. Experiments is carried on a personal computer with
(AM) and conventional numerical method (CNM) are used to perform 3.2 GHz Intel Core CPU and 8 GB RAM. Note that, the three methods
the eigenvalue sensitivity analysis for comparison. In CNM, eigenvalue are performed in serial execution with only one computing thread.
sensitivity is evaluated by (18), where S0 + Sk is computed by implicitly Overall time costs of the three methods are given in Table 4. It is ob-
restarted Arnoldi method at perturbed operating point. served that, MNM obtains the best computational efficiency for all test
systems.
d S0+ Sk S0 For locating the computational bottlenecks of the three method,
dSk Sk (18) time costs on sub-routines are measured in application to test system 4,
and the results are provided by Table 5. It is observed that the most
computational expensive sub-routine is computing critical eigenvalue
4.2. Computational effectiveness at perturbed operating points for CNM. While for AM, the most ex-
pensive part is evaluating sensitivity of  . For the proposed MNM, it is
Damping ratio sensitivity of test system 1 is evaluated by MNM, AM formulation of  that cost the most time.

Table 1
Profile of test systems.
System No. Number of buses Number of generation units Number of transmission lines Total power demands Total power generations

1 39 10 46 6254 MW 6298 MW
2 2383 327 2896 24,558 MW 25,108 MW
3 9241 1445 16,049 312,350 MW 320,350 MW
4 13,659 4092 20,467 381,430 MW 390,540 MW

4
Y. Li, et al. Electric Power Systems Research 174 (2019) 105859

Fig. 4. Damping ratio sensitivity obtained by three methods for system 1.

Fig. 5. Damping ratio sensitivity obtained by three methods for system 2.

For assuring fairness of the above comparison, calculation in step 6 partial eigenvalue algorithms. Critical eigenvalue and its eigenvector at
of CNM is optimized. The calculation of (18) is implemented based on initial operating point are passed to eigs function as the shift point for
eigs function in Matlab software. The core algorithm is Implicitly Re- matrix transformation and starting vector. Only a few iterations are
started Arnoldi [16], which is one of the most efficient non-Hermitian required to make perturbed eigenvalues converge. However, the

Fig. 6. Damping ratio sensitivity obtained by three methods for system 3.

Fig. 7. Damping ratio sensitivity obtained by AM and MNM for system 4.

5
Y. Li, et al. Electric Power Systems Research 174 (2019) 105859

Table 3
Maximal relative error of eigenvalue sensitivity for CNM and MNM.
Test system Case 39 Case 2383 Case 9241 Case 13,659

e 2.02e−06 1.41e−05 7.39e−07 6.05e−08

Table 4
Overall computational time of three methods.
Test system CNM/s AM/s MNM/s

39-bus 0.39 0.15 0.12

2383-bus 67.69 91.94 5.51
9241-bus 933.01 2083.32 91.80 Fig. 8. Parallel speedup under the parallelization scheme.
13,659-bus 2614.01 / 330.51
t1
r=
tm (20)

Table 5
The speedup curves are provided by Fig. 8. It is observed that
Computational time (seconds) on Sub-routines for Test System 4. system with larger scale obtains better speedup. A theoretical analysis
of the speedup is given as follows. For MNM, the overall time cost
Sub-routines CNM AM MNM
contains three parts, which are serial execution, parallel execution and
1. Eigen-analysis at S0 0.674 0.441 0.585 management of parallel threads. Serial execution includes step 1 and
2. LU factorization of J0 0.049 0.032 0.043 step 2 in Table 6, and parallel execution includes step 3–step 6. For
3. Solving perturbed power flow 5.100 / 4.423 different test systems, the main factors influencing computation com-
0.674 / 83.040
4. Initializing and formulating  plexity are test system scale and the size of operational parameter set S.
5. Computing dA ˆ /dSk / 2082.375 2.857
For the serial part, computational complexity of initial eigen-analysis
6. Computing dλ/dSk 831.427 0.912 0.853
and LU factorization is assumed to be O(n Âa ) , where n  is the dimension
Total time 933.010 2083.320 91.800
of  and a is obviously larger than 1. For the parallel part, computa-
tional complexity is assumed to be O(nS n Âb ) , where nS is the size of
operational parameter set and b is larger than 1. As commonly ac-
Table 6 cepted, complexity of eigenvalue computation is higher than numerical
Parallel computing time (seconds) of the Proposed MNM. evaluation of state matrix sensitivity. Hence, it satisfies that a > b. For
managing parallel threads, time cost is nearly fixed for different scale
Test system Number of used computing threads
systems with the same number of threads. Its complexity is approxi-
1 2 4 8 16 32 mately O(m), where m is the number of threads.
39-bus 0.553 0.483 0.542 0.508 0.508 0.491 For larger scale systems, there are generally more operational
2383-bus 6.272 3.225 2.022 1.368 0.989 0.937 parameters to be considered, and the parallel part is more dominant in
9241-bus 85.304 44.013 23.713 12.647 7.229 4.474
13,659-bus 307.294 155.982 82.693 42.817 23.303 13.829
overall computational complexity. Assume that the overall time cost
with one thread is approximated by C1 n Aˆ a + C2 nS n Aˆ b + C3 , where C1,
C2, and C3 are coefficients that match the proportions of the three parts.
computational complexity of (18) with the optimization is still higher The parallel speedup is able to be expressed as (21). With fixed m
than evaluating (10) with the readily obtained eigenvalue and eigen- parallel threads, larger system scale and larger number of parameters,
vector at initial operating point. For MNM, the most computational i.e. bigger n and bigger k, would make parallel speedup r better and
expensive step 4 is able to be further optimized by fine-grained paral- closer to its ideal value. The number of operational parameters for the
lelization. At perturbed operating points, initializing variables of system four test systems are 20, 100, 200, 400 respectively. The results of
dynamic models and formulating sub-blocks of state matrix related with parallel test are consistent with the theoretical analysis.
each component is fully parallelizable. Component-level fine-grained t1 C1 n Aˆ a + C2 nS n Aˆ b + C3 m
parallelization is able to be integrated into the presented coarse-grained r= =
tm C1 n Aˆ a + C2 nS n Aˆ b/ m + C3 m
parallel framework in Fig. (2). This fine-grained parallelization could be
m 1
also applied to AM and CNM to help deal with complex dynamic =1+ C 1 C 1
m2 C3 + m C1 b a + 1
component in realist power systems, but it would bring rigorous de- b
2 nS n A
ˆ 2 nS n A
ˆ (21)
mands on computing platform.

5. Conclusions
4.4. Parallel performance

An efficient numerical method with well parallelization capability is

Eigenvalue sensitivity computations are decoupled by different
proposed for computing operational eigenvalue sensitivity of large-
groups of operational parameters. Each group is assigned as equal as
scale power systems. Sufficient experiment results demonstrate its well
possible number of parameters to assure good load balancing among
performance in the aspect of computational accuracy, efficiency and
computing threads. Parallelization test is performed on an SMP
scalability. The numerical approach to evaluating state matrix sensi-
(Symmetric Multi-Processing) workstation with Linux operating
tivity makes the proposed method integratable with functional modules
system. Up to 32 CPU cores, Intel E7-8837 with 2.67 GHz frequency,
of existing power system analysis software, e.g. eigenvalue analysis,
are available. 1, 2, 4, 8, 16, and 32 computing threads are used to
power flow solving, state matrix formulation. The main contributions of
perform the test, respectively. Computational time costs under different
the reported work are summarized as follows.
number of computing threads are summarized in Table 6.
Parallel speedup is defined as (20), where t1 is the time with one
(1) Modifications are made on conventional numerical methods, which
computing thread and tm is with m computing threads.

6
Y. Li, et al. Electric Power Systems Research 174 (2019) 105859

features reuse of LU factorization in solving perturbed power flow (2) Comprehensive comparisons of the proposed method with analy-
equations, adoption of sparse matrix calculations and utilization of tical method and conventional numerical method are performed.
parallel computing techniques. Remarkable enhancement on com- The computational accuracy and effectiveness are validated, which
putational efficiency is achieved. lays foundation for its practical application.

Appendix A. Generator model

Differential equations of generator dynamical model and its parameters are given by (A.1) and Table A.1 , respectively.

A.1
Parameters of generator model.
Set No. 1 2 3 4

xd 1.25 1.7 1.8 1.81

xd 0.232 0.245 0.3 0.3
xd 0.12 0.185 0.25 0.23
td0 4.75 5.9 8 8
td0 0.06 0.033 0.03 0.03
xq 1.22 1.64 1.7 1.76
xq 0.715 0.38 0.55 0.65
xq 0.12 0.185 0.25 0.25
tq0 1.5 0.54 0.4 1
tq0 0.21 0.076 0.05 0.07
TJ 10.12 7.92 13 7

d
= *( 0)
dt
d 1
= (Pmech Pe )
dt TJ
deq 1 Xd Xd Xd Xd
= ( eq + eq + vf )
dd Td0 Xd Xd Xd Xd
deq 1
= {eq eq (Xd Xd ) i d}
dt Td0
ded 1 Xq Xq Xq Xq
= ( ed + ed )
dt Tq0 Xq Xq Xq Xq
ded 1
= {(Xq Xq ) iq + ed ed }
dt Tq0 (A.1)

Appendix B. Excitation system model

The block diagram of excitation system and coefficients of control loops are given by Fig. B.1 and Table B.1 , respectively.

Fig. B.1. Block diagram of excitation system.

7
Y. Li, et al. Electric Power Systems Research 174 (2019) 105859

B.1
Parameters of excitation system model.
Set No. 1 2 3 4

vrmax 5 4.38 7.32 5

vrmin −5 0 0 −5
Ka 20 40 200 20
Ta 0.2 0.02 0.02 0.055
Kf 0.063 0.001 0.002 0.125
Tf 0.35 1 1 1.8
Te 0.314 1.98 0.2 0.36
Tr 1.00E−03 1.00E−03 1.00E−03 5.00E−02
Ae 3.90E−03 6.00E−04 6.00E−04 5.60E−03
Be 1.555 0.9 0.9 1.075

Appendix C. Turbine governor model

The block diagram of turbine governor and coefficients of control loops are given by Fig. C.1 and Table C.1 , respectively.

Fig. C.1. Block diagram of turbine governor.

C.1
Parameters of turbine governor model.
Set No. R Tmax Tmin Ts Tc T3 T4 T5

1 0.0005 1.2 0.3 0.1 0.45 0.2 12 50.0

2 0.0200 1.2 0.3 0.1 0.45 0.2 12 50.0
3 0.0400 100.0 0.0 20.0 0.20 0.2 0 0.2

References [10] K. Wang, C. Chung, C. Tse, K. Tsang, Multimachine eigenvalue sensitivities of power
system parameters, IEEE Trans. Power Syst. 15 (2) (2000) 741–747.
[11] H.-K. Nam, Y.-K. Kim, K.-S. Shim, K.Y. Lee, A new eigen-sensitivity theory of aug-
[1] C. Tse, K. Wang, C. Chung, K. Tsang, Robust PSS design by probabilistic eigenvalue mented matrix and its applications to power system stability analysis, IEEE Trans.
sensitivity analysis, Electr. Power Syst. Res. 59 (1) (2001) 47–54 2001. Power Syst. 15 (1) (2000) 363–369.
[2] D. Sumina, N. Bulić, M. Mišković, Parameter tuning of power system stabilizer using [12] T. Smed, Feasible eigenvalue sensitivity for large power systems, IEEE Trans. Power
eigenvalue sensitivity, Electr. Power Syst. Res. 81 (12) (2011) 2171–2177. Syst. 8 (2) (1993) 555–563.
[3] C. Tse, S. Tso, Design optimisation of power system stabilisers based on model and [13] S. Li, X. Liang, W. Xu, Dynamic load modeling for industrial facilities using template
eigenvalue-sensitivity analyses, IEE Proceedings C-Generation, Transmission and and PSS/E composite load model structure CLOD, Industrial and Commercial Power
Distribution, vol. 135, IET, 1988, 1988, pp. 406–415. Systems Technical Conference (I&CPS), 2017 IEEE/IAS 53rd, IEEE, 2017, 2017,
[4] C. Li, J. Deng, X.-P. Zhang, Coordinated design and application of robust damping pp. 1–9.
controllers for shunt FACTS devices to enhance small-signal stability of large-scale [14] F. Milano, An open source power system analysis toolbox, IEEE Trans. Power Syst.
power systems, CSEE J. Power Energy Syst. 3 (4) (2017) 399–407. 20 (3) (2005) 1199–1206.
[5] R. Zárate-Miñano, F. Milano, A.J. Conejo, An OPF methodology to ensure small- [15] X.-F. Wang, Y. Song, M. Irving, Modern Power Systems Analysis, Springer Science &
signal stability, IEEE Trans. Power Syst. 26 (3) (2011) 1050–1061. Business Media, 2010.
[6] J. Cao, W. Du, H. Wang, Z. Chen, H. Li, A novel emergency damping control to [16] D.J. Kim, Y.H. Moon, Application of the implicit restarted Arnoldi method to the
suppress power system inter-area oscillations, IEEE Trans. Power Syst. 28 (3) small-signal stability of power systems, J. Electr. Eng. Technol. 2 (4) (2007)
(2013) 3165–3173. 428–433.
[7] T. Parreiras, S. Gomes, G. Taranto, K. Uhlen, Closest security boundary for im- [17] Y. Li, G. Geng, Q. Jiang, An efficient parallel Krylov-Schur method for eigen-ana-
proving oscillation damping through generation redispatch using eigenvalue sen- lysis of large-scale power systems, IEEE Trans. Power Syst. 31 (2) (2016) 920–930.
sitivities, Electr. Power Syst. Res. 160 (2018) 119–127. [18] Z. Du, C. Li, Y. Cui, Computing critical eigenvalues of power systems using inexact
[8] E.A. Zamora-Cárdenas, C.R. Fuerte-Esquivel, Computation of multi-parameter sen- two-sided Jacobi-Davidson, IEEE Trans. Power Syst. 26 (4) (2011) 2015–2022.
sitivities of equilibrium points in electric power systems, Electr. Power Syst. Res. 96 [19] R.D. Zimmerman, C.E. Murillo-Sánchez, R.J. Thomas, et al., MATPOWER: steady-
(2013) 246–254. state operations, planning, and analysis tools for power systems research and
[9] J. Ma, Z.Y. Dong, P. Zhang, Eigenvalue sensitivity analysis for dynamic power education, IEEE Trans. Power Syst. 26 (1) (2011) 12–19.
system, PowerCon 2006. International Conference on Power System Technology,
IEEE, 2006, 2006, pp. 1–7.