J. Inst. Eng. India Ser.

B
https://doi.org/10.1007/s40031-021-00707-1

ORIGINAL CONTRIBUTION

Robust Incrementalized Least-Squares Estimation Method

in State Estimation with Phasor Measurement Unit Measurements
S. Bindu1 • S. Ushakumari2 • J. S. Savier3

Received: 1 October 2020 / Accepted: 15 December 2021

Ó The Institution of Engineers (India) 2022

Abstract Power system operators need real-time informa- Keywords Power system  State estimation 
tion to reliably maintain delivery of energy to the consumers Sparse least-squares minimization 
in a modern electric power network characterized by Phasor measurement unit 
dynamically varying power flow and infrastructure avail- Supervisory control and data acquisition (SCADA)
ability. State estimation involves minimizing a sparse non-
linear sum-of-squares objective function. The conventional
method of solution is Gauss–Newton method which is sub- Introduction
jected to divergence issues, particularly in the presence of
bad data and topology error. In this paper, numerical opti- SE is used to estimate the voltage magnitudes and
mization techniques like Powell’s dog leg (PDL), indefinite respective phase angles for all nodes of the electric net-
Gauss–Newton Powell’s dog leg (IGNPDL) and robust work. The estimated values shall have minimum error, and
incrementalized least-squares estimation are presented. prior information regarding the states of the power system
IGNPDL is an incrementalized version of the PDL method is required to estimate the present state.
which is generally used suitably in sparse least-squares SE is presented as one of the key features in energy
minimization. These methods are robust to nonlinearity of management systems, which helps in building a network
objective function. The robustness is extendable to numeri- real-time model [1]. Rapid convergence and accuracy are
cal ill-conditioning also which arises due to topology errors expected from a modern robust state estimator.
and bad data. Data from phasor measurement units are also GN method is the standard technique followed for the
used in the SE to enhance the performance. The proposed solution of SE problems. The conventional SE based on
algorithm is simulated and evaluated on IEEE 14-bus and GN method is found to diverge under many situations. The
IEEE 118-bus systems. The proposed methods converge in ill-conditioning of the gain matrix leads to numerical
the presence of topology errors, and the error percentage in complexities, and the GN method is found to work well
the estimated values is considerably reduced in the presence only when the initially guessed value is close to the solu-
of bad data. The simulation results bring out the effectiveness tion. Though an incrementalized variant of GN method [2]
of these methods. is found to be more efficient, its performance deteriorates
when significant nonlinearity is in the objective function.
A robust SE algorithm shall be capable of overcoming a
& S. Bindu couple of problems. Several methods using artificial
[email protected]
intelligence techniques are discussed in the earlier resear-
1
Government Engineering College, BartonHill, ches [3] where particle swarm optimization technique is
Thiruvananthapuram, India discussed as an alternative for the conventional GN tech-
2
Mar Baselios College of Engineering, Thiruvananthapuram, nique. But this method does not find a solution to the
India numerical ill-conditioning problem of the gain matrix.
3
College of Engineering Trivandrum, Thiruvananthapuram,
India

123
 J. Inst. Eng. India Ser. B

The issue due to numerical ill-conditioning is partially State Estimation using Gauss–Newton Method
solved with orthogonal transformation method which is
referred to as QR-GN [4, 5]. But, even QR-GN method GN method is considered as the most common and popular
fails to give a converged solution under certain conditions. technique for solving SE problems. In this method, the
Trust region (TR) method provides solution to the afore- components of the objective function are approximated to a
mentioned issue. In the TR method, the SE is articulated as linear function in the neighbourhood of the estimate ‘x’.
a nonlinear problem where a quadratic model is trusted to The measurement equation for a given set of measured
be a sufficient model of the objective function in the trust parameters is expressed as (1) in SE:
region of the problem space. The radius of the TR is 2 3 2 h ðx ; x ; x ; . . .x Þ 3 2 3
Z1 m1 1 2 3 n e1
controllable and is considered as a measure of progress
6 Z2 7 6 hm2 ðx1 ; x2 ; x3 ; . . .xn Þ 7
7 6 e2 7
toward convergence. Application of the TR method for 6 Z 7 6 6 hm3 ðx1 ; x2 ; x3 ; . . .xn Þ 7 6 e3 7
Z¼6 6 7 7þ6 7
7¼6 ð1Þ
3
power system SE is backed by its strong theoretical 4 .. 5 6
. .. 6
7 4 .. 7
5
4 . 5 .
framework, practical feasibility, numerical efficiency and
robustness [6–10]. Zm hmm ðx1 ; x2 ; x3 ; . . .xn Þ em
Performance of the SE method depends on the technique where the measurement vector is defined as
applied and the quality of measurements. SCADA system
provides data measured in real time to classical SE Z ¼ ½ Z1 Z2 Z3 . . . Zm  T
schemes. The measured data include power injections, line hm ð x Þ ¼ ½ hm 1 hm 2 hm3 . . . hmm T
flow data, magnitude of voltages at different nodes, etc.
hm(x) is a vector that defines the linkage between the state
The precision of a SCADA-dependent SE method is low as
vector and measurement vector. The vector hm(x) has ele-
the accuracy, and update rate of parameters is less. SCADA
ments which are nonlinear functions.
cannot provide measured data like voltage, current and its
State vector to be measured is expressed as
phase angles in real time. The synchronized measurement
of phasors (voltage and current) by PMUs can greatly X ¼ ½x1 x2 x3 . . . xn T
improve the efficiency, quality and accuracy of power
The measurement error vector ‘r’ is expressed as
system SE methods [14–18]. Considering the cost and
infrastructure overheads for a PMU-monitored electric r ¼ ½e1 e2 e3 . . .em T
network, the number of PMUs is to be optimized ensuring
The rank of system Jacobian matrix should be ‘n’ and
adequate visibility.
m  n where m is the number of measurements.
In this paper, a hybrid SE method incorporating PMU
‘x’ is the optimal state estimate vector which is evalu-
measurements along with TR methods like IGNPDL and
ated by minimizing the objective function, i.e. the sum of
RISE is proposed to solve the issues in conventional SE
weighted squares of residuals as in (2):
using the GN technique. The accuracy of estimate is found
to improve in hybrid methods which integrates these SE Min J ð xÞ ¼ ½z  hm ð xÞT W ½z  hm ð xÞ ð2Þ
methods with synchronized phasor measurements [19].
where ‘W’ is the weighting factor, the inverse of which is
The researchers [20] introduced a linear state estima-
the measurement variance.
tion with conventional and PMU measurements. The power
The minimum can be found out by equating the first
system is modelled as a fully linear equivalent circuit. A
derivative of J (x) to zero.
suitable optimization problem is formulated to derive the
system states. Application of deep neural network for oJ ð xÞ
g1 ð x Þ ¼ ¼ H T ð xÞW ½z  hm ð xÞ ¼ 0 ð3Þ
monitoring of power system in real time is addressed in the ox
literature [24], and physics-based deep learning method is The measurement Jacobian matrix (H(x)) has dimensions
discussed by the researchers [25]. These methods claim to m 9 n and
be easily trainable and computationally inexpensive, but
the drawback is the large computation time.

123
J. Inst. Eng. India Ser. B

ohm ð xÞ
H ð xÞ ¼ ð4Þ
ox
The linearization of the nonlinear function hm(x) can be
expressed as
hm ðx þ DxÞ  hm ð xÞ þ H ð xÞDx ð5Þ
An iterative approach is used to solve Eq. (3) is given as:
ðH T R1 HÞDx ¼ H T W ½z  hm ð xÞ ð6Þ
xkþ1 ¼ xk þ Dx ð7Þ
G = HT R-1 H is a symmetric matrix, and it is also called
gain matrix or information matrix. Iteration index is
denoted as ‘k’.
In weighted least square estimation algorithm, normal
equation is expressed as in the literature (6).
A flat start with the assumed value of phase voltage 1.0
p.u., for all phase voltages, is the method followed for state Proposed Methods
variable initialization. The termination of iterative steps is
done when the mismatch in estimates reaches a set A more robust and reliable SE scheme which can effec-
threshold. tively handle problems like topological errors in power
The reference bus (bus 1) has a set value of phase angle system models can be implemented using a globally con-
equal to zero degree. Thus, there are (2n - 1) elements in vergent modification of GN method using the TR method.
the state vector. This includes ‘n’ bus voltages and (n - 1) A least square solver with the robustness of steepest des-
phase angles as expressed in (8). cent (SD) method having a convergence rate of GN method
xT ¼ ½d2 ; d3 ; . . .; dn ; v1 ; v2 ; v3 . . .vn  ð8Þ will be highly promising in SE problems. Among trust
region algorithms, IGNPDL and RISE methods are suit-
Magnitude of voltage and value of real/reactive power
able for sparse least square minimization problems
injections and power flows are the elements in the
[11–13]. These methods ensure robustness when there
measurement vector. Thus, the measurement Jacobian (H)
exists nonlinearity and numerical ill-conditioning with the
has a structure as given as follows:
2 3 objective function.
oVmag A model of the objective function is constructed on the
6 oP0 oV 7 neighbourhood of the current estimate in each iteration of
6 inj
oPinj 7
6 7 the Newton’s method. The step length which minimizes the
6 od 7
6 oQinj oV 7 objective function is then determined. But this depends on
6 oQinj 7
H¼6 6 od
7
7 ð9Þ the local behaviour of the objective function near the
6 oPij oV 7
6 oPij 7 current estimate and the first and second derivatives of the
6 od 7
6 oQ oV 7 objective function. When the update step leaves the region
4 ij
oQij 5
od where the objective function is well approximated, this
oV becomes a problem. In the proposed methods, this problem
where Vs is the magnitude of bus voltage, Pinj—real power, is tackled by maintaining a region of trust which is centred
Qinj—reactive power, Pij—real power flow between bus on the current estimate where the objective function is well
i and j, Qij—reactive power flow between bus i and bus j. approximated. It is important that the problem of diver-
gence needs to be avoided by carefully choosing the update
steps that minimize the objective function while being
enclosed in the trust region.
As the method advances, the radius of the region is
tuned based on the gain ratio. Gain ratio is the ratio
between actual reductions in the value of objective function
to the reduction predicted. The objective function is con-
sidered to be performing well if the gain ratio magnitude is

123
 J. Inst. Eng. India Ser. B

near to unity. Then, the radius and therefore step length can
be increased so that faster convergence is achieved.
The main challenge with the accuracy of GN method is
when the gain matrix has numerical ill-conditioning. This
situation mainly occurs in the case of conditions like
topological/parameter errors and bad data. The proposed
methods-based SE schemes are found to be very reliable
under these conditions also.

Powell’s Dog Leg Method (PDL)

The lack of robustness of GN method even under ideal

conditions is a disadvantage of the method. PDL algorithm
When the step length (hdl) is obtained, the value of the
addresses this shortcoming. This method works by com-
state variable is updated as
bining GN method and the steepest descent (SD) direc-
tions. The radius of trust region (TR) controls the choice of xkþ1 ¼ xk þ hdl
the step length. Step length in each iteration is updated as a The maximum value of hdl is then compared with the
function of TR radius. This algorithm performs signifi- tolerance to decide whether to continue the iterations or
cantly faster with reasonable accuracy for sparse least- not. If the value of hdl is less than tolerance, the iterations
squares minimization problems. PDL possesses quick are stopped, otherwise the iterations are continued till
convergence property of GN algorithm and global con- convergence is obtained.
vergence contributed by SD methods. Based on the gain ratio (q), the radius of TR is adap-
The region of trust in PDL is assumed as a ball of radius tively varied while proceeding with PDL
(D), which has a centre at the current linearization point
‘X’. The radius of the TR is considered as the measure of J ð xÞ  J ðx þ hdl Þ
D¼ ð12Þ
progress, and this is a quantity that can be controlled. The Lð0Þ  Lðhdl Þ
correctness of predicted behaviour of the objective func- where
tion, by the local model, decides whether to expand or
compress the TR. L ¼ HT h þ r ð13Þ
While estimating a dog leg (DL) step hdl, GN updates J is the objective function to be minimized. And the step
are chosen as it leads to better speed and reliability. It is to size hdl is derived from hgn.
be ensured that the updates are falling in the TR. The reduction in the value of the objective function is
The size of the SD step hsd can be considered as a vector estimated by selecting the proposed DL step, hdl. If the
parallel to the residue value of the gain ratio is near unity, it can be assumed that
hsd ¼ 0:5  delðJ Þ ð10Þ the algorithm works well near the current point. This can
be followed by increment of steps in larger magnitude to
The GN step hgn is computed as
  achieve the convergence rapidly. If the gain ratio is close to
hgn ¼ B1  g1 ð11Þ zero, the approximation is poor and the TR radius is to be
decremented accordingly.
where B is the Hessian matrix and g1 is the first derivative
of the objective function.

123
J. Inst. Eng. India Ser. B

Indefinite Gauss–Newton–Powell’s Dog Leg Method

(IGNPDL)

IGNPDL method is based on PDL with GN steps. In this

method, a model or the approximation of the objective
function is minimized within a specific region.
In the case of a full rank J(x), it can be seen that the
Hessian matrix B is greater than zero (B [ 0) and hdl can
be calculated in terms of hsd. If the objective function is not
of full rank and B is not positive definite, hdl cannot be
evaluated. In such a case, a Cauchy’s step can be used. If
the minimum of the model is found along the Cauchy arc,
the point at which the largest decrease obtained is referred
as Cauchy point (hc).
A unique Cauchy’s step always exists if g1 (first
derivative of the objective function) is not equal to zero.
This algorithm is thus derived from the PDL algorithm
selecting PDL steps for a full rank J(x) and Cauchy steps if
not.
The Cauchy step is defined as
hc ¼ hdl ¼ ja ð14Þ
g1 ¼ H T r ð15Þ Robust Incrementalized Least-Squares Estimation
8 ( ) (RISE)
>
>
> D jjg1 jj2
< min ; ; jjH ð xÞg1 jj [ 0
jjg1 jj 2jjH ð xÞg1 jj2
j ¼ RISE is an extended version of IGNPDL algorithm. When
>
> D
>
: b; jjH ð xÞg1 jj ¼ 0 B is non-singular, the IGNPDL algorithm estimates the
jjg1 jj approximate TR update step from the GN step hgn and the
ð16Þ SD step hsd. There is already an efficient method to com-
pute hgn; therefore, it remains only to compute hsd and hc
jjg1 jj2
a¼ ð17Þ efficiently. hsd and hc are calculated in terms of the gradient
jjH ð xÞg1 jj2 direction vector g1 and a and j, the scale factors.
RISE method uses the QR decomposition matrix to
evaluate the factors, which makes it simple and faster.
Calculation of ‘g1’ in RISE is as follows:
If [k, c] can be assumed to represent the QR decompo-
sition of the Jacobian matrix H
½k; c ¼ QRðH Þ ð18Þ
d ¼ kT  r ð19Þ
g1 ¼ c T d ð20Þ
Since the matrix multiplication is faster as c is a sparse
matrix with smaller dimension than that of H, it is
advantageous to use (20). To compute ||H(x)g1||2, R factor
of the QR decomposition is used, and since Q is
orthogonal,
jjH ð xÞg1 jj ¼ jjRg1 jj ¼ jjcg1 jj
jjg1 jj2 ð21Þ
a¼ 2
jjcg1 jj
and

123
 J. Inst. Eng. India Ser. B

8 ( )
> D jjg1 jj2 The placement of PMUs at appropriate buses needs to
>
>
< min ; ; jjcg1 jj [ 0 ensure complete visibility of the system. Genetic algorithm
jjg1 jj jjcg1 jj2
j¼ ð22Þ [23] is used to find the nodes of the network where the
>
> D
>
: ; jjcg1 jj ¼ 0 PMUs are to be deployed for complete visibility. It is
jjg1 jj identified that 4 nodes (nodes 2, 8, 10 and 13) require PMU
The algorithm of RISE method is summarized as follows: deployment in an IEEE 14-bus system and 32 nodes
require PMU in an IEEE 118-bus system to ensure com-
plete observability.
The proposed measurement model with PMU data is
therefore written as,
M ¼ Hc Vest þ e ð23Þ
Here, M is the measurement vector. The superscript ‘c’
is used for unique identification of the coefficient matrix H.
sest is the real and imaginary parts of the state estimates.
Gaussian error in the measurements is denoted as e.
Equation (24) describes the measurement model.
Different measurements used for constructing the mea-
surement vector (M) are
1. State output directly from SE using PDL/IGNPDL/
RISE method
2. Voltage measurements with PMU
3. Current measurements with PMU
The subscripts R and I are used to denote the real and
imaginary parts of voltages and current measurements.
Hc is the coefficient matrix with its elements being
Hm11 ¼ I; Hm12 ¼ 0:
Hm21 ¼ 0; Hm22 ¼ I;
State Estimation with PMU Measurements
The identity matrix I has an order n 9 n, n being the
PMUs are becoming popular in monitoring and control number of buses in the system. Null matrix (n 9 n) is
applications in electric power networks. The number of represented as 0.
PMUs required needs to be optimized due to high cost and If P is the number of PMUs used for measurements in
infrastructure requirement. the power network, Hm31 will be a (P 9 n) matrix. Each
There are two main approaches to combine PMU data row of Hm31 will indicate the presence of PMU as given
with the conventional measurement data—pre-processing below. The elements in any row with a PMU will be all
technique and post-processing technique. Both measure- zeroes except in the jth column where it will be corre-
ments are handled together in pre-processing technique, sponding to the index of VRkpmu representing PMU
whereas PMU measurement data are added to the output of measurements.
the existing state estimator in the post-processing technique
[20–22]. Pre-processing technique is difficult to be
employed in the existing software as the whole program
has to be modified to include the PMU measurements.
Post-processing technique is used in this paper as it is easy
to incorporate the changes in the existing algorithm.
The state estimates from the PDL, IGNPDL and RISE
methods are taken together with measurements of voltage
and current phasors with PMU to form the measurement
vector (M). Direct measurements of angle from PMUs are
also included. This eventually enhance the performance of
the SE method. Fig. 1 Model (pi) of a transmission branch

123
J. Inst. Eng. India Ser. B

Table 1 Comparison of MAE for different methods while estimating 0

voltage magnitude -2 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Voltage angle (degree)

IEEE 14-bus IEEE 118-bus -4
system system -6
-8
GN method 0.0539 0.135
-10
PDL method 0.0134 0.044
-12
PDL method with PMU 0.0112 0.035
-14
IGNPDL method with 0.0055 0.032
-16
PMU
-18
RISE method with PMU 0.005 0.032 Bus number

PDL with PMU True value IGNPDL with PMU

Fig. 3 True and estimated values of voltage angle in an IEEE 14-bus

Table 2 Comparison of MAE for different methods while estimating
system
voltage angle
IEEE 14-bus IEEE 118-bus
system system
0.08

Voltage magnitude error(p.u.)

GN method 1.369 1.65 0.07
PDL method 0.356 0.71 0.06
PDL method with PMU 0.323 0.685 0.05
IGNPDL method with 0.313 0.0789 0.04
PMU 0.03
RISE method with PMU 0.187 0.0785 0.02
0.01
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14
1.1 Bus number
Gauss Newton PDL with PMU RISE with PMU
1.08
Voltage magnitude (pu)

1.06 Fig. 4 Comparison of voltage magnitude errors for an IEEE 14-bus

1.04
system

1.02
transmission branch as depicted in Fig. 1, the current Ipq
1 can be expressed as
0.98  
Ipq ¼ Ep Ypo þ Ep  Eq Ypq ð25Þ
0.96
0.94 Hm51 is a (P 9 n) matrix. Elements of Hm51 are pure
1 2 3 4 5 6 7 8 9 10 11 12 13 14 conductances. Hm52 is a (P 9 n) matrix where elements are
Bus Number pure susceptances. It is to be noted that Hm61 is also a
True value PDL with PMU (P 9 n) matrix with only susceptance elements. Hm62 is a
IGNPDL with PMU (P 9 n) matrix where all elements are conductances.
Using the relation (26), the state estimation problem is
Fig. 2 True and estimated values of voltage magnitude in an IEEE
14-bus system solved directly without iteration:
 1
Hm31k ¼ ½0 0 . . . 0:: 1:: 0 Vest ¼ HcT ðR1Þ1 Hc HcT ðR1Þ1 M ð26Þ
Hm32 ¼ 0 ðP  nÞ
Hm41 ¼ 0 ðP  nÞ The diagonal covariance matrix is defined by (27), and it is
Similarly, Hm42 is a (P 9 n) matrix. As explained assumed that it has a zero mean.
before, the elements in any row with a PMU will be all nh io
R1 ¼ diag r2jVRSE j ; r2jVISE j ; r2jVRPMU j ; . . .; r2jIIPMU j ð27Þ
zeroes except in the jth column where it will be corre-
sponding to the index of VIkpmu in the vector representing
Each element of R1 is diagonal sub-matrices, which is
PMU measurements.
shown as follows:
Hm42k ¼ ½0 0 . . . 0:: 1:: 0
Hm51, Hm52, Hm61 and Hm62 are elements of Hc repre-
senting PMU current measurements. From the model of the

123
 J. Inst. Eng. India Ser. B

Voltage angle error(degree) 4 1.2

Voltage magnitude (pu)

1
3
0.8

2 0.6

0.4
1
0.2

0 0

101
106
111
116
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
1 2 3 4 5 6 7 8 9 10 11 12 13 14
Bus number True value Bus number
Gauss Newton PDL with PMU RISE with PMU IGNPDL with PMU

Fig. 5 Comparison of voltage angle error for an IEEE 14-bus system Fig. 8 True and estimated values of voltage magnitude in an IEEE
118-bus system
1.8
Voltage magnitude (pu)

15
1.6

Voltage angle (degree)

1.4 10
1.2
5
1
0.8 0

1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
0.6 -5
0.4
-10
0.2
0 -15
101
106
111
116
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
1
6

-20
Bus number
-25
True value PDL with PMU Gauss Newton Bus number
True value IGNPDL with PMU
Fig. 6 True and estimated values of voltage magnitude in an IEEE
118-bus system Fig. 9 True and estimated values of voltage angle in an IEEE
118-bus system
15
Voltage angle(degree)

1.2
Voltage magnitude(pu)

10

5 1

0 0.8
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116

-5 0.6
-10 0.4
-15
0.2
-20
Bus number 0
101
106
111
116
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
1
6

-25
True value PDL with PMU Gauss Newton
Bus number
True value RISE with PMU
Fig.7 True and estimated values of voltage angle in an IEEE 118-bus
system Fig. 10 True and estimated values of voltage magnitude in an IEEE
118-bus system
nh io
r2jVRSE j ¼ diag r2jVR1 j ; r2jVR2 j ; r2jVR3 j ; . . .r2jVRN j ð28Þ
Results and Discussion
The SE is performed in two parts. In the first part,
SCADA measurements which are non-synchronized are The SE techniques proposed in the paper are simulated and
used for SE by PDL, IGNPDL and RISE methods. Linear evaluated using MATLAB. As discussed, the simulation
SE is done in the second part incorporating synchronized studies are carried out on IEEE 14-bus and IEEE 118-bus
measurements. The accuracy of SE is improved by using systems. The results are compared with the results of power
PMU measurements. flow studies. SE results obtained with inclusion of PMU
measurements and without PMU are also analysed. The
minimum number of PMUs that ensures complete visibility

123
J. Inst. Eng. India Ser. B

Voltage angle(degree) 15 12
10 10

log||(z-h(x))||2
5 8
0
6
1
6
11
16
21
26
31
36
41
46
51
56
61
66
71
76
81
86
91
96
101
106
111
116
-5
4
-10
2
-15
0
-20 1 2 3 4 5 6
-25
Bus number iteraon
True value RISE with PMU
Fig. 14 Convergence curve for state estimation with topology error
Fig. 11 True and estimated values of voltage angle in an IEEE in an IEEE 14-bus system using RISE method
118-bus system

Table 3 Comparison of computational time in SE using different methods

GN method PDL with PMU IGNPDL with PMU RISE with PMU

Computational time for an IEEE 118-bus system 1.2831 s 0.0159 s 0.0178 s 0.0148 s

25 25
log||(z-h(x))||2

20 20
log||(z-h(x))||2

15
15
10
10
5
5
0
1 3 5 7 9 11 13 15 17 19 21 23 25
0
iteraon 1 2 3 4 5 6 7 8 9 10 11 12
Iteraon
Fig. 12 Convergence curve for state estimation with topology error
in an IEEE 14-bus system using GN method Fig. 15 Convergence curve for state estimation with topology error
in an IEEE 118-bus system using RISE method
30
The performance of the three methods is compared
25
based on the mean absolute errors (MAE) given by each
log||(z-h(x))||2

20 and the execution time. Tables 1 and 2 show the MAE of

state estimates when different methods are used for com-
15
putation. The RISE method is faster than the other methods
10 because matrix multiplication is faster in RISE method as
matrix c is a sparse matrix and is smaller in dimension.
5
Figures 2 and 3 show the comparison of true values with
0 estimated values using the PDL method with PMU and the
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
IGNPDL method with PMU for voltage magnitude and
Iteraon angle in an IEEE 14-bus system. The errors are less for
IGNPDL method, and the estimated values from the
Fig. 13 Convergence curve for state estimation with topology error
in an IEEE 118-bus system using GN method
IGNPDL method follow the true values.
The errors in voltage magnitude and voltage angle
of the system is obtained through genetic algorithm. PMUs estimates using conventional GN and RISE methods at
are placed in these buses, and the voltage phasors and different buses in an IEEE 14-bus system are illustrated in
current phasors are measured. Figs. 4 and 5. As given by the MAE values, the errors are

123
 J. Inst. Eng. India Ser. B

very high for conventional GN method as compared to matrix is tackled in the IGNPDL and RISE methods by
PDL method with PMU and RISE methods. In the RISE finding the Cauchy step. This ensures convergence.
method, the lowest error is obtained for magnitude and Errors and bad data may be unintentional or intentional.
angle of voltage phasors and the execution time is also the Communication failure, transducer failure or tampering by
lowest. The proposed methods show better accuracy hackers are the reasons for the occurrence of bad data in
because of the property that the values are always held measurements. There are post-estimation bad data pro-
within the trust region. The execution time is also reduced cessing methods to detect, identify and eliminate bad data.
because the convergence is made faster due to the same But these methods have high computational cost and can-
reason. not be efficiently applied at fast rates. The state estimation
The estimated values using GN and PDL methods with methods itself should be robust enough to deal with bad
PMU data and the true values for an IEEE 118-bus system data. The IGNPDL and RISE methods can handle bad data
are shown in Figs. 6 and 7. The values obtained from the due to the application of Cauchy step in the case of sin-
GN method deviate much from the actual values, but the gularities in the presence of bad data.
values from the PDL method with PMU show less errors. Active power injection measurement at bus 25 and
The comparison of true values with the calculated values active power flow measurement between buses 5 and 6 are
from the IGNPDL method for an IEEE 118-bus system is set to zero for an IEEE 118-bus system in order to intro-
shown in Figs. 8 and 9. Voltage magnitude values and duce bad data.
voltage angles are compared in these figures. The error percentage in mean square error of prediction
RISE with the PMU method is also used to estimate the (MSEP) is then calculated as.
states of an IEEE 118-bus system, and the results are MSEP value is nearly 25% for Gauss Newton state
shown in Figs.10 and 11. The errors are the lowest in this estimator due to its vulnerability to bad data, while it is
method, and the time taken for estimation is also the lowest only around 5% for IGNPDL- and RISE-based SE.
for this method as shown in Table 3.
The computational time taken for the estimation of
states by various methods for an IEEE 118-bus system is Conclusion
shown in Table 3. The time taken is the highest for Gauss–
Newton method due to computational complexities. The Modified state estimation using Powell’s dog-leg-based
computation becomes simple in the PDL method, IGNPDL algorithm (PDL), indefinite Powell’s dog leg numerical
method and RISE method. The computation time is the optimization technique (IGNPDL) and robust incremen-
lowest in RISE method due to the sparsity of c matrix. talized least-squares estimation (RISE) method incorpo-
Another advantage of the proposed methods is their rating conventional measurements and synchronized
performance in the presence of bad data and topology phasor measurements are proposed and evaluated in this
errors. The gain matrix in the GN method is ill-conditioned paper. It is observed that the accuracy of the state esti-
in the presence of bad data and topology errors. This mation is improved in the proposed methods when com-
problem is tackled in the proposed methods. pared with the conventional SE. The computation time is
The results also illustrate the advantage of the proposed also found to improve with the proposed methods. The
methods which can adapt with topology error where the proposed methods also show better performance even in
GN method is unable to reach a commendable solution. In the presence of bad data and topology errors. The robust-
the presence of topology errors, a converged solution ness and effectiveness of the proposed methods are verified
(although not accurate) is required before any error successfully by comparison with the results of conventional
detection algorithms can be implemented. The GN method GN method with and without bad data and topology errors.
fails to converge when there are topology errors, whereas The simulation results show good accuracy and better
the proposed methods (PDL, IGNPDL and RISE methods) convergence time in the estimated states and improved
converge even in the presence of a topology error. performance with bad data and topology errors.
Topology error is introduced in the system by the exclusion
of transmission line 6–11 in an IEEE 14-bus system and
transmission line 60–61 in an IEEE 118-bus system. From Funding No funding.
Figs. 12 and 13, it is seen that the GN method fails to Declarations
converge, whereas the RISE method with PMU converges
in a few iterations as illustrated in Figs. 14 and 15. Conflict of interest The authors declare that there is no conflict of
IGNPDL and RISE methods work even when there are interest.
topology errors or bad data. Ill-conditioning in the Hessian

123
J. Inst. Eng. India Ser. B

References 14. R.F. Nuqui, A.G. Phadke, Hybrid linear state estimation utilizing
synchronized phasor measurements. IEEE Lausanne Power
1. A. Monticelli, Electric power system state estimation. Proc. IEEE Technol. 56, 1665–1669 (2007)
88, 262–282 (2000) 15. K. Jamuna, K.S. Swarup, Two stage state Estimator with phasor
measurements, in 2009 International Conference on Power Sys-
2. J. Nocedal, S. Wright, Numerical Optimization, 2nd edn.
(Springer, New York, 2006) tem, pp. 1–5 (2009).
3. S.K. Sikarwar, V. Kumar, R. Naresh, Power system state esti- 16. L.T. Valeriu Presad, M. Eremia, Modified state estimation in
mation using particle swarm optimization. J. Adv. Res. Electr. presence of PMU measurements, in 2014 U.P.B. Sci. Bull., Series
C, vol. 76 (2014).
Electron. Eng. 4(3), 1–35 (2017)
4. P. Yang, Z. Tan, A. Wiesel, A. Nehorai, Power system state 17. J.S. Thorp, A.G. Phadke, K.J. Karimi, Real-time voltage-phasor
estimation using PMUS with imperfect synchronization. IEEE measurements for static state estimation. IEEE Power Eng. Rev.
Trans. Power Syst. 28, 4162–4172 (2013) PER-5, 32–33 (1985)
5. A. Simoes-Costa, V.H. Quintana, A robust numerical technique 18. A. A. Antonio Gomez-Exposito, On the use of PMU in power
for power system state estimation. IEEE Trans. Power Appl. Syst. system state estimation, in 17th Power system computation con-
PAS-100, 691–698 (1981) ference, Sweden (2011).
6. V. Seshadri Sravan Kumar, D. Thukaram, State estimation in 19. G. Korres, An integer-arithmetic algorithm for observability
power systems using linear model infinity norm-based trust analysis of systems with Scada and PMU measurements. Electr.
region approach. IET Gener. Transm. Distrib. 7(5), 500–510 Power Syst. Res. 81, 1388–1402 (2011)
(2013) 20. A. Jovicic, M. Jereminov, L. Pileggi, G. Hug, A linear formula-
tion for power system state estimation including RTU and PMU
7. M. Zhou, V. Centeno, J. Thorp, A. Phadke, An alternative for
including phasor measurements in state estimators. IEEE Trans. measurements, in IEEE PES ISGT Conference Europe (2019).
Power Syst. 21, 1930–1937 (2006) 21. M.R. Wagner, M. Jereminov, A. Pandey, L. Pileggi, A proba-
8. S. Chakrabarti, E. Kyriakides, G. Ledwich, A. Ghosh, Inclusion bilistic approach to power system state estimation using a linear
of PMU current phasor measurements in a power system state algorithm, in Proceedings of 19th IEEE International Conference
estimator. IET Gener Trans. Dist. 4, 1104–1115 (2010) on Environment and Electrical Engineering, Italy (2019).
9. A.R. Conn, N.I.M. Gould, P.L. Toint, Trust region methods. 22. T. Jin, X. Shen, A mixed WLS power system state estimation
Society for Industrial and Applied Mathematics and Mathematics method integrating a wide-area measurement system and SCADA
Programming Society (2000). technology. Energies J. (2018).
10. S. Pajic, K.A. Clements, Power system state estimation via 23. N.P. Theodorakatos, Optimal phasor measurement unit placement
globally convergent methods. IEEE Trans. Power Syst. 20, for numerical observability using branch-and-bound and a binary-
coded genetic algorithm. Electr. Power Comp. Syst. 47, 4–5
1683–1689 (2005)
11. M. Powell, On the global convergence of trust region algorithms (2019)
for unconstrained minimization. Math. Program. 29(3), 297–303 24. L. Zhang, G. Wang, G.B. Giannaki, Real-time power system state
(1984) estimation and forecasting via deep unrolled neural networks.
IEEE Trans. Signal Process. 67(15), 5–60 (2019)
12. G. Shultz, R. Schnabel, R. Byrd, A family of trust-region-based
algorithms for unconstrained minimization with strong global 25. L. Wang, Q. Zhou, S. Jin, Physics-guided deep learning for power
convergence properties. SIAM J. Numer. Anal. 22(1), 47–67 system state estimation. J. Mod. Power Syst. 8(4), 56–80 (2020)
(1985)
13. D.M. Rosen, M. Kaess, J.J. Leonard, RISE: an incremental trust Publisher’s Note Springer Nature remains neutral with regard to
region method for robust online sparse least-squares estimation. jurisdictional claims in published maps and institutional affiliations.
IEEE Trans. Robot. 30(5), 1091–1108 (2014)

123