Received: 16 July 2020 Revised: 27 September 2020 Accepted: 1 November 2020

DOI: 10.1002/2050-7038.12714

RESEARCH ARTICLE

An improved algorithm for cubature Kalman filter based

forecasting-aided state estimation and anomaly detection

Zhaoyang Jin1 | Saikat Chakrabarti2 | James Yu3 | Lei Ding1 |

Vladimir Terzija4

1
School of Electrical Engineering,
Shandong University, Jinan, China
Summary
2
Department of Electrical Engineering, This article proposes a new algorithm for forecasting-aided state estimation
Indian Institute of Technology, Kanpur, based on the cubature Kalman filter (CKF) and new methods for detecting and
India
identifying data anomalies. In this article, through extensive simulations, the
3
Future Networks, SP Energy Networks,
CKF was compared to four different types of forecasting-aided state estimators
Glasgow, UK
4
Center for Energy Science and (FASEs) including extended Kalman filter (EKF), iterated EKF, second-order
Technology (CEST), Skolkovo Institute of Kalman filter and unscented Kalman filter under normal operation and bad
Science and Technology (Skoltech), data conditions. Identifying the challenge that the estimation accuracy of the
Moscow, Russia
existing CKF-based estimator is significantly lower than that of the other
Correspondence FASEs in the cases of sudden load change, and sudden topology change caused
Lei Ding, School of Electrical
by faults an attempt to improve the CKF accuracy has been undertaken. The
Engineering, Shandong University, 17923
Jingshi Road, Jinan, China. existing detection methods cannot accurately detect and distinguish those
Email: [email protected] anomalies, and they cannot identify the anomaly location. This article pro-

Handling Editor: Jalilzadeh-

poses an improved algorithm for CKF-based FASE that overcomes the draw-
Hamidi Reza backs of the existing CKF-based FASEs using a novel anomaly detection
algorithm. The simulation results show that the new anomaly detection
method is superior to the two existing anomaly detection algorithms. The sim-
ulations are performed in the above-mentioned four cases in IEEE 14 and
118 bus test systems in MATLAB.

KEYWORDS
phasor measurement unit, power system protection, synchrophasor, wide area measurement,
wide area protection

1 | INTRODUCTION

The power system requires persistent monitoring of the operating states, described by voltage angles and magnitudes of
all buses in the network, to guarantee secure and economic operation. In the early 1970s, the concept of static state esti-
mation (SSE) was firstly proposed by Schweppe.1-3 Since then, the SSE has been widely applied in power systems to pro-
vide accurate information of the states based on a single set of measurements. However, SSE requires a full set of

List of Symbols and Abbreviations: CKF, cubature Kalman filter; EKF, extended Kalman filter; FASE, forecasting-aided state estimator; IEKF,
iterated extended Kalman filter; NI, normalized innovation; NR, normalized residual; PMU, phasor measurement unit; SCADA, supervisory control
and data acquisition; SIR, skewness to innovation ratio; SKF, second-order Kalman filter; SSE, static state estimator; UKF, unscented Kalman filter;
WAMS, wide area monitoring system.

Int Trans Electr Energ Syst. 2021;31:e12714. wileyonlinelibrary.com/journal/etep © 2021 John Wiley & Sons Ltd 1 of 23
https://doi.org/10.1002/2050-7038.12714
2 of 23 JIN ET AL.

measurements to estimate the states, and is susceptible to bad data or loss of data in the measurements. The drawbacks
of SSE can be overcome using the forecasting-aided state estimator (FASE) that uses both the measurements and pre-
dictions. The FASE is especially suitable for distribution systems where measurements are usually insufficient. There-
fore, the FASE is becoming increasingly important in modern power systems, especially with the development of a
wide area monitoring system (WAMS).4
The FASE is composed of a dynamic model that generates predictions based on the previous states, and a nonlinear
filter to estimate the states from the predictions and measurements. The dynamic model for FASE developed from over-
simplified models.5-7 The predictions provided by these methods are not accurate, but the basic structure of the FASE
was established. Other simplified methods include the tracking estimators8-10 in which the current states are assumed
to be the same with their previous states. In Reference 11, the Holt's exponential smoothing technique was proposed,
which made the predictions based on the trend of past states. This method is very accurate if the states of the power sys-
tem progress slowly in a stable trend. Later, the method with artificial neural network and pattern analysis was pro-
posed in References 12 and 13. This method can capture the hourly, daily changing pattern of the loads, which makes
prediction based on the pattern. Recently, a series of papers14-17 proposed to use the synchronous machine dynamic
equations to produce the predictions. The predictions made through this method can capture the transient dynamic
process of the network. Thus, they are called dynamic state estimators to distinguish FASEs.18 In spite of its advantages,
the dynamic state estimator is unable to forecast the long-term or even the short-term change of the states. Therefore,
dynamic state estimator will not be further discussed in this article.
In literature, nonlinear Kalman filters are adopted for most of the proposed FASEs because of its simplicity and unbi-
asedness. The most widely applied nonlinear Kalman filters for state estimation include the extended Kalman filter
(EKF),19 and unscented Kalman filter (UKF).20 To minimize the loss of accuracy due to the truncated higher-order terms
in the Taylor expansion in EKF, iterated EKF (IEKF)21 and second-order Kalman filter (SKF)22 have been proposed. UKF
suffers from the problems such as numerical inaccuracy and unavailability of a square root solution due to negatively
weighted sigma points.23 Therefore, cubature Kalman filter (CKF)23,24 with optimal location of sigma points have been
proposed recently to solve the problem. Although IEKF, SKF and CKF claimed their advantages in their published papers,
their performance has not been thoroughly compared in all cases where the FASE might work and in different scale net-
works. Recently, the ensemble Kalman filter25 and particle filter26 have been reported to be more resistant to anomalies,
but their execution time is much longer.26 Here, anomalies mean the naturally happened events that can degrade the esti-
mation accuracy of FASE, including bad data, sudden load change and sudden topology change due to open circuit faults.
Under normal conditions, the state change of the power system is slow and predictable.27 However, sudden change of
states may occur due to load change or network topology change. In such cases, the estimation accuracy of FASEs would be
even lower than that of SSE. Therefore, FASEs need to detect these anomalies, distinguish them from the bad data and then
to reject or down-weigh the erroneous data. References 28 and 29 proposed a method that automatically down-weights the
anomalies based on the normalized residuals (NRs) of the measurements. The term NR is defined as the normalized value
of the difference between the measurements and the estimated measurements. However, NRs are unable to distinguish
between sudden state changes and bad data. Both predictions and measurements will be down-weighed if any sudden state
change occurs. In References 30 to 32, a new anomaly detection method based on innovation analysis was proposed. The
term innovation analysis is defined as the process of comparing the normalized innovations (NIs) of the measurements and
the skewness of the NIs with their corresponding thresholds to determine the type of anomaly happened. Here, the NIs are
the normalized value of the difference between the measurements and the predicted measurements. This method was
widely adopted by most of the papers on FASE.11,20,21,28,29,33-37 However, as pointed out in Reference 18, the detection
criteria of this conventional innovation analysis method are unreliable—they are different with different network sizes and
measurement noise levels. The latest works on anomaly detection use robust FASEs.16,17,38-42 However, this method still can-
not detect the sudden load change if the change of states is small in comparison to the change of NI of the measurements.
In addition, both the innovation analysis method and the robust FASE have not considered the detection of the topology
change after fault, and the identification of the anomalies. This article proposed an improved innovation analysis algorithm
that can accurately detect, discriminate all three types of anomalies and identify the location of the anomalies.
This article has the following two contributions:

1. Proposition of a new CKF-based FASE algorithm that overcomes the drawbacks of the existing CKF-based FASE
algorithm that its estimation accuracy is much worse in a long period of time after load sudden change; It is also
more accurate than the other FASEs in four different power system operation cases, including normal operation
case, bad data case, sudden load change case and fault case.
JIN ET AL. 3 of 23

2. Proposition of an improved innovation analysis method that is much more accurate than the conventional innova-
tion analysis method and the robust FASEs. In addition, new methods are also proposed to identify the locations of
sudden load change and fault based on pseudo injection measurements innovation analysis and branch power flow
variation ratios, respectively.

Note that this article only considers anomalies caused by nonhuman factors. The anomalies caused by attackers
such as leverage points or false data injection attacks43,44 are not considered.
This article is structured as follows. Section 2 introduces the system dynamic model based on which the FASE pre-
dicts the states; Section 3 presented the formulations of the five above-mentioned state estimators; Section 4 discusses
the proposed anomaly detection and identification method; Section 5 presents the proposed FASE algorithm based on
CKF; Section 6 gives the simulation settings; Section 7 presents the results and analysis of the simulations in four cases;
Finally, Section 8 concludes this article.

2 | SYSTEM DYNAMIC MODEL

This section introduces mathematical models used in FASEs including the dynamic model of the system based on
which the prediction of the states is produced.
The general form of the system state transition function can be represented as,

x k + 1 = F k x k + gk + q k ð1Þ

where x is the state vector, F is the transition matrix, g is the input vector, q is the process error vector and the subscript
k denotes the instant.
The derivation of the parameters is determined by the system dynamic model used. To investigate the performance
of the FASE when an anomaly occurs, Holt's linear exponential smoothing technique that is more accurate for short-
term forecasting is adopted in this article. The parameters, F and g, are obtained as follows:

Fk = αk ð1 + βk ÞI ð2Þ

gk = ð1 + βk Þð1 −αk Þxk− −βk ak − 1 + ð1 −βk Þbk − 1 ð3Þ

where I is the identity matrix, and αk and βk are factors lying in the range from 0 to 1, xk− is the predicted state at time
k. The parameter vectors a and b at time k can be calculated as:

ak = αk xk + ð1 −αk Þxk− ð4Þ

bk = βk ðak −ak − 1 Þ + ð1 −βk Þbk − 1 ð5Þ

In all simulations presented in this article, for all k, αk = 0.8, βk = 0.5. Note that these two values are chosen empiri-
cally, based on various simulation results.

Remark. State forecasting is different from pseudo measurements in conventional state estimation, and is much more
accurate. This is because it uses more up-to-date data, and makes predictions based on all of the states rather than
a given measurement.

3 | N O N L I N E A R KA L M A N F I L T E R

In this section, the mathematical formulations of the four above-mentioned nonlinear Kalman filters (EKF, UKF, CKF,
IEKF and SKF) are presented. All of these nonlinear Kalman filters can be used to solve the forecasting-aided state esti-
mation problem as formulated below:
4 of 23 JIN ET AL.

xk = f ðxk − 1 , k − 1Þ + qk − 1 ð6Þ

zk = hðxk , k Þ + rk ð7Þ

σ 2 ðqk − 1 Þ = Qk − 1 ð8Þ

σ 2 ðr k − 1 Þ = R k − 1 ð9Þ

where f is the general form of the system state transition matrix as detailed in (1) to (5), q is the zero-mean process error
vector, z is the measurement vector, h is the nonlinear measurement function and r is the zero-mean measurement
error vector.

3.1 | Extended Kalman filter

The EKF algorithm can be divided into two steps of prediction and correction, as summarized in the following
equations:
Prediction:

xk− = f ðxk − 1 , k −1Þ ð10Þ

Pk− = Fk Pk − 1 FTk + Qk − 1 ð11Þ

Correction:

−1
Kk = Pk− HTk Hk Pk− HTk + Rk ð12Þ



xk+ = xk− + Kk zk −h xk− ð13Þ

Pk+ = ðI− Kk Hk ÞPk− ð14Þ

where Pk− is the covariance matrix of xk−, F is the Jacobian matrix of f with respect to x, H is the Jacobian of h with
respect to x, Kk is the Kalman gain, zk is the measurement vector, xk+ is the final estimate of the states after correction,
I is the identity matrix and Pk+ is the covariance matrix of xk+.
The estimate, xk, as given in (13) may be improved by calculating xk through iterations, linearizing about the most
recent estimate at each iteration. This variation of EKF is named Iterated EKF. Another way to reduce the estimation
error of EKF during linearization is to introduce higher-order terms of the Taylor expansion of the measurement func-
tion. The most widely used high-order EKF is the second-order Kalman filter (SKF). Because of the similarities among
them, EKF, IEKF and SKF are called the EKF family. Details of the process of IEKF and SKF are not given in this arti-
cle. The readers can find the formulations of the two FASEs in References 21 and 22, respectively.

3.2 | Unscented Kalman filter

The UKF is based on the Unscented Transform theory. Instead of linearizing the measurement function using the Tay-
lor expansion, UKF propagates a statistical distribution of the states through the nonlinear functions. This characteristic
allows UKF to estimate the states and the covariance matrix without the need to calculate the Jacobian matrices of the
nonlinear functions. Thus, UKF could increase both the accuracy and speed of the DSE, especially in the case the func-
tions are highly nonlinear. The process of UKF can be divided into five steps, including sigma point generation, state
JIN ET AL. 5 of 23

prediction, sigma point generation, measurement prediction and correction. The mathematical formulations of each of
the steps as given below.
Sigma point generation:
A set of 2n + 1 (n is the number of states) sigma points are generated. They can be aggregated into a single matrix
as given in (15).
  pffiffiffi pffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffi 
Xk − 1 = ½χ0   χ2n  = xk+− 1   xk+− 1 + 3 0 Pk+− 1 − Pk+− 1 ð15Þ

State prediction:

Xk− = f ðXk − 1 ,k −1Þ ð16Þ

^ k ωm
xk− = X ð17Þ

T
Pk− = Xk− W Xk− + Qk − 1 ð18Þ


W = ðI− ½ ωm   ωm Þdiag W c0   W c2n ðI − ½ ωm   ωm ÞT ð19Þ

where Xk− is the resulted matrix after propagating Xk through f, W is the weight matrix, the definitions of the weighing
vector, ωm, and W0c … W2nc are given below:

 m T

ωm = W m
0 W 1   W 2n
m
ð20Þ

0 = 1 −n=3
Wm ð21Þ

Wm
i = 1=6, i = 1, 2,…, 2n ð22Þ

W c0 = 3 −n=3 ð23Þ

W ci = 1=6, i = 1, 2, …, 2n ð24Þ

Measurement prediction:


Γk− = h Xk− , k ð25Þ

zk− = Γk− ωm ð26Þ

−

T
Pz,k = Γk− W Γk− + Rk ð27Þ

−

T
Pxz,k = Xk− W Γk− ð28Þ

where zk− is the predicted measurement vector, Pz,k− −

is the covariance of the predicted measurements and Pxz,k is the
covariance between the predicted states and the predicted measurements.
Correction:

−1
Kk = Pk− HTk Hk Pk− HTk + Rk ð29Þ



xk+ = xk− + Kk zk −h xk− ð30Þ
6 of 23 JIN ET AL.

Pk+ = ðI− Kk Hk ÞPk− ð31Þ

3.3 | Cubature Kalman filter

The CKF is a highly efficient and accurate nonlinear filter. It is named after the spherical-radial cubature rule,23 the
key technique that makes it possible to calculate multivariate moment integrals within the Bayesian filtering paradigm.
Under the assumption that both the process noises and measurement noises are Gaussian noises, the process of CKF
can be summarized into 3 steps as follows:
State prediction:
ð


xk− = f ðxk − 1 , k −1ÞN xk − 1 ;xk+− 1 ,Pk+− 1 dxk − 1 ð32Þ
ℝ n

ð
Pk− = f ðxk − 1 , k −1Þf T ðxk − 1 , k −1Þ
ℝn ð33Þ


T
N xk − 1 ; xk+− 1 , Pk+− 1 dxk − 1 −xk− xk−

where N(x;μ,P) is the symbol for Gaussian density for variable x with the mean value of μ and covariance of P.
Measurement prediction:
ð


zk− = hðxk , k ÞN xk ;xk− , Pk− dxk ð34Þ
ℝ
n

ð
−


T
Pz,k = hðxk , k ÞhT ðxk , k ÞN xk ;xk− , Pk− dxk −zk− zk− ð35Þ
ℝ
n

ð
−


T
Pxz,k = xk hT ðxk ,k ÞN xk ; xk− , Pk− dxk −xk− zk− ð36Þ
ℝn

Correction:

−1
Kk = Pk− HTk Hk Pk− HTk + Rk ð37Þ



xk+ = xk− + Kk zk −h xk− ð38Þ

Pk+ = ðI− Kk Hk ÞPk− ð39Þ

It is usually very difficult to directly calculate the multivariate integrals as given in (32) to (36). However, they can
be approximated using the spherical-radial cubature rule as explained in the following equations:

ð X
2n  pffiffiffiffi 
I N ðf Þ = f ðxÞN ðx; μ,PÞdx ’ ω i f μ + P ξi ð40Þ
ℝn i=1

pffiffiffi
− nei for i = 1, …, n
ξi = pffiffiffi ð41Þ
nei − n for i = n + 1, …, 2n
JIN ET AL. 7 of 23

1
ωi = , i = 1,…,2n ð42Þ
2n

where ei is the ith column of the identity matrix, ξi, i = 1, …, 2n, are the cubature points, ωi, i = 1, …,2n, are the weights
of the cubature points.
The cubature points generated in CKF are very similar to the sigma points generated in UKF. However, they are dif-
ferent in terms of both values and weights. For example, consider a four-state vector with 0 mean and unity covariance,
the expressions of the cubature points and the sigma points and their corresponding weights are given in (43) to (46). It
is shown clearly in (46) that one of the weights of the sigma points is below 0. According to (27), the covariance matrix,
−
Pz,k , may become nonpositive definite, which could lead to filter instability, that is, for a bounded input, the output of
the filter may not be bounded.23

ξ = ½ ξ1 ξ2 … ξ8  = ½ −I4 × 4 I4 × 4  ð43Þ

1 1 1
½ ω1 ω2 … ω8  = … ð44Þ
8 8 8

pffiffiffi
X = ½ χ0 χ1 χ2 … χ8  = 3½ 04 × 1 − I4 × 4 I4 × 4  ð45Þ

1 1 1 1
½W0 W1 W2 … W8  = − … ð46Þ
3 6 6 6

4 | ANOM ALY I DENTIFIC ATION

In this section, the proposed anomaly detection and discrimination method will be discussed in Section 4.1. The anom-
aly identification method will be discussed in Section 4.2.

4.1 | Anomaly detection and discrimination

4.1.1 | The conventional innovation analysis method

The proposed anomaly detection and discrimination method is developed based on the conventional innovation analy-
sis method. The implementation of the conventional innovation analysis method is given in Figure 1. In Figure 1, the
NI of any measurement can be calculated in (47), while the skewness of any NI is calculated in (50). The algorithm
involves the following two steps:

Step 1: Detect the anomalies by examining whether the largest NI is greater than 3.
Step 2: Distinguish bad data from sudden load change by examining whether the skewness of the largest NI is greater
than 3.

The conventional innovation analysis method has two drawbacks. First, it is unable to detect sudden topology
change after fault. Second, the threshold in the second step is unreliable—it will change in different networks and with
different measurement configurations. To overcome these two drawbacks, an improved innovation analysis method is
proposed, as will be discussed in the next sub-section.

 qffiffiffiffiffiffiffi
ffi
τk = zk − zk− = Pz,k
−
ð47Þ
8 of 23 JIN ET AL.

−
where the definition of the predicted measurement zk− , and its covariance, Pz,k , of UKF and CKF have already been
−
given in (26), (27), (34) and (35). The definitions of zk− and Pz,k of EKF, IEKF and SKF are given in (48) and (49),
respectively.


zk− = h xk− ð48Þ

−
Pz,k = Hk Pk− HTk + Rk ð49Þ

ψ k = M 3,k =σ 3k ð50Þ

where M3,k represents the third central moment, and σ k is the SD, whose values can be obtained as given below:
   
M 3,k = E ðτk Þ:3 −3vk E ðτk Þ:2 + 2v3k ð51Þ

 
σ 2k = E ðτk Þ:2 −v2k ð52Þ

vk = E½τk  ð53Þ

where the operator,. *n, denotes element-wise exponential calculation with the exponent n.

FIGURE 1 Diagram of the anomaly detection algorithm. Ck,

max is the maximum skewness in the normal case, which can be
determined offline
JIN ET AL. 9 of 23

4.1.2 | The improved innovation analysis method

The implementation block diagram of the proposed improved innovation analysis method is illustrated in Figure 2. The
NIs and the skewness can be calculated with the same equations as listed in (47) to (50). The algorithm involves five
steps, each of the steps is explained below:

Step 1: Detect whether or not an anomaly happens by counting the number of NIs that are greater than the pre-set
threshold of 3. Anomaly is detected if this number Sk is greater than 0, or it is a normal case.
Step 2: Detect the anomaly caused by single bad data in measurements by examining if Sk = 1. This is because except
for bad data, both sudden load change and sudden topology change due to fault will cause multiple NIs to be greater
than the threshold.
Step 3: Detect the sudden topology change after fault. A sudden topology change will cause consistent large errors to
the predictions. Thus, two consecutive largest NIs are examined. A larger threshold of 6 is selected. Static state estima-
tion with the measurements received at Instant k is performed. The NR is calculated. The impact of sudden load change
on the NR of SSE is very small. The NR is usually under the value of 6. Even a sudden cut-off of the heaviest load would
only make NR slightly over the value of 10. In contrast, the impact of sudden topology change on the NR of SSE is very

F I G U R E 2 Diagram of the
anomaly detection algorithm
10 of 23 JIN ET AL.

large. The NR is usually larger than 1000. Therefore, a threshold of 30 for NR is set to distinguish the sudden topology
change from the sudden load change.
Step 4: Detect the sudden load changes that evenly affect all of the predictions and thus the innovations. This type of
sudden load changes is characterized by very small skewness. The detection threshold is selected to be 3, which is the
same with that in the conventional innovation analysis method.
Step 5: Distinguish the anomaly caused by multiple bad data and the sudden load changes that have significant
impacts on only a small part of the predictions and thus the innovations. In this article, a large number of simu-
lations show there is a clear threshold to distinguish bad data and sudden load change, SIRth = min(1/3, 3σ max).
Here SIR is the skewness to the largest NI ratio, and σ max is the maximum SD of the measurement noise. The
simulations show that if bad data occur, SIR will be greater than SIRth. If a sudden load change happens, SIR will
be less than SIRth.

Remark 1. The threshold of 3 is used to distinguish the anomaly case from the normal cases. Based on the central limit
theorem, the NIs can be demonstrated to follow standard Gaussian distribution. According to the standard Gauss-
ian distribution, only less than 0.3% of the data's absolute value is greater than 3. Hence, we set 3 as a threshold to
detect anomalies.

Remark 2. The thresholds of 6 and 30 are for distinguishing the fault cases from the other anomaly cases. The thresh-
olds are set based on the unique characteristics of the faults that they will cause very large persistent errors to
both CKF and SSE. These values are chosen empirically based on the statistical distribution from the simulation
results.

4.2 | Identification of the anomalies

The identification of bad data in FASE is straightforward. The measurement that has the largest NI is identified
to be the bad data, since there is only one NI greater than the normal value. However, both sudden load change
and fault will cause multiple NIs to become greater than the normal value. Hence, the location where these
events happen may not be identified directly. Two methods are proposed here to identify the location of the two
kinds of anomalies.

4.2.1 | Identification of sudden load change bus

A pseudo injection method using CKF is proposed to identify the bus where the sudden load change happens, as
described in the steps below:

Step a: Obtain the average value of the load, select buses whose load is larger than certain percentage of the average
value.
Step b: Calculate the predicted active power injection measurements, zl−，and its covariance, Pl−, as given in (54).
ð


zl− = hl ðxk , k ÞN xk ;xk− , Pk− dxk ð54Þ
ℝn

where hl are the active power injection functions corresponding to the selected buses in Step a.
Step c: Calculate the corrected active power injection measurements, zl+, as given in (55):
ð


zl+ = hl ðxk , k ÞN xk ; xk+ , Pk+ dxk ð55Þ
ℝn

Step d: Calculate the covariance between zl− and zl+, Pl:

JIN ET AL. 11 of 23

ð


T
Pl = hl ðxk , k ÞhTl ðxk , k ÞN xk ;xk+ ,Pk+ dxk −zl+ zl+ ð56Þ
ℝn

Step e: Calculate the NI vector of the pseudo injection measurements, τl.

pffiffiffiffiffi
τl = zl− −zl+ = Pl ð57Þ

Step f: Find the maximum element of τl; the corresponding bus is where the sudden load change happens.

4.2.2 | Identification of the removed branch after fault

Step A: Select a unidirectional set of power flow measurements that covers all branches of the network, M.
Step B: After receiving the measurements at instant k, store the measurements  M into a vector, mk.
Step C: If any fault is detected, calculate the variation ratio, rm, as given in (58). The corresponding branch of the largest
element of rm is identified to be the removed branch.

rm = jlnðjmk − 1 =mk jÞj ð58Þ

5 | FORECASTING-AIDED STATE ESTIMATION USING CKF

Based on the cubature spherical rule and the new anomaly detection and identification algorithms presented in the pre-
vious section, a new forecasting-aided state estimation algorithm using CKF is proposed as follows. The corresponding
flow chart is shown in Figure 3.
Begin: Find fk and Qk according to (1) to (7) find hk and Rk according to (8) and (9), respectively. a0 = x−1, b0 = 0,
x0+ = x0− = x0, Pk0+ = Q0. Set k = 1.
While (k ≥ 1).
{

Step 1: Generate cubature points and the corresponding weights according to (41) and (42).
Step 2: Predict the states using (32) and (33).
Step 3: Predict the measurements using (34) to (36).
Step 4: Obtain the Kalman estimate using (37) to (39).
Step 5: Detect the anomaly using the algorithm in Figure 3.
Step 6: According to the type of anomaly detected, identify the bad data using the traditional largest NI method, identify
the sudden load change bus following Step a to Step e, and the removed bus after the fault, following Step A to Step C.
Step 7: If any bad data is identified, remove the bad data and go back to Step 5. Else, if any sudden load change is identi-
fied, reject the prediction and execute static state estimation. Else, if any branch is identified to be missing, then reject
the prediction, update the topology and execute static state estimation.
Step 8: Output xk+ and Pk+. Set k = k + 1.}

6 | SIMULATION SETTINGS

CKF based forecasting-aided state estimation algorithm proposed in Section 5 is tested and compared with the other
four nonlinear Kalman Filters discussed in Section 4 for IEEE 14 and 118 bus test systems. To guarantee normal opera-
tion of the anomaly detection functions, a redundant set of conventional measurements as well as a redundant set of
12 of 23 JIN ET AL.

F I G U R E 3 Flowchart of the forecasting-aided state estimation

using cubature Kalman filter algorithm

PMU measurements are selected for the simulations in both IEEE 14 and 118 bus test systems, whose network diagram
and data are available in Reference 44. The measurement placements are given in Tables 1 and 2, respectively. Here,
measurement redundancy means that the system can still perform state estimation if any of the measurement is
missing.
The assumption that the system operates in quasi steady state is simulated with a gradual linear increase of some
selected loads by 20% over 50 instants. The small fluctuations of the changes are simulated by adding a random walk
signal whose SD is 3% of the magnitude of the change. In response to the load increase, the power generations of the
generators increase according to their participation factors, that is, the percentage of power generation before the
change of load. For example, if the total load increase is 10 MW and the two generators (GA and GB) have participation
factors of 0.3 and 0.7, respectively, the generated power of GA and GB will be 3 MW and 7 MW, respectively.
Without loss of generality, the covariance of the process error, Q, is assumed to be 1 × 10−6 × I, where I is the iden-
tity matrix. The measurement errors are Gaussian errors whose SD is set to be a certain percentage of the measure-
ments according to the types of the measurements including supervisory control and data acquisition (SCADA) system
and phasor measurement units (PMUs):

• SCADA: P, Q measurements (0.02) and V (0.002)

• PMU: 0.002 for magnitudes, and 0.01 for angles
JIN ET AL. 13 of 23

T A B L E 1 Measurement placement Injection measurements at bus # 1;3;6;7;8;9;11;13;14

for IEEE 14 bus test system
Flow measurements at branch # 1;2;4;5;6;7;13;18;19;34
Voltage measurements at bus # 2;3;8;10;12
PMU placed at bus # 2;6;7;9

Abbreviation: PMU, phasor measurement unit.

TABLE 2 Measurement placement for IEEE 118 bus test system

Injection measurements at bus # 1, 4, 7,10, …, 118

Flow measurements at branch # 1, 3, 5, 7, …, 185
Voltage measurements at bus # 1, 6, 11, 16, …, 116
PMU placed at bus # 1;5;10;12;13;17;21;25;28;34;37;41;45;49;53;56;62;64;72;73;75;77;80;85;87;91;94;101;105;110;114;116

Abbreviation: PMU, phasor measurement unit.

For example, if an active power injection measurement with the value of 0.5 p.u., the SD of its error is set to be
0.5 × 0.02 = 0.01 p.u.
The initial condition assumes the knowledge of the previous two states before the first state is to be estimated
(x−1 and x0), so that the Holt's method can work for the first prediction. The performance of the FASE is assessed using
the overall performance index (OPI) defined as follows:
P +
z − zk,true
Jk = P k ð59Þ
jzk − zk,true j

An intuitional explanation of the OPI is the rate of reduction of the measurement errors after state estimation. Thus,
a small OPI indicates good performance of a FASE.
Four cases are considered for the simulation, including the normal operation case, presence of bad data, presence of
sudden load change, and presence of fault. The simulation settings are the same for all cases except for the introduction
of the anomaly at instant 25 for the last three cases.

7 | S I MULATION RESULTS A ND ANALYSIS

In this section, the results and analysis of the simulations in IEEE 14 and 118 bus test systems in four cases, including
normal operation, bad data, sudden load change and fault, are presented. The results in the four cases are discussed in
Sections 7.1 to 7.4, respectively. In each sub-section, the performances of the different nonlinear filters will be discussed
first; then the results of the anomaly detection methods will be discussed. In Sections 7.3 and 7.4, the results of anomaly
identification are also discussed in the final part of the sub-section. Because of the excessively long execution time in
IEEE 118 bus test system, simulations of SKF are only performed in IEEE 14 bus test system. All results presented in
this section are obtained after 100 Monte Carlo simulations.

7.1 | Normal operation case

In this sub-section, the simulation results in the normal case in IEEE 118 bus test system are presented. The sim-
ulation in IEEE 14 bus test system shows almost the same results as that in IEEE 118 bus test system. The OPI,
skewness and NI of SKF are similar to that of EKF and IEKF. Because of the similarity, these results are not
shown in this article.
14 of 23 JIN ET AL.

7.1.1 | Performance comparisons of the nonlinear filters

The OPIs of the four FASEs introduced in Section 6 over the 50 instants are illustrated in Figure 4. All of FASEs have
shown smaller OPIs than SSE, demonstrating the effectiveness of the Holt's method and the high estimation accuracy
of FASEs over SSE when the dynamics of the system are slow. It is also shown clearly that CKF has the lowest OPI,
with UKF being about 10% larger, EKF and IEKF being about 25% larger, and SSE being about 30% larger. Such
improvements in estimation accuracy by using CKF are significant, considering that the same measurements and the
same prediction process are used for the other FASEs.
In Figures 5 and 6, it is shown that in normal operation case the average skewness and the average NIs are well
within the thresholds. Therefore, it is demonstrated that the skewness threshold of 3 and the NI threshold of 3 in the
proposed method are correct for the normal operation case.

7.1.2 | Anomaly detection

Table 3 shows the test results of the three types of anomaly detection method, including the method proposed in this
article, the conventional NI method30-32 and the robust FASE.38
The test results show that the proposed method correctly detect 85 cases as normal operation cases; it also detects
14 cases as bad data cases and 1 case as sudden load change. Since Gaussian errors of the measurements might occa-
sionally cross the threshold, the subsequent identification and elimination of these measurements would enhance the
estimation accuracy. In contrast, the detections as sudden load change are mal-detections, which would lead to elimina-
tion or down-weighting the predictions, and eventually result in a decrease in estimation accuracy. Hence, the proposed
method has a 99% rate of good detections, while the conventional NI method's good detection rate is only 85%. As for
the robust FASEs, its threshold for bad data is very small such that it can down-weight the measurements with larger
errors. Thus, the robust FASE can work very well in the normal operation case with 100% good detections.

F I G U R E 4 Average overall performance indexes (OPIs) of

forecasting-aided state estimators (FASEs) and static state estimator
(SSE) in normal operation case in IEEE 118 bus test system

F I G U R E 5 Average skewness of forecasting-aided state

estimators (FASEs) in normal operation case in IEEE 118 bus test
system

F I G U R E 6 Average normalized innovation of the forecasting-

aided state estimators (FASEs) in normal operation case in IEEE
118 bus test system
JIN ET AL. 15 of 23

TABLE 3 Results of the anomaly detections methods

Method The proposed method Conventional NI method Robust FASE

Num of detections 85 85 0
(Normal operation)
Num of detections 1 15 0
(Sudden load change)
Num of detections 14 0 100
(Bad data)
Num of detections 0 0 0
(Fault)

Abbreviations: FASE, forecasting-aided state estimator; NI, normalized innovation.

F I G U R E 7 Average overall performance indexes (OPIs) of

forecasting-aided state estimators (FASEs) and static state estimator
(SSE) in bad data case in IEEE 118 bus test system

7.2 | Presence of bad data

In this sub-section, the presence of the first type of anomaly, the bad data, is considered in the simulation study. A gross
bad data (the measurement is set to 0) is introduced in Measurement 83, the measurement of the active power flow
between Bus 4 and Bus 5, at instant 26.

7.2.1 | Performance comparisons of the nonlinear filters

In Figure 7, a significant drop can be seen for both FASEs and SSE. It is because the size of the overall error after any
type of estimation is smaller than the magnitude of the gross error itself. However, OPI quantifies the improvement of
data quality rather than the absolute values of the errors. Therefore, the bad data needs to be detected and identified
since it does significantly smear the state estimate.
To illustrate the effect of bad data on NI and skewness, the average skewness of different instants and the average
NIs at instant 26 are shown in Figures 8 and 9, respectively. The average NI of Measurement 83 is much greater than
the threshold of 3. The average skewness at instant 26 is about 23, which is clearly greater than 1/3 of the max jNIj
(about 45).

7.2.2 | Anomaly detection

The results demonstrate that the NI threshold of 3, skewness threshold of 3 and SIR threshold of 1/3 in the proposed
method are correct. The test results in Table 4 show that all three methods compared in this article can detect the bad
data with 100% accuracy. Since its corresponding NI is the only one that is greater than the threshold of 3, the bad data
is identified to be Measurement 83.
The simulation results in IEEE14 bus test are also almost the same to that in IEEE 118 bus test system. SKF also
has similar OPI, skewness and NI compared to EKF and IEKF. Thus, these results are not presented here.
16 of 23 JIN ET AL.

F I G U R E 8 Average skewness of forecasting-aided state

estimators (FASEs) in bad data case in IEEE 118 bus test system

F I G U R E 9 Average normalized innovation of forecasting-

aided state estimators (FASEs) in bad data case in IEEE 118 bus test
system

TABLE 4 Results of the anomaly detections methods

Method The proposed method Conventional NI method Robust FASE

Num of detections 0 0 0
(Normal operation)
Num of detections 0 0 0
(Sudden load change)
Num of detections 100 100 100
(Bad data)
Num of detections 0 0 0
(Fault)

Abbreviations: FASE, forecasting-aided state estimator; NI, normalized innovation.

7.3 | Presence of sudden load change

7.3.1 | IEEE 14 bus test system

The third simulation case involves the presence of sudden load change: a 40% load curtailment is imposed at the load at
Bus 2 at instant 26.

Performance comparisons of the nonlinear filters

Unlike the almost equal effect of the bad data, the sudden load change affects FASEs with different nonlinear filters dif-
ferently: As can be seen in Figure 10, CKF is the most affected filter with its peak at about 0.7, UKF is also significantly
affected but to a smaller extent. SSE is not affected since it does not use the predictions. EKF family is slightly affected.
Its peak OPI value is only slightly larger than that of SSE. In addition, the OPI of CKF will rise again after instant
26 and reaches a second pea at instant 29 even if no more anomaly happens after instant 26. Hence, it is very important
to accurately detect and identify the sudden load change and take appropriate actions to guarantee that the FASE can
make accurate estimates with the high-performance nonlinear filters, especially CKF.
To illustrate the effect of sudden load change in a small network on NI and skewness, the average skewness of dif-
ferent instants and the average NIs at instant 26 are shown in Figures 11 and 12, respectively. The sudden load change
in a small network like IEEE 14 bus test system might affect every state to more or less the same extent. Therefore, the
average skewness of FASEs increases slightly at instant 26, which is still within the normal value. In contrast, the aver-
age NIs of several measurements have exceeded the threshold of 3. These results correspond with the detection criteria
given in the proposed method.
JIN ET AL. 17 of 23

Anomaly detection
The results of the anomaly detection methods after 100 Monte Carlo simulations are given in Table 5. It shows that
both the proposed method and the conventional NI method successfully detected the sudden load changes in all
100 simulations; while the robust FASE failed in all simulations. This is because in these cases the magnitude of the NI
of the measurements are much larger than the errors in the predicted states. Therefore, only the measurements are
detected as outliers, indicating it is a bad data case.

Anomaly identification
Although the five measurements whose absolute values are greater than 3, V4, Ireal,1-2, Ireal,1-5, Iimag,1-2 and Iimag,2-4, are
near Bus 2, it is not obvious to directly identify the location where the sudden load change happens. Following the steps
given in Section 4.2.1, the NI vector of the pseudo injection measurements, τl, for CKF is calculated as included in
Table 6. The data shows clearly that Load 1 at Bus 2 has the largest τl. Thus, the sudden load change is correctly identi-
fied to be located at Bus 2.

7.3.2 | IEEE 118 bus test system

Another simulation is conducted in a large network of IEEE 118 bus test system, where a 40% load curtailment is also
imposed at Bus 2 at instant 26.

Performance comparisons of the nonlinear filters

As shown in Figure 13, the effect of sudden load change on the estimation accuracy, in this case, is similar to that in
IEEE 14 bus test system with CKF being the most affected and EKF family being least affected. In addition, the estima-
tion accuracy of CKF is worse than that before the sudden load change until about 10 instants after the sudden load
change, which is also similar to the previous case. What is notable is that the OPI level of SSE stays unchanged. This is
because SSE does not utilize the predictions for producing estimation.

F I G U R E 1 0 Average overall performance indexes (OPIs) of

forecasting-aided state estimators (FASEs) and static state estimator
(SSE) in sudden load change case in IEEE 14 bus test system

F I G U R E 1 1 Average skewness of forecasting-aided state

estimators (FASEs) in sudden load change case in IEEE 14 bus test
system

F I G U R E 1 2 Average normalized innovation (NI) of

forecasting-aided state estimators (FASEs) in sudden load change
case in IEEE 14 bus test system
18 of 23 JIN ET AL.

TABLE 5 Results of the anomaly detections methods

Method The proposed method Conventional NI method Robust FASE

Num of detections 0 0 0
(Normal operation)
Num of detections 100 100 0
(Sudden load change)
Num of detections 0 0 100
(Bad data)
Num of detections 0 0 0
(Fault)

Abbreviations: FASE, forecasting-aided state estimator; NI, normalized innovation.

T A B L E 6 Normalized innovations
Load Bus τl
of the pseudo injection measurements
1 2 1.25
2 3 0.0579
3 4 0.0256
4 5 0.1026
5 6 0.1595
6 9 0.2753
7 10 0.7035
8 11 0.1135
9 12 0.0748
10 13 0.1226
11 14 0.1997

F I G U R E 1 3 Average overall performance indexes (OPIs) of

forecasting-aided state estimators (FASEs) and static state estimator
(SSE) in sudden load change case in IEEE 118 bus test system

To have an overview of the effect of sudden load change in a larger network of IEEE 118 bus test system, the average
skewness of different instants and the NIs at instant 26 are shown in Figures 14 and 15, respectively. The skewness is clearly
over the threshold of 3 at instant 26 when the sudden load change happens. This is very similar to the skewness in the bad
data case. But the peak, at the value of about 5, is smaller than 1/3 of the max jNIj, which is about 30. Finally, the NIs of many
measurements are greater than the threshold of 3, which is clearly different from the bad data case where only a few data
might be corrupted. The similarity of skewness to the bad data case can be explained by the fact that in a larger network, the
sudden load change at any bus can have significant effect on smaller number of states, and hence predicted measurements.
Hence, the effect of sudden load change would be similar to the effect of a group of neighboring bad data with smaller errors.

Anomaly detection
The detection results show that the proposed method can successfully detect all 100 sudden load changes, while the conven-
tional NI method wrongly detects all of them as bad data since the skewness is greater than the threshold of 3. What is nota-
ble is that the robust FASE has labeled both the related measurements and related predictions as outliers. Thus, the
subsequent down-weighting of both the measurements and predictions might not improve the estimation accuracy. Therefore,
in this case, only the proposed method has perfect anomaly detection performance, while both the two existing methods are
JIN ET AL. 19 of 23

unable to detect sudden load changes. The subsequent identification of the sudden load change can be done following the
same procedures as described in Section 4.2.1. Note that the two considered sudden load change cases, it is assumed that the
events can significantly increase the magnitude of the NIs of the measurements. It is true that the impact of the sudden load
change is small, the largest NI might be slightly over 6, while the skewness might be reduced to some value smaller than 3. In
these cases, the threshold given in the article might not be correct (Table 7).

7.4 | Presence of fault

The final case to be studied in this article is in the presence of fault. The transient characteristics of the network in
response to the fault are not studied, as the state estimators update its estimate (about a few minutes interval) much
slower than the transients. Rather, it is assumed that the topology of the network is changed after clearing the fault by
removing one of the branches. In this case, in IEEE 14 bus test system, the topology change happens at instant 26, when
Branch 6 that connects Bus 3 and Bus 4 is removed. If the topology change is not updated in time, the state estimators
will not be able to produce the correct estimate of the states. Because of the results' similarity, the results in IEEE
118 bus test system are not given in this article.

7.4.1 | Performance comparisons of the nonlinear filters

As shown in Figure 16, the OPIs of the state estimators are much greater than 1 after the fault clearance, indicating that the
estimations are much worse than the measurements. Nevertheless, the persistent error caused by topology error results in

F I G U R E 1 4 Average skewness of forecasting-aided state

estimators (FASEs) in sudden load change case in IEEE 118 bus test
system

F I G U R E 1 5 Average normalized innovation of forecasting-

aided state estimators (FASEs) in sudden load change case in IEEE
118 bus test system

TABLE 7 Results of the anomaly detections methods

Method The proposed method Conventional NI method Robust FASE

Num of detections 0 0 0
(Normal operation)
Num of detections 100 0 100
(Sudden load change)
Num of detections 0 100 100
(Bad data)
Num of detections 0 0 0
(Fault)

Abbreviations: FASE, forecasting-aided state estimator; NI, normalized innovation.

20 of 23 JIN ET AL.

F I G U R E 1 6 Average overall performance indexes (OPIs) of

forecasting-aided state estimators (FASEs) and static state estimator
(SSE) in fault case in IEEE 14 bus test system

F I G U R E 1 7 Average skewness of forecasting-aided state

estimators (FASEs) in fault case in IEEE 14 bus test system

F I G U R E 1 8 Average normalized innovation of forecasting-

aided state estimators (FASEs) in fault case in IEEE 14 bus test
system at time 26

F I G U R E 1 9 Average normalized innovation of forecasting-

aided state estimators (FASEs) in fault case in IEEE 14 bus test
system at time 27

persistent disturbance to the NIs of the measurements and its skewness, which can be used to detect the topology change, as
shown in Figures 17 to 19. Figure 17 shows that the skewness of CKF decreases sharply from the normal value of around 1 to
about 0.1 at instant 26, then return back to 1 at instant 27, and then further increase to over 4 after instant 28. As shown in
Figure 18, at instant 26, the NIs of several measurements are much greater than the threshold of 3. At instant 27, as illustrated
in Figure 19, there are still a number of NIs that are greater than the threshold of 3.

7.4.2 | Anomaly detection

Applying the proposed method and the two existing methods, the anomalies detected are summarized in Table 8. It
shows clearly that the proposed method can correctly identify all 100 topology changes after fault, while the other two
existing methods are unable to detect the fault.

7.4.3 | Anomaly identification

To identify the removed branch, the variation ratios of the branches at instant 25 are calculated as listed in Table 9.
Branch 6 is successfully identified to be the removed branch as its rm is the largest.
JIN ET AL. 21 of 23

TABLE 8 Results of the anomaly detections methods

Method The proposed method Conventional NI method Robust FASE

Num of detections 0 0 0
(Normal operation)
Num of detections 0 100 100
(Sudden load change)
Num of detections 0 0 100
(Bad data)
Num of detections 100 0 0
(Fault)

Abbreviations: FASE, forecasting-aided state estimator; NI, normalized innovation.

T A B L E 9 Variation ratios of the

Branch # rm
branches after the fault in IEEE 14 bus
test system 1 0.0116
2 0.0124
3 0.0121
4 0.0120
5 0.0068
6 1.4350
7 0.0063
8 0.0435
9 0.0323
10 0.0002
11 0.0345
12 0.0561
13 0.0169
14 0.7173
15 0.0374
16 0.0129
17 0.0208
18 0.0088
19 0.0105
20 0.0305

8 | C ON C L U S I ON S

The article proposed a novel algorithm for CKF-based FASE. It solved the problem of the existing CKF-based FASE
algorithm in the sudden load change cases. The proposed algorithm includes a set of actions for the FASE to follow for
each type of anomaly and a new anomaly detection and identification algorithm that is much more accurate than the
existing anomaly detection algorithms.
The simulation results obtained show that in normal operation cases, as well as bad data cases, the existing CKF
algorithm has the highest estimation accuracy compared to the other four FASEs. However, in cases described by sud-
den load changes and faults, the grade is totally opposite, the existing CKF is the most inaccurate.
The simulations also show that the proposed anomaly detection algorithm can correctly detect the considered
anomalies, by using a new detection threshold, skewness to the largest NI ratio (SIR). However, existing innovation
analysis methods cannot distinguish bad data and sudden load change, which might lead to wrong actions for the FASE
to follow. The robust FASE methods supress the weights of both predictions and measurements in the cases of sudden
22 of 23 JIN ET AL.

load change and sudden topology change due to fault. Since the correct measurements could also be down-weighted,
the estimation accuracy improvement by the robust FASE are limited in these cases. In addition, the proposed algo-
rithm is shown to be capable of detecting sudden topology changes after open circuit faults by using the largest NR of
the static state estimator. Two new methods to identify the location of sudden load change, and the location of the
removed branch after fault clearance also show satisfactory accuracies.
The limitation of the proposed algorithm is that when the magnitude of anomaly is small, there might not be clear
threshold for distinguishing sudden load change and bad data. Further research will focus on improving the proposed
algorithm or finding new methods to solve this problem.

P EE R R EV IE W
The peer review history for this article is available at https://publons.com/publon/10.1002/2050-7038.12714.

DATA AVAILABILITY STATEMENT

Data openly available in a public repository that issues datasets with DOIs.

ORCID
Zhaoyang Jin https://orcid.org/0000-0003-1876-2153

R EF E RE N C E S
1. Schweppe FC, Wildes J. Power system static-state estimation, part I: exact model. IEEE Trans Power App Syst. 1970;PAS-89(1):120-125.
2. Schweppe FC, Rom DB. Power system static-state estimation, part II: approximate model. IEEE Trans Power App Syst. 1970;PAS-89(1):
125-130.
3. Schweppe FC. Power system static-state estimation, part III: implementation. IEEE Trans Power App Syst. 1970;PAS-89(1):130-135.
4. Terzija V, Valverde G, Cai D, et al. Wide-area monitoring, protection, and control of future electric power networks. Proc IEEE. 2011;99:
80-93.
5. A. S. Debs and R. E. Larson, “A dynamic estimator for tracking the state of a power system,” IEEE Trans Power App Syst, vol. PAS-89,
pp. 1670–1678, Sep./Oct. 1970.
6. K. Srinivasan and Y. Robichaud, “A dynamic estimator for complex bus voltage determination,” IEEE Trans Power App Syst, Sep./Oct.
1974, PAS-93, 1581, 1588.
7. Allam MF, Laughton MA. Power-system dynamic-state estimation by optimal variable incremental displacements. Proc Inst Elect Eng C.
1976;123(5):433-436.
8. R. D. Masiello and F. C. Schweppe, “A tracking static state estimator,” IEEE Trans Power App Syst, vol. PAS-90, pp.1025–1033,Mar./Apr.
1971.
9. Handschin E. Real-time data processing using state estimation in electric power systems. Real Time Control of Electric Power Systems.
Amsterdam, The Netherlands: Elsevier; 1972:29-57.
10. Falc~ao DM, Cooke PA, Brameller A. Power system tracking state estimation and bad data processing. IEEE Trans Power App Syst. 1982;
PAS-101:325-333.
11. Leite da Silva AM, Do Coutto Filho MB, de Queiroz JF. State forecasting in electric power systems. IEE Proc Gener Transm Distrib. 1983;
130(5):237-244.
12. Alves da Silva AP, Quintana VH, Pang GKH. Solving data acquisition and processing problems in power systems using a pattern analysis
approach. Proc Inst Elect Eng C. 1991;138(4):365-376.
13. Alves da Silva AP, Quintana VH. Pattern analysis in power system state estimation. Elect Power Energy Syst. 1995;17(1):51-60.
14. Singh AK, Pal BC. Decentralized dynamic state estimation in power systems using unscented transformation. IEEE Transactions on
Power Systems. 2014;29(2):794–804. https://doi.org/10.1109/TPWRS.2013.2281323.
15. Ariff MAM, Pal BC, Singh AK. Estimating dynamic model parameters for adaptive protection and control in power system. IEEE Trans-
actions on Power Systems. 2015;30(2):829-839. https://doi.org/10.1109/TPWRS.2014.2331317.
16. Zhao J, Netto M, Mili L. A robust iterated extended Kalman filter for power system dynamic state estimation. IEEE Trans Power Syst.
2017;32(4):3205-3216.
17. Zhao J, Mili L. A robust generalized-maximum likelihood unscented Kalman filter for power system dynamic state estimation. IEEE J
Sel Topics Signal Process. 2018;12(4):578-592.
18. Zhao J, Qi J, Huang Z, et al. Power system dynamic state estimation: motivations, definitions, methodologies, and future work. IEEE
Trans Power Syst. 2019;34(4):3188-3198.
19. Mandal JK, Sinha AK, Roy L. Incorporating nonlinearities of measurement function in power system dynamic state estimation. IEE Proc
Gener Transm Distrib. 1995;142(3):289-296.
20. Valverde G, Terzija V. Unscented Kalman filter for power system dynamic state estimation. IET Gener Transm Distrib. 2011;5(1):29-37.
21. Bretas NG. An iterative dynamic state estimation and bad data processing. Int J Electr Power Energy Syst. 1989;11(1):70-74.ISSN 0142-
0615.
JIN ET AL. 23 of 23

22. Gelb A. Applied Optimal Estimation. Cambridge, MA: MIT Press; 1974.
23. Arasaratnam I, Haykin S. Cubature Kalman filters. IEEE Trans Automat Contr. 2009;54(6):1254-1269.
24. Sharma A, Srivastava SC, Chakrabarti S. A cubature Kalman filter based power system dynamic state estimator. IEEE Trans Instrum
Measure. 2017;66(8):2036-2045.
25. Akhlaghi S, Zhou N, Huang Z. Hybrid approach for estimating dynamic states of synchronous generators. IET Gener Transm Distrib.
2019;13(5):669-678.
26. Zhou N, Meng D, Huang Z, Welch G. Dynamic state estimation of a synchronous machine using PMU data: a comparative study. IEEE
Trans Smart Grid. 2015;6(1):450-460.
27. Sinha AK, Mondal JK. Dynamic state estimator using ANN based bus load prediction. IEEE Trans Power Syst. 1999;14(4):1219-1225.
28. Huang SJ, Shih KR. Dynamic-state-estimation scheme including nonlinear measurement-function considerations. IEE Proc Generat
Transm Distrib. 2002;149(6):673-678.
29. Shih KR, Huang SJ. Application of a robust algorithm for dynamic state estimation of a power system. IEEE Trans Power Syst. 2002;17
(1):141-147.
30. Nishiya K, Takagi H, Hasegawa J, Koike T. Dynamic state estimation or electric power systems - introduction of a trend factor and detec-
tion of innovation processes. Elect Eng Jpn. 1976;96(5):79-87.
31. Nishiya K, Takagi H, Hasegawa J, Koike T. Dynamic state estimation including detection of innovation process for electric power sys-
tems. Elect Eng Jpn. 1978;98(1):52-61.
32. Nishiya K, Hasegawa J, Koike T. Dynamic state estimation including anomaly detection and identification for power systems. Proc Inst
Elect Eng C. 1982;129(5):192-198.
33. Do Coutto Filho MB, Leite da Silva AM, Cantera JMC, da Silva RA. Information debugging for real-time power systems monitoring. Proc
Inst Elect Eng C. 1989;136(3):145-152.
34. da Silva AML, Filho MBDC, Cantera JMC. An efficient dynamic state estimation algorithm including bad data processing. IEEE Power
Eng Rev. 1987;PER-7(11):49-49.
35. Do Coutto Filho MB, Stacchini de Souza JC, Mattos RSG, Schilling MT. Revealing gross errors in critical measurements and sets via
forecasting-aided state estimators. Int J Elect Power Syst Res. 2001;57(1):25-32.
36. Jayakumar Geetha S, Chakrabarti S, Rajawat K. Hierarchical parallel dynamic estimator of states for interconnected power system. IET
Gener Transm Distrib. 2018;12(10):2299-2306.
37. Geetha SJ, Chakrabarti S, Rajawat K, Terzija V. An asynchronous decentralized forecasting-aided state estimator for power systems.
IEEE Trans Power Syst. 2019;34(4):3059-3068.
38. Zhao J, Zhang G, Dong ZY, La Scala M. Robust forecasting aided power system state estimation considering state correlations. IEEE
Trans Smart Grid. 2018;9(4):2658-2666.
39. Wang Y, Sun Y, Dinavahi V. Robust forecasting-aided state estimation for power system against uncertainties. IEEE Trans Power Syst.
2020;35(1):691-702. https://doi.org/10.1109/TPWRS.2019.2936141.
40. Ma W, Qiu J, Liu X, Xiao G, Duan J, Chen B. Unscented Kalman filter with generalized correntropy loss for robust power system
forecasting-aided state estimation. IEEE Trans Indus Inform. 2019;15(11):6091-6100. https://doi.org/10.1109/TII.2019.2917940.
41. Dang L, Chen B, Wang S, Ma W, Ren P. Robust power system state estimation with minimum error entropy unscented Kalman filter.
IEEE Trans Instrum Measure. https://doi.org/10.1109/TIM.2020.2999757. 2020;69 8797–8808.
42. Ji X, Yin Z, Zhang Y, et al. Real-time robust forecasting-aided state estimation of power system based on data-driven models. Int J Electr
Power Energy Syst. 2020;125:106412, ISSN 0142-0615. https://doi.org/10.1016/j.ijepes.2020.106412.
43. Majumdar A, Pal BC. Bad data detection in the context of leverage point attacks in modern power networks. IEEE Trans Smart Grid.
2016;9(3):2042-2054.
44. Zhao J, Zhang G, Dong ZY, Wong KP. Forecasting-aided imperfect false data injection attacks against power system nonlinear state esti-
mation. IEEE Trans Smart Grid. 2016;7(1):6-8. https://doi.org/10.1109/TSG.2015.2490603.

How to cite this article: Jin Z, Chakrabarti S, Yu J, Ding L, Terzija V. An improved algorithm for cubature
Kalman filter based forecasting-aided state estimation and anomaly detection. Int Trans Electr Energ Syst. 2021;
31:e12714. https://doi.org/10.1002/2050-7038.12714