Nomenclature

 bus admittance matrix of the AC power system

Ii and Vi net current and voltage at the ith bus

 voltage vectors for PMU and non-PMU buses,

 bus admittance sub-matrices for various PMU and

non-PMU buses

 net current vector.

x ,y real vectors

 weights associated with the bad data detection

 ,affine relationship between the mean-variance.

I Imaginary Value

R Real value

 output of convolutional layer

 optional parameters for convolutional layer

 used for numerical stability

W represents the layer weights

 represents the norm of weights for the output channel

 denotes the convolution

Power system security is greatly improved by accurately

analyzing the current operating status in real-time. A power grid

is a network of interconnected power lines that transmit energy

from generators to users. Power networks may cover whole

nations or regions and come in a range of sizes. Because of the

growing need for energy, power distribution networks must be

transformed into networks that combine diverse kinds of energy

generation as well as distribution. Loads and generators

connected to the grid both have an impact on power

performance. Due to their speed, synchronism, and precision,

Power Data Quality Improvement Through PMU Bad Data Detection

Based on Deep Complex Network

1PREETI KABRA, 2D SUDHA RANI

1K L University, Vijaywada, Andhra Pradesh, INDIA

2Sri vasavi college of Engineering, Tadepalligundem, Andhra Pradesh, INDIA

Abstract- Phasor Measurement Units (PMUs) enable the switching of devices in various power signal modes.

A jitter or glitch in a signal cause bad data and also the PMU data will spike due to a disturbance or a transmitting

data mistake. As a result of these difficulties, PMU data suffer from different degrees of data quality problems.

To detect the bad data, several approaches have been already utilized however it provides some disadvantages

such as complexity due to the utilization of dual identical systems separately for analyzing both real and

imaginary values of PMU. Likewise, the bad data due to the topology variations have not been optimally

identified. To overcome these issues a Robust Bad Data Detection Technique has been proposed in which a

Deep complex neural network (DCNN) is incorporated to process the complex number having both voltage

magnitude and phase angle. Deep complex Networks are also proposed with the conjunction of topology

processor and AC state estimator (SE). Moreover, instead of Batch normalization weight normalization is

altered due to the fusion of recurrent timestamps for measuring voltage magnitude and phase angle. The

comparative analysis is done in terms of accuracy , Bad data detection capability , bad data detection range and

running time with existing techniques The proposed technique provides accuracy of about 99.5% which is

higher than the existing techniques.

Keywords- Phasor Measurement Units (PMUs), Bad Data, Deep Complex Neural Network (DCNN), State

Estimator(SE), Weight Normalization, Batch Normalization.

Received: August 28, 2021. Revised: April 21, 2022. Accepted: May 21, 2022. Published: June 25, 2022.

1. Introduction

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

phasor measurement units (PMUs) have become a fundamental

mechanism utilized in pervasive electronics to accomplish state

perception [1]. A phasor measuring unit (PMU) is a tool utilized

within a power grid that determines both the magnitude and

phase angle of such an electrical phasor element (like voltage

or current) utilizing a shared time source for timing. PMUs may

also supply real-time phasor time data for important power

system applications including corrective action schemes,

oscillation detection, and condition estimates [2-6].

Numerous academics have examined the false data

issue in PMU assessments out from the viewpoints of cyber

assault and signal processing flaws, and have proposed several

solutions. However, it is also stated that in spite of repeated

information assaults, both computation cost or temporal

difficulty of a filter would rise, rendering it inappropriate for

real-time use. PMU data is sensitive to a variety of variables

due to the complexity of the components [7]. A jitter in a global

positioning system (GPS) signal, for example, might produce

phase angle variation. It's also conceivable that PMU data will

surge as a result of interference or a data transmission error.

PMU data suffer from varying degrees of data quality concerns

as a result of these issues. According to the California

Independent System Operator's (ISO) Five-Year Plan, around

10% to 17% of PMU data in US has issues [8]. PMUs are

utilized in Energy Management Systems to improve the state

estimation functionality (EMS). The simulation model starts by

collecting completely viewable assessment data from the

system, i.e. necessary qualities of voltages and currents in the

system are detected and given to the State Estimator to compute

the fundamental values of the overall network. PMU

measurements are commonly distorted as a result of (a)

deliberate data corruption by a cyber-assault or (b) inadvertent

data corruption over the digital data processing, storage, and

retrieval stage [9–11]. Data quality concerns make the system

less visible, impair the effectiveness of state estimates and

parameter identification using PMUs, and potentially

jeopardize power system safety and stability. PMU poor data

detection has become a significant issue, and it is vital for

increasing data quality and guaranteeing correct state

perception.

Several approaches for detecting bad data in power systems

have been presented. The singularity of the impedance matrix

and the sparsity of the error vector are exploited in [12] to

propose a new technique for detecting measurement errors in

DC power flow. It makes use of the power system's structure to

correctly compute measurement errors. A poor data detection

technique based on state estimate is given in [13]. By

identifying angle biases and current scaling errors, the phasor-

measurement-based state estimator enhances data consistency.

[14] introduces a time-series prediction model with a Kalman

filter and smoothing method for cleaning poor data. [15]

provides a technique for detecting incorrect data in real-time

that uses an unscented Kalman filter in combination with a state

estimation algorithm. Various academics had concentrated on

developing reduced and highly secure connections for such

purposes. Nevertheless, because of constraints in legacy

technology used in many power stations, bad data detection is

typically only offered at the central level. A linear weighted

least square-based state estimation method may detect incorrect

data from defective current transformers, according to [16].

To identify fake data in PMU measurements, most false data

detection algorithms need SCADA measurements.

Furthermore, current research indicates that PMU

measurements are not completely safe against cyber assaults

[17–19]. An attacker manipulates the measurements of the

PMU as well as the adjacent SCADA meters (RTUs) in the case

of an FDIA on the PMU measurements such that corrupt values

pass the measurement residual-based checks [20] at the energy

control centre and the attack is disguised [21-22]. As a result,

standard techniques for detecting, identifying, and correcting

incorrect data are ineffective in detecting purposeful misleading

data in PMU measurements.

In the case of accidental data corruption, the inherent high

noise level in SCADA measurements, low sample rate, and lack

of time-stamping render SCADA measurements unsuitable for

faulty data detection in sparsely located PMU measurements.

Apart from cyber assaults, accidental FDI can occur owing to

defective current or potential transformers, noise in the

communication channel, GPS jamming, and other factors [23-

25]. As a result, an alternate method for detecting both

deliberate and inadvertent misleading data in PMU

measurements has to be developed.

The Contribution of this paper includes,

 a deep complex neural network to handle complex

values include both voltage magnitude as well as

phase angle as a whole.

 Rather than using batch normalization, weight

normalization has already been added as a result of

the merging between repeated timestamps for

monitoring voltage magnitude and phase angle,

which may significantly enhance the model's

training performance.

 A deep complex network has been utilized with the

conjunction of a topology processor and AC state

estimator, topology processors detect substantial

information about the network topology to

recognize faulty data caused by topology change

caused by disruptions.

The content of the paper is organized as follows: section 1

represents the introduction; section 2 presents the related work;

the novel solutions are presented in section 3; the

implementation results and its comparison are provided in

section 4; finally, section 5 concludes the paper.

Amutha, et al [26] aimed at detecting abnormalities in

streaming PMU multivariate information in smart grid by

taking into account all attributes utilizing the Density

Estimation Technique depending on Gaussian Mixture Theory.

The suggested architecture has been evaluated in both offline as

well as online forms of data streams, but also the research

findings show that the suggested technique performs

2. Literature Survey

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

competently in identification. This methodology has a lower

false positive as well as false-negative rate, it may be used for

real-time anomaly monitoring. The suggested model is

additionally validated for streaming data scenarios by being

evaluated to current research in respect of the accuracy, recall,

and F1 score.

Zhou, et al [27] Utilizing online learning and a multivariate

data-drift detection technique, a unique device-level deep

learning-based data-driven strategy for anomaly detection,

localization, and classification over streaming PMU data is

proposed. Dynamic data Change Driven Learning (DCDL) but

also Continuity Driven Learning (CDL) are presented as well

as contrasted as PMUNET variations. The DCDL technique

surpasses the CDL as well as similar standard approaches. Thus

the suggested methodology in this paper efficiently detects the

data anomalies.

Rehan, et al [28] An efficient attack strategy has been

proposed to detect the false data injected into the power system.

Using linear regression the false data is injected into the power

system,also it is used by FDI to create any assault which could

overcome BDD but also SVM-based defence technologies.

Monitoring intelligent grid assaults allows us to ensure that

command choices in the control room are predicated upon

trustworthy assessments and that deceptive information is

eliminated resulting in computing complexity.

Yang, et al [29] A data-driven PMU poor data identification

technique based on spectral clustering utilizing single PMU

data is presented to improve PMU data quality. The topology

and characteristics of the system are not required by the

suggested approach. First, utilizing the slope characteristic of

each data, a data identification technique based on a decision

tree is presented to identify event data from bad data. Then,

using spectral clustering, a technique for detecting faulty data

is devised. This approach can detect faulty data with a minor

variation by evaluating the weighted relationships among all the

data.

Jovicic, et al [30] A linear approach is provided for state

estimate of power systems that are monitored using both

conventional and synchrophasor measurements, including bad

data detection. Both forms of data are processed at the same

time, with states approximated in rectangular coordinates. The

linear weighted least square approach is used to create the

suggested estimator. The network is represented in terms of

voltages and currents in the rectangular form to permit the

generation of linear measurement functions, and pseudo-

measurements are employed to simulate traditional

measurements. Furthermore, to detect faulty data, the biggest

normalized residual test is utilized.

In [26] difficulty in handling bad data .[27] detects data

anomalies.[28] analyzes about the computing complexity . The

paper [29] detect the faulty data with some minor variation

only. To detect faulty data , the bigget normalized residual test

is used which is complex [30]. To overcome the above

mentioned issues a robust methodology is proposed which is

explained in the upcoming section.

PMUs are used to measure the phasor quantity in the grid

system. Because of the glitch or ripple, it produces false or bad

data at the output.These anomaly may reduce the accuracy of

the of the system and leads to performance degradation . Dual

identical models were used for analyzing the real and imaginary

data PMU independently which increases the difficulty in

computational process also the phase angle variations may

cause false data occurrence. Similarly, poor data owing to

topology changes have not been adequately detected using

simply the AC state estimator. As a result, a Robust Bad Data

Technique was designed which uses a deep complex neural

network to process complex numbers with both voltage

magnitude and phase angle overall. Weight normalization has

been integrated into the deep complex neural network owing to

the fusion of recurrent timestamps for monitoring voltage,

magnitude and phase angle, which may further enhance model

training efficiency. With the help of prior data measured of

smaller complexity, the proposed deep complex neural network

recognizes false information in PMU assessments.As a result,

the suggested deep complex network has been used in

conjunction with a topology processor and an AC state

estimator, with the topology processor recognizing substantial

changes in the network topology to achieve incorrect data

identification owing to topology variation caused by

disturbances. Bad data in the multiple PMU power grid

networks may be rectified correctly using the proposed

methodology. The Data neural Network consists of the hidden

layer which may perform the exactly needed function of the

proposed so that the required output can be obtained without

any fault. The Robust bad data detection technique with its

concept which is incorporated in it has been clearly shown

below in figure 1.

Figure 1: Robust bad data detection Technique overview

The Bad data deviates from the normal data. After evaluating

a substantial portion of data obtained, it was discovered that the

majority of the bad data occurs on its own and that the amount

of adjacent bad data seems to be no over than three. It is also

mentioned that the exceptions are all separated not even in

order. As an example, consider the magnitude. This may be

used for calculating amplitude, frequency, and rate of change in

frequency, where magnitude includes voltage as well as current

magnitude. The schematic representation of Multiple Bad data

occurrences is shown in figure 2

3. Bad Data Detection Using

Deep Complex Network

3.1 Bad Data Occurrence

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

(a)

(b)

(c)

Figure 2: A diagram of (a) 1 Bad data (b) 2 Bad data and

The bad data occurs in the input data itself due to the

occurrence of many inteferences such as glitch etc.The bad data

may includes normal data, irregularity data with higher

magnitude and irregularity data with lower magnitude.These

bad datas are differentiated by different colours.The blue colour

signifies normal data. The yellow colour signifies the

irregularity data with a higher magnitude and the orange colour

represents the irregularity data with a lower amplitude. The

number of contiguous bad data maybe 1 or 2 or 3 as represented

in figure 2. This is how the bad data initially occur in the source

system.

For n bus, AC power structure, the nodal expression

processed utilizing Ohm’s Law is as follows



















































 

 

 

 

 





























……… (1)

where Ii and Vi were the net current and voltage at the ith

bus, respectively. Pn×n denotes the bus admittance matrix of

the AC power system having (i,j) th element being Pij. If there

are m PMUs mounted in the system. Here equation 1 can be

written as 󰇟󰇠󰇟󰇠󰇣

󰇤…………… (2)

Here m×1 and (n m)×1 represents the voltage

vectors for PMU and non-PMU buses, correspondingly. The

admittances between the various PMU as well as non-PMU

buses were characterized by the associated bus admittance sub-

matrices PU n×m and n×(nm). I n×1 is the net current

vector. Then formula 2 can be rewritten as

  ……………….. (3)

If ( ) can be substituted by M and also  which

is the mean and variance vector correspondingly.

…………………. (4)

Although equation 4 depicts a non-linear network, this

simply offers a linear input connection between the voltages of

the PMU buses. Those must be stressed since the basic link

between M and in equation 4 is unknown. The deterministic

optimization method for determining M and  in equation 2 is

to utilize the historical real-time/dynamic SE results from the

corresponding mean and variance vectors of and I.

A PMU provides the phasor quantity such as the voltage

magnitude and the associated phase angle of the voltage To deal

with the complex voltages the deep complex neural networks

along with the AC system model has been incorporated. Each

deep complex neural network model is trained in a way so that

it can linearly express the ith PMU bus voltage measurement in

terms of other (m - 1) PMU bus voltage measurements.

The relationship with the AC system model with the bad data

detection is given using the expression,

󰇩





 󰇪

………….. (5)

Hence the above-mentioned equation has been replaced by

the weight terms and the affine relationship between the

variance were given as,







 … (6)

Where  are the weights associated with the bad data

detection  and were the affine relationship between the

mean-variance.

Normally, the system gets complicated due to the utilization

of 2m number of systems from the number of PMUs. Hence in

this proposed model we are utilizing m models from PMUs in

deep neural network through the mathematical concept called

the convolution of complex numbers for both Real as well as

imaginary numbers. Hence the complexity has been reduced.

Every Complex number has both the real part as well as the

imaginary part. Here the convolution of the complex filter

matrix along with the complex vector is performed where

C=A+iB………………………….. (7)

And the complex vector is given as h= x+iy. In this A and B

are the Real matrices and x and y are the real vectors. The

convolution operator is always distributive

(A * x – B * y) + i (B * x + A * y)………. ( 8)

Rewriting the convolution terms employing Matrix as

follows,

 󰇛󰇜

 󰇛󰇜󰇣 

  󰇤󰇣

󰇤……. (9)

Here R means the Real value represents the Value of the

1st

2nd

3rd

3.2 Model of an AC System

)XQFWLRQRI'HHS1HXUDO1HWZRUNV

ZLWK6WDWH(VWLPDWRU

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

Voltage Magnitude and then the I, the Imaginary Value

represents the Phase angle value in the above-mentioned

equation.

By using this convolution form in the deep neural networks,

providing the Voltage magnitude and the phase angle in the

single system provides the output efficiently. Hence the

complexity gets reduced due to the availability of a single

system in the proposed methodology. The Bad data Occurrence

may happen at any time so the detection of Bad data process is

time-dependent. It has to be detected in a periodical manner.

The Batch Normalization approach towards speeding neural

network training by normalizing its signal density for every

level. But in this proposed system we are concern about the

periodicity in the detection of bad data that is not achieved by

the BN so we are utilizing the Weight normalization method.

The WN is an alternative for BN. It utilizes implicit

normalization, that leads inside the standard of both the

outcome having nearly identical to the current standard of the

source. Weights, for instance, should be normalized and

multiplied by a learnt scaling parameter for the convolutional

layer:

 

   ………………. (10)

where represents the output,  were the optional

parameters, used for numerical stability, W represents the

layer weights,  represents the norm of weights for the output

channel and  denotes the convolution. The flow chart explains

the entire proposed system as follows,

Weight normalization has been used in the proposed

methodology due to the fusion of recurrent timestamps for

measuring voltage magnitude and phase angle. We have

overcome the disadvantage by using weight normalization. The

Flow chart for the entire Bad data processing for the novel

technique is given below.

Figure 3: Flow Chart for bad data detection process

The Deep complex neural network produces the output of the

predicted output for the voltage magnitude and the phase as



 .The phasor quantities measured by the PMU were

given as 

. Bad data can also be identified using the

topology processor. It detects the bad data that occurred due to

the variations in the topology. Numerous measurements shown

whenever an according to with of a PMU bus is adjusted by

efficiency%, the influence of this measurement on the forecast

of PMU bus voltage utilizing generates a prediction error that

is consistently smaller than efficiency %. As a result, when the

predicted bus voltage  is compared to its

corresponding observed equivalent the greatest

difference between the anticipated and measured bus voltages

is seen.

An AC power distributor along with the Topology processor

was incorporated for the betterment of the system. Thus by

introducing the novel ideology of a Deep complex neural

network, the topology processor incorporated with an AC

estimator enhances the system by reducing the Disadvantages

like complexity and increases efficiency. Thus the proposed

approach increases the efficiency, accuracy, decreases the

complexity. the bad data detection capability and detection

range or the rate of bad data detection have been increased. The

running time of the operation also decreased in the increase in

performance of the proposed system.

The upcoming section explains the results obtained by

MATLAB simulation and also the performance parameters

were compared and analyzed.

This section provides a description of various

implementation results and the performance analysis of our

proposed model and also the comparison section to ensure

enhancement of our proposed system.

This work has been implemented and the simulation of the

system was then done using MATLAB with the following

system specification and the simulation results are discussed

below,

Platform : MATLAB

OS : Windows 8

Processor: Intel Core i5

RAM : 8GB RAM

The following diagram depicts the design of the proposed

system. The Simulated design depicts the process carried over.

Initially the voltage of three-phase , where the three phases

are denoted as a,b,c respectively along with the phasor

quantities like the magnitude of the voltage and also the phase

of the voltage was obtained per unit (PU). Here the Phasor

Measurement Unit (PMU) designed is based on phase locked

loop-based positive –sequence. A PLL is a closed-loop system

with a control mechanism to reduce any phase error that may

occur.

Then the Deep Complex Neural System incorporated with

the topology processor and the SE (State Estimator) performs

the required function to produce the required results. Finally,

the Bad data has been detected from the outputs obtained.

Source i/p

(Pmu o/p)

 (DCNN o/p)

State Estimator and

Topology processor

No bad data

Bad data detected

PMU o/p =DCNN o/p

4. Results and Discussions

4.1 Experimental Setup

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

Figure 4: Simulated design for bad data detection

The Simulated design depicts the process carried over.

Initially the voltage of three-phase , where the three phases

are denoted as a,b,c respectively along with the phasor

quantities like the magnitude of the voltage and also the phase

of the voltage was obtained per unit (PU). Here the Phasor

Measurement Unit (PMU) designed is based on Phase Locked

Loop-based positive –sequence which is a closed-loop system

with a control mechanism to reduce any phase error that occurs.

Then the Deep Complex Neural System incorporated with the

topology processor and the SE (State Estimator) performs the

required function to produce the required results. Finally, the

Bad data has been detected from the outputs obtained. The next

diagram deals with the intrinsic functioning that happens in the

proposed system.

Figure 5: Simulated Design of Source

Figure 5 shows the internal functioning of the source. From

the source, only the three-phase voltage along with the phasor

quantity such as the voltage magnitude, phasor angle in degrees

and also the frequency in hertz can be derived.

Figure 6: Internal Function of Deep Complex

Neural Network

The output from the PMU, the voltage magnitude and the

phase were given to the DCNN. As we already know that the

DCNN consists of a hidden layer and they perform some

functions so that the output is obtained from the other side as

y2 depicted in figure 6.

Figure 7: Output of Three Phase voltage

Three-phase voltage is initially received from the source

input power grid. So finally the output obtained is also three-

phase voltage.The output voltage obtained is maximum at time

sequence 1s to 1.5s . The graph shows the final output .The

three phases are differentiated by various colours in figure 7.

Figure 8: Simulated Output of Magnitude

The graphical representation in figure 8 has been drawn

concerning time in seconds. This depicts both the active and the

reactive magnitude sequence, where active represents the used

magnitude reactive represents the unused magnitude. The

active magnitude range reaches a maximum of 2.25 V and

constant throughout the graph figure 8.

Figure 9: Simulated out of phase sequence in bad data detection

Graph in figure 9 depicts the simulated output of the phase

/angle in degree obtained finally. Here also both the reactive, as

well as the nonreactive angles, are shown in different colours.

In figure 9 range of the phase sequence for both reactive and

nonreactive angles reaches the maximum 180 .

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

Figure 10: Simulated output of Frequency in bad data detection

Graph figure 10 depicts the simulated output of the frequency

in hertz. Here both the reactive, as well as nonreactive

frequencies were differentiated for the evaluation purpose. In

our proposed system , Weight normalization have being

incorporated into to the deeply complex neural network as a

result of the integration of recurrent timestamps for monitoring

voltage, magnitude, and phase angle, which improve model

training efficiency even more.Thus the simulated output

obtained shows better performance due to the AC state

estimator and the Deep complex neural network in the proposed

methodology .

The comparative analysis provides better performance

parameters when compared to the existing methods such as

SVR, BP Neural Network, Spectral Clustering (SC), Ensemble

Method(EM) and Density-Based Clustering (DBSCAN) [29].

In the following analysis, the parameters such as the

Accuracy in percentage, Bad data Detection capability, Bad

Data Detection range in terms of the deviation of phase angle,

Running time in terms of time window as well as the data points

were compared and the comparison graphs are given in the

corresponding order.

Figure 11: Accuracy Comparision

The Novel proposed technique using a Deep Complex neural

network provides higher performance when compared to the

previously existing techniques such as Decision Tree, SVM, BP

Neural Technique has been highlighted in figure 11

Sl.N

Techniques

Accuracy(

BP Neural Technique

Support Vector

Machine

Decision Tree

Deep Neural

Technique

97%

97.2%

98.4%

99.5%

The Comparision of the detection capability of the bad data

occurrence by the existing method with the proposed system is

systematically represented utilizing Graphical manner in figure

Figure 12: Comparison of Bad data detection Capability

Considering the maximum ratio of detection capability is 20

%. This graph shows that when the bad data detection ratio is

higher than 11% the DBSCAN cannot detect completely. When

the ratio reaches above 12% EM cannot detect the bad data.

when it reaches above 15% also the spectral clustering won't

detect it completely. But the proposed system shows betterment

by detecting the bad data that occurred above the ratio level of

about 18 %. Thus our novel approach produces better results in

the detection capability.

The next parameter analysed for the comparison is the

detection range according to the deviation angle. The bad data

also occurs due to the phase angle deviation based on the phase

angle deviation the bad data detection range is observed. Here

considering the maximum deviation phase angle is 5. The

graphical representation of the detection range in terms of

deviation phase angle is given below.

Figure 12: Comparison of detection range

The DBSCAN will not detect any bad data when the below

deviation angle of 5. EM will not detect any bad data when the

deviation angle ranges below 1. The spectral clustering method

will not detect any bad data when the deviation angle ranges

4.2 Comparative Analysis

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

below 0.8. In our proposed system, the bad data can be detected

from 0.2 deviations of phase angle to the maximum of 5 . Hence

it is again proved that our proposed system shows more

improvement than the old methodologies.

The next parameter compared is the Running time. The

running time is calculated in seconds. Hence we are using

Weight normalization the parameter called the running time is

reduced for our proposed system. This is compared for both the

time window as well as the data points. The graphical

representation of the running time in terms of the time window

and the data points were compared and analyzed in figure 13.

A decrease in running time increases the efficiency of the

system.

Figure 13: Comparison of Running time with Time windows

Based on the time windows the data points are calculated. So

for time window 1, the data points appeared 50. For Time

window 2 the data points given are 100. Thus increase in the

time window increases the Datapoint. The Running time of the

system in the old existing techniques such as the EM, SC, DB

were much higher for the data points analyzed. The same data

points, as well as the time windows, are analyzed for our

proposed system of Novel bad detection techniques. The

Running time of the system is lower when compared.

The running time for the EM technique is about 0.05 seconds.

Running time for DB is 0.029 seconds and for SC technique is

about 0.028 seconds. But the proposed system produces a lesser

running time of 0.024 seconds which is comparatively lesser

than the previously existing technique. The graphical

representation of running time for both the time window and the

data points were shown as a graphical representation in figure

13 and 14.

Figure 14: Comparison of Running time along with Datapoints

Thus the simulated output for the robust Bad data

detection technique was explained along with the functional

diagram of the source and Deep complex neural network. Also,

the simulated analysis and the output for the voltage of three-

phase (Vabc), Magnitude sequence of the voltage ( |u| ), Phase

angle of the voltage, Frequency in Hertz were derived.

Then the comparative analysis of the various parameters is

also analyzed and depicted in the corresponding figures.

The Proposed Power Data Quality improvement through

PMU Bad Data Detection Based on Deep Complex Network

approach was tested effectively and its superiority over other

models was determined. The current strategy provides better

accuracy of 99.5% over other existing models. Also, the

complexity of the basic system has been reduced by introducing

the Deep complex neural network also the detection of bad data

due to the phase angle deviation is also achieved by the novel

approach. Lesser training time is also achieved by using weight

normalization. It also shows betterment by detecting the bad

data that occurred above the ratio level of about 18 % of 20% .

The running time for the execution of 200 data points is reduced

to 0.023 s. The bad data can be detected from 0.2 deviations of

phase angle to the maximum of 5. Due to the implemented

Novel approaches, the efficiency of our system increased with

higher data detection rate when compared with the other

existing systems.

[1]. O.I. Abiodun, E.O. Abiodun, M. Alawida, R.S.

Alkhawaldeh, H. Arshad, “A Review on the Security of the

Internet of Things: Challenges and Solutions,” Wireless

Personal Communications, 2021, pp. 1-35.

[2]. Jia, Ke, et al. "Transient high-frequency impedance

comparison-based protection for flexible DC distribution

systems." IEEE Transactions on Smart Grid, vol. 11, no. 1

pp. 323-333, 2019.

[3]. S. Shankar, K.B. Yadav, A. Priyadarshi, and V. Rathore,

“Study of phasor measurement unit and its applications,”

In Recent Advances in Power Systems, 2021, pp. 247-

257). Springer, Singapore.

[4]. Usman, Muhammad Usama, and M. Omar Faruque,

"Applications of synchrophasor technologies in power

systems,” Journal of Modern Power Systems and Clean

Energy, vol. 7, no. 2, pp. 211-226, 2019.

[5]. Jia, Ke, et al., "Sparse voltage measurement-based fault

location using intelligent electronic devices,” IEEE

Transactions on Smart Grid, vol. 11, no. 1, pp. 48-60, 2019.

[6]. Liu, Hao, et al. "Data-Driven Subsynchronous Oscillation

Identification Using Field Synchrophasor

Measurements." IEEE Transactions on Power

Delivery, 2021.

[7]. D. Ghosh, “Novel trends in resilience assessment of a

distribution system using synchrophasor application: A

literature review,” International Transactions on Electrical

Energy Systems, vol. 31, no. 8, pp. e12934, 2021.

[8]. Mian, Ishaq, "Probability Estimation Models to Predict

Risk Occurrence in Smart Grid Projects." PhD diss., The

George Washington University, 2021.

5. Conclusion

References

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022

[9]. Li, Jian, Chaowei Sun, and Qingyu Su, "Analysis of

cascading failures of power cyber-physical systems

considering false data injection attacks,” Global Energy

Interconnection, vol. 4, no. 2, pp. 204-213, 2021.

[10]. C. Pei, Y. Xiao, W. Liang, X. Han, “A Deviation-Based

Detection Method Against False Data Injection Attacks in

Smart Grid,” IEEE Access, vol. 12, no. 9, pp. 15499-509,

2021.

[11]. X. Luo, M. Pajic, and M.M. Zavlanos, “An optimal graph-

search method for secure state

estimation,” Automatica, vol. 123, p.109323, 2021.

[12]. S. Hopkins, E. Kalaimannan, C.S. John, “Sub-Erroneous

Outlier Detection of Cyber Attacks in a Smart Grid State

Estimation System,” In2020 11th IEEE Annual Ubiquitous

Computing, Electronics & Mobile Communication

Conference (UEMCON), 2020, pp. 0447-0454. IEEE.

[13]. Pagnier, Laurent, and Michael Chertkov, "Embedding

Power Flow into Machine Learning for Parameter and

State Estimation,” arXiv preprint arXiv:2103.14251, 2021.

[14]. A. R. Kulkarni, & M. S. Ballal, “Synergism of

synchrophasor measurements and data analytics for

enhancing situational awareness of power grid,”

Computers & Electrical Engineering, vol. 93, pp. 107231,

2021.

[15]. Lu, Jiyong; Wang, Weizhen, Li, Li, Guo,

Yanping, “Unscented Kalman filtering for nonlinear

systems with sensor saturation and randomly occurring

false data injection attacks,” Asian Journal of Control,

2019, asjc.2262–.

[16]. Gu, Yingzhong, Zhe Yu, Ruisheng Diao, and Di Shi,

"Doubly-fed Deep Learning Method for Bad Data

Identification in Linear State Estimation,” Journal of

Modern Power Systems and Clean Energy, vol. 8, no. 6,

pp. 1140-1150, 2020.

[17]. Shereen, Ezzeldin, et al., "Feasibility of time-

synchronization attacks against PMU-based state

estimation," IEEE Transactions on Instrumentation and

Measurement, vol. 69, no. 6, pp. 3412-3427, 2019.

[18]. S. V. S. Chauhan, & G. X. Gao, “Synchrophasor Data

Under GPS Spoofing: Attack Detection and Mitigation

Using Residuals,” IEEE Transactions on Smart Grid, vol.

12, no. 4, pp. 3415–3424, 2021.

[19]. Sundararajan, Aditya, et al., “Survey on synchrophasor

data quality and cybersecurity challenges, and evaluation

of their interdependencies,” Journal of Modern Power

Systems and Clean Energy, vol. 7, no. 3, pp. 449-467,

2019.

[20]. M. Dorier, G. Frigo, A. Abur, M. Paolone, “Leverage Point

Identification Method for LAV-Based State Estimation,”

IEEE Transactions on Instrumentation and Measurement,

vol. 70,pp.1-0, 2021.

[21]. Aljohani, Nader, and Arturo Bretas, "A Bi-level model for

detecting and correcting parameter cyber-attacks in power

system state estimation,” Applied Sciences, vol. 11, no. 14,

pp. 6540, 2021.

[22]. B. Li, R. Lu, & G. Xiao, “Detection of False Data Injection

Attacks in Smart Grid Cyber-Physical Systems,” Springer,

2020.

[23]. L. Kumar, & P. K. Kherdikar, “An overview of different

static state estimation techniques with bad data detection,

identification and elimination in power

system,” International Journal of Ambient Energy, pp. 1-5,

2021.

[24]. Khare, Gaurav, Mohapatra, Abheejeet, S.N. Singh, “A

Real-Time Approach for Detection and Correction of False

Data in PMU Measurements,” Electric Power Systems

Research, vol. 191, pp. 106866–,2021.

[25]. M. Shahraeini, A. Alvandi, & S. Khormali, “Behavior

Analysis of Random Power Graphs for Optimal PMU

Placement in Smart Grids,” 2020 10th International

Conference on Computer and Knowledge Engineering

(ICCKE), 2020.

[26]. A. L. Amutha, R. Annie Uthra, J. Preetha Roselyn, & R.

Golda Brunet, “Anomaly detection in multivariate

streaming PMU data using density estimation technique in

wide area monitoring system,” Expert Systems with

Applications, vol. 175, pp. 114865, 2021.

[27]. A. Ahmed, K. S. Sajan, A. Srivastava, & Y. Wu, “Anomaly

Detection, Localization and Classification Using Drifting

Synchrophasor Data Streams,” IEEE Transactions on

Smart Grid, vol. 12, no. 4, pp. 3570–3580, 2021.

[28]. Nawaz, Rehan, Rabbaya Akhtar, Muhammad Awais

Shahid, Ijaz Mansoor Qureshi, and Muhammad Habib

Mahmood, "Machine learning based false data injection in

smart grid,” International Journal of Electrical Power &

Energy Systems, vol. 130, pp. 106819, 2021.

[29]. Yang, Zhiwei, et al., "Bad data detection algorithm for

PMU based on spectral clustering," Journal of Modern

Power Systems and Clean Energy, vol. 8, no. 3, pp. 473-

483, 2020.

[30]. Jovicic, Aleksandar, and Gabriela Hug, "Linear state

estimation and bad data detection for power systems with

RTU and PMU measurements,” IET Generation,

Transmission & Distribution, vol. 14, no. 23, pp. 5675-

5684, 2020.

Creative Commons Attribution License 4.0

(Attribution 4.0 International, CC BY 4.0)

This article is published under the terms of the Creative

Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en_US

International Journal of Electrical Engineering and Computer Science

DOI: 10.37394/232027.2022.4.5

Preeti Kabra, D. Sudha Rani

E-ISSN: 2769-2507

Volume 4, 2022