This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS 1

False Data Injection Attacks on LFC Systems: An

AI-Based Detection and Countermeasure Strategy
Zhixun Zhang , Jianqiang Hu , Member, IEEE, Jianquan Lu , Senior Member, IEEE,
Jinde Cao , Fellow, IEEE, and Jie Yu, Member, IEEE

Abstract— The development of integrated energy system has enable real-time monitoring, dynamic control, and information
promoted the development of the power system into the largest services through seamless integration of computing systems,
and most complex cyber-physical energy system, where the communication networks, and physical entities [4], [5], the
change in the openness of the load frequency control (LFC)
system has also increased its susceptibility to false data LFC system has evolved into an open system. Due to its
injection attacks (FDIAs). This paper porposes an artificial openness, the LFC system may be exposed to various potential
intelligence (AI) based detection and countermeasure strategy safety hazards, primarily because the phasor measurement
to protect LFC systems from FDIAs. First, The levenberg- unit (PMU) in the LFC system engages in high-speed, real-
marquarelt-back propagation (LM-BP) neural networks (NNs) is time data collection, and complex cryptography, which can
trained by collecting the historical data of frequency deviation,
power deviation of contact line, and active power load deviation. impact the PMUs’ performance. Applying complex cryptog-
Second, the output control signal of LM-BP NNs is tested against raphy to PMUs is not a feasible solution, rendering the LFC
the output of the system controller for residuals to determine system susceptible to cyber attacks [6].
whether has FDIAs. Furthermore, to resist the negative effects Malicious cyber attacks have the potential to go undetected
of FDIAs, the control signal calculated by the LM-BP NNs by the power system, leading to various negative conse-
will replace the traditional proportion integration differenti-
ation (PID) output. Finally, The advantages of the proposed quences, such as heightened generation costs, transmission
strategy are illustrated by the two interconnected power systems. line outages, and cascading failures [7], [8]. These attacks
come in different forms, including Denial of Service (DoS)
Index Terms— LFC system, LM-BP NNs, false data injection
attack, detection, countermeasure. attacks [9], topology attacks [10], and false data injection
attacks (FDIAs) [11]. Among these, FDIAs are considered
I. I NTRODUCTION particularly challenging as they pose a significant threat to
power system stability.
T HE system frequency is a significant metric for measuring
the power condition of the smart grid, as its stability
ensures the power system’s efficient operation and the delivery
Regarding FDIAs, extensive research has been dedicated
to detecting and devising countermeasure strategies for var-
of high-quality power to consumers [1]. The load frequency ious application scenarios in smart grids, including DC/AC
control (LFC) system, an integral element of power sys- transmission or estimation systems [12], [13], [14], [15],
tems’ energy management, plays a critical role in the normal distribution systems [16], and LFC systems [17]. In one
functioning of the power system [2], [3]. However, with the study [18], a greedy algorithm was developed to defend
advent of modern power cyber-physical systems (CPS), which against FDIAs through the strategic placement of secure PMU
equipment. Jiang et al. [19] monitored changes in measure-
Manuscript received 31 May 2023; revised 25 July 2023; ments and states between two time slots as a means of
accepted 16 August 2023. This work was supported in part by the
National Natural Science Foundation of China under Grant 61703095, Grant
detecting FDIAs. Additionally, some effective deep learning
61833005, Grant 61973078, and Grant 51977032; in part by the Natural and neural network methods were proposed for the detection
Science Foundation of Jiangsu Province of China under Grant BK20231416; and defense against FDIAs [16], [20], [21]. Furthermore,
and in part by the Zhishan Youth Scholar Program from Southeast University
under Grant 2242022R40042. This article was recommended by Associate
from the attackers’ perspective, FDIAs have been known to
Editor G. Chen. (Corresponding author: Jianqiang Hu.) exploit incomplete power grid information or reduced network
Zhixun Zhang is with the School of Cyber Science and Engineering, South- information to target power systems [10], [22], [23].
east University, Nanjing 210096, China (e-mail: [email protected]).
Jianqiang Hu is with the School of Mathematics, Southeast University, In the LFC system, Bad data detection (BDD) systems
Nanjing 211189, China (e-mail: [email protected]). are commonly employed as the primary defense against
Jianquan Lu is with the School of Mathematics, Southeast University, attacks [24]. However, crafty attackers may attempt to design
Nanjing 211189, China, and also with the School of Electronic Information
and Electrical Engineering, Chengdu University, Sichuan Chengdu 610106 FDIAs in a way that circumvents BDD detection by lever-
P.R. China (e-mail: [email protected]). aging the system’s structure and status [25]. Of all the cyber
Jinde Cao is with the School of Mathematics, Southeast University, Nanjing attacks on LFC systems, FDIAs significantly impact the stable
211189, China, and also with the Yonsei Frontier Laboratory, Yonsei Univer-
sity, Seoul 03722, South Korea (e-mail: [email protected]). operation of the smart grid. These attacks can disrupt the state
Jie Yu is with the School of Electrical Engineering, Southeast University, estimator and manipulate it to produce erroneous commands in
Nanjing 210096, China. the power system [26]. Furthermore, FDIAs exploit loopholes
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TCSI.2023.3307821. in the LFC system’s bad data detection process, enabling
Digital Object Identifier 10.1109/TCSI.2023.3307821 malicious tampering with state estimation results. This, in turn,
1549-8328 © 2023 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

leads to incorrect commands issued by the control center, pos-

ing serious risks to the safety of the power system [27]. Even
a small amount of data injection resulting from a successful
FDIA can trigger instability in the LFC system [28], [29].
Thus, investigating methods for FDIA detection and defense
becomes both necessary and urgent.
Literature along this line can be traced back to [8], where an
optimal FDIA against automatic generation control (AGC) was
studied. Chen et al. [30] considered the effect of coordination
attacks in the LFC system and used the threshold-based detec-
tion method to detect these kinds of attacks. Mohan et al. [31] Fig. 1. The LFC system model.
proposed a covert attack, which can maintain the relative
offset between the LFC system’s actual frequency deviation the model’s uncertainty and decrease the error in defense
and take control of it by employing a covert reference signal. and detection. The LM-BP NNs is trained on historical LFC
With in-depth research on the security issues of the LFC system data, and the BP NNs enhanced by the LM algorithm
system, there are a myriad of methods against FDIAs. A neu- can be utilized to handle complex LFC system problems. The
ral network-based detection method to estimate and detect function of detecting and defending FDIA can be realized
FDIAs injected to sensing loop of the system was proposed after the LM-BP NNs training process is completed. The main
in [32]. This method has sufficient accuracy to detect in real contributions of this paper include:
time and estimated FDIA. Then Khalaf et al. [33] took into (1) From the attacker’s perspective, two different FDIAs
consideration the AGC system’s nonlinearity and suggested a under the LFC system are analyzed, which involve attack-
brand-new particle filter-based method to find FDIAs in AGC ing only the frequency deviation and attacking both the
systems while taking into account both the communication frequency deviation and the contact line power deviation
time-delay and governor dead-band nonlinearities. In [17], simultaneously.
a unique mechanism based on the GAN network was proposed (2) An AI-based detection and countermeasure strategy that
to detect and defend FDIAs. Abbaspour et al. [11] proposed leverages the LM-BP NNs is proposed. The strategy per-
a novel resilient control system based on attack estimation forms residual testing between the output control signal of
that can effectively identify FDIAs and mitigate their negative the LM-BP NNs and the output of the proportion integra-
impacts. tion differentiation (PID) controller to detect the presence
The current approaches on detecting FDIAs in LFC sys- of FDIAs. Additionally, when FDIAs are detected, the
tems fall into two main categories: model-based and data- predicted control input can replace the original controller
driven approaches. Although the model-based approach does output effectively to provide a robust defense mechanism.
not require training or additional memory space to store a (3) By replacing the conventional PID controller output with
large number of samples, the system parameters and model the control signal from the LM-BP NNs, the negative
requirements of this approach are its main drawbacks. Small effects of FDIAs are effectively counteracted.
uncertainties in the system parameters may have an impact on
The outline of this paper is given as follows. Section II
the detection performance. On the other hand, the data-driven
describes the models of the LFC system and FDIAs.
method outperforms other methods in terms of detecting and
Section III illustrates the AI-based detection and counter-
defending FDIAs, but if convergence is slow, the system may
measure strategy. Validation studies are given in Section IV.
be burdened with computational work. In [17], training the
Section V gives the Discussion, and Section VI concludes
GAN network needs to reach Nash equilibrium, and an optimal
this paper.
strategy to achieve Nash equilibrium has not been found,
so training the GAN network is unstable. Although the resilient
II. M ODELS OF LFC S YSTEM AND FDIA S
control strategy proposed by [11] incorporates model-based
and data-based approaches with good defense performance, A. System Framework
its model complexity is too high and requires more system Frequency is an important performance indicator of the
resources. power system, and the main function of the LFC system
Due to the accuracy and efficiency of the neural network to is to maintain the frequency stability of the power system
detect FDIAs, In this study, an an artificial intelligence (AI) under control [34]. The frequency deviation mainly results
based method using levenberg-marquarelt-back propagation from the unbalance of the load and generator power. For the
(LM-BP) neural networks (NNs) is proposed to protect the purpose of preventing frequency deviation, the controller in the
LFC system against FDIAs, which has the advantages of LFC system will send an adjustment signal 1Pci to change
nonlinearity, non-limitation, and very qualitative, and the the turbine speed according to the load change, so that the
LM-BP algorithm’s computational speed can be significantly prime mover’s active output deviation 1Pmi can change the
increased by switching from the conventional BP gradient generator speed by following the load change 1PLi , which
descent approach to the LM algorithm, thus the detection can be seen from the control structure Fig. 1.
efficiency can be improved. In addition, the LM-BP neural The LFC system is an essential component of the AGC
network-based strategy suggested in this paper can get around system. In the island smart grid, the main goal of the LFC

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHANG et al.: FDIAs ON LFC SYSTEMS: AN AI-BASED DETECTION AND COUNTERMEASURE STRATEGY 3

system is to keep the frequency stable at a set value. While

in multibus smart grids, the LFC system must consider not
only the change of frequency, but also the influence of the
interconnection between smart grids. The LFC system aims to
stabilize the frequency of each control area at a predetermined
value and maintain the exchange power between areas at a
certain value. In the LFC system, the area control error (ACE)
is in the form of a linear equation which can be expressed by:
AC E i = 1Ptie,i + βi 1 f i , (1)
where 1Ptie,i is the tie-line power deviation of the ith power
area. βi is the frequency bias factor, and 1 f i is the frequency
deviation of ith area.
When the ACE signal is calculated, the LFC system will Fig. 2. Two-area LFC system model with potential FDIAs.
calculate the system power adjustment through the PID con-
the droop coefficient, the turbine’s time constant and the
troller. The system power adjustment P I Di = 1PCi of ith
governor’s time constant of the ith control area, Ti j is the
area can be expressed by:
synchronous power factor between the area i and j.
1PCi = f (AC E i )
Z
B. FDIA on the LFC System
= K pi AC E i + K I i AC E i + K di AC˙ E i , (2)
Most of the false data is introduced into the LFC system by
where K pi is the proportional gains; K I i is the integral gains, attackers to deceive the control center [35], then the control
and K di is the Derivative gains. center made the wrong decision, which in turn made the smart
Then the dynamic model of the LFC system are given by: grid in an unsteady state. As shown in Fig. 2, 1 f and 1Ptie
( are most vulnerable to be attacked, and the attacker may inject
ẋi (t) = Axi (t) + Bu i (t) + Cvi (t),
(3) the false data to change 1 f or 1Ptie . ACE signal is the
yi (t) = K xi (t), combination of these two signals, the false data are often
aggregated at the ACE signal and the controller may make
where xi (t) = [1 f i , 1Pmi , 1Pgi , 1Ptie,i ]T , and 1Pgi , 1Pmi
wrong adjustment decision because of receiving the incorrect
are the turbine valve position and the generator mechani-
ACE signal.
cal power deviation, respectively. u i (t) = 1PCi (t) is the
From the perspective of the attackers, the FDIA needs to
input vector of the system of ith area; vi (t) = 1PLi (t)
bypass the BDD system in the LFC system. BDD system
is the controlled input of ith area; the output y(t) =
typically uses the largest normalized residual test to detect
[AC E i , AC E i , AC˙ E i ]T ; A, B, C, K represent the state,
R
the presence of FDIAs in the system [17]:
input, load disturbance and output matrices: (

−Di 1 1
 ||z − H x||2 ≤ λ, no attack
0 − 0 0 (4)
 Mi Mi Mi ||z − H x||2 > λ, attack alarm
 
 1 1 
where z is the measurement variables, λ is the threshold set
 0 − 0 0 0

 Tti Tti 
 by the system in advance and H is the Jacobian matrix of the
 1 1 
smart grid.
 − 0 0 0 0
A= Ri Tgi Tgi Considering that the attacker has access to the current
 

n
configuration data of the power system, if attacker wants to
 P 
 2π Ti j 0 0 0 0 0
bypass the BDD system, the attack vector is set as follows:
 
 j=1, j̸=i 
βi
 
0 0 1 0 0
∥(z a − H xa )∥2 = ∥(z − H x) + (a − H c)∥2

n βi Di βi βi
 
P
2π Ti j − 0 − 00 ≤ ∥z − H x∥2 + ∥a − H c∥2 ≤ λ
 
Mi Mi Mi (5)
j=1, j̸=i

1
T From Eq.(5), if a = H c or ∥a − H c∥2 ≤ λ−∥z − H x∥2 , the
B= 00 000 attacker can successfully bypass the BDD system to launch
Tgi
FDIAs. After FDIAs bypass the bad data detection system,
βi T
 
1 there are two different scenarios of FDIAs and the AC E signal
C= − 0000−
Mi Mi can be expressed as:
β (1) The attacker only attacks the frequency signal 1 f :
 
0 0 1 00
0 0 0 0 1 0
AC E i = 1Ptie,i + βi (1 f i + 1 f a ),

K =  Pn βi Di βi βi

 (6)
2π Ti j − 0− 00
j=1, j̸=i M i M i M i (2) The attacker attacks both 1Ptie,i and 1 f at the same
time:
where Di , Mi , Ri , Tti , Tgi are respectively the generator equiv-
alent damping coefficient, the moment of inertia of generator, AC E i = (1Ptie,i + 1Pa ) + βi (1 f i + 1 f a ), (7)

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

of global and local convergence compared to traditional BP

algorithm [38]. In the actual inspection process, in order to
reduce the non-optimum singularity problem, the approximate
quadraticity of the approximate second derivative around the
extreme point is mainly utilized to speed up the optimization
and convergence process when the objective function is close
to the optimal point [39], [40]. Compared with the gradient
method, the LM algorithm is much faster. The iterative process
of the LM algorithm is given:
(1) The error indicator function can expressed as:
q
1X
e(k) = (di (k) − yi (k))2 , (9)
2
i=1

where di (k) and yi (k) represent the expected output and

actual output.
(2) Let w(k) be the vector composed of the threshold and
Fig. 3. The BP NNs structure.
weight, e(w) is the error function. According to LM
calculation method:
Then the PID controller calculates the incorrect 1PCai
h i−1
W (k + 1) = W (k) − J T (wk ) J (wk ) + µk I
regulation signal, which can be expressed as:
× J (wk ) e (wk ) , (10)
1PCai = 1PCi + a, (8)
where I is an identity matrix, and J is a Jacobian matrix.
where 1PCi denotes the original actual output and a is the
u k is a scale factor. If the scale factor is very huge, it is
deviation value resulted from the input attack vector.
similar to the gradient descent method; if the scale factor
Remark 1: In the above equations, two forms of FDIAs can
is zero, it is identical to the Gauss-Newton method.
be distinguished based on two factors. Firstly, each type of
FDIA targets different components. The first type of FDIA Remark 2: The trial parameter µ is used in this section,
focuses on attacking the frequency of the LFC system, whereas and for a given parameter, the variation of µ depending on
the second type targets both the frequency and tie-line power whether the computed threshold of this variable reduces the
of the LFC system. Secondly, these two types of FDIAs induce error function. Due to the utilization of data from the second-
distinct deviations in the LFC system’s frequency and tie-line order derivatives, The LM algorithm can be computed much
power. faster than the traditional BP method.

III. AI-BASED D ETECTION AND C OUNTERMEASURE B. The LM-BP NNs

S TRATEGY AGAINST FDIA S
The iterative process of LM-BP NNs is summarized:
To effectively defend against FDIA, an AI-based detection
(1) Initialize the parameters of the network structure, the
and countermeasure strategy against FDIAs is proposed. This
weight, and threshold vector. Furthermore, let k = 0,
approach involves utilizing LM-BP NNs in place of the
µ = µ0 , initialize the error allowable value ε, µ, and the
existing PID controller to measure the modification signal
constant β satisfies (0 < β < 1).
when the LFC system comes under FDIAs. By comparing
(2) Calculate network output and error function e;
the P I Dc adjustment calculated by the LM-BP NNs with
(3) Calculate Jacobian matrix J T (W ):
the P I D output computed by the PID controller, the detec-
∂e1 (W ) ∂e1 (W ) ∂e1 (W ) ∂e1 (W )
 
tion function is achieved. Subsequently, the calculated P I Dc
adjustment is used to replace the P I D, thereby accomplishing  ∂W1 · · · ∂W2 ∂θ11
···
∂θ L2 
 11 LL 
the countermeasure function.  ∂e (W ) ∂e2 (W ) ∂e2 (W ) ∂e2 (W ) 
 2
··· ···

The BP NNs is a multi-layer feedforward network [36],  ∂ W11
 1 ∂ W 2 ∂θ 1 ∂θ 2 
LL 1 L

it achieves the learning process through forward propagation .. .. .. .. .. ..
 
. . . . . .
 
of information and backward propagation of errors. The input  
 ∂e L (W ) ∂e L (W ) ∂e L (W ) ∂e L (W ) 
 
layer, hidden layer and output layer are the three layer structure · · · · · ·
of BP NNs [37], which can be shown in the Fig. 3. ∂ W11 1 ∂ W L2 L ∂θ11 ∂θ L2
(4) Calculate the modified weight 1w;
A. LM Algorithm (5) If e < ε, the iteration ends. Conversely, go to step (6);
According to the BP algorithm’s drawbacks, including its (6) Adjust new weights and threshold vectors by (10),
local minimum point and sluggish convergence speed, the and calculate error function e. If e(wk+1 ) < e(wk ), let
BP algorithm is improved by adopting the LM algorithm k = k + 1, µ = µβ, go back to the step (2); otherwise
based on optimization theory. LM algorithm has the advantage let µ = µ/β, go back to the step (4).

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHANG et al.: FDIAs ON LFC SYSTEMS: AN AI-BASED DETECTION AND COUNTERMEASURE STRATEGY 5

Fig. 5. The LM-BP NNs detection flowchart.

Algorithm 1 NN-Based Detection and Countermeasure

Strategy
1: Input: Real-time data: x = [1 f, 1Ptie , 1PL ].
Fig. 4. The LM-BP NNs training flowchart.
2: Train the LM-BP NNs.
3: for t=1,2,. . . ,T do
C. The Detection and Defense Strategy 4: Caculate P I Dc .
Based on LM-BP NNs 5: if |P I Dc − P I D| > ε
6: The LFC system is under attack.
This section focuses on the application of the proposed
7: Replace the P I D adjustment of the system with the
LM-BP NNs method to the LFC system. Firstly, data collect
calculated P I Dc .
is prepared to train the LM-BP NNs and the input data is
8: Issue control signals to actuators.
selected as xi = [1 f i , 1Ptie,i , 1PLi ], and the target data is
9: end if
yi = P I Di . The specific Procedures are present in Fig. 4.
10: end for
After the training of the LM-BP NNs is completed, the
11: Output: The security of the LFC system.
detection and defense of FDIAs can be carried out. The
LM-BP NNs first collects x at each moment of the LFC
system and calculates the corresponding adjustment P I Dc .
If |P I Dc − P I D| ≤ ε, it is considered that there is no attack,
otherwise, the LFC system was attacked, and the calculated
P I Dc is used to replace the P I D adjustment of the system.
Remark 3: The value of ε is determined according to the
system size and parameters. The grid frequency usually fluc-
tuates within 50 ± 0.2hz, so if |P I Dc − P I D| ≤ ε is satisfied,
it is considered that the grid frequency fluctuation does not
exceed 50 ± 0.2hz. On the contrary, it is considered that the
frequency fluctuation is large, and the system may be under
attack at this time, and a defense strategy should be taken in
time to keep the system stable. The specific steps are present
in Fig. 5.

IV. VALIDATION S TUDIES

In this section, the efficiency of the suggested strategy
is demonstrated using a simulation of two interconnected
LFC systems. which can be shown in Fig. 6. the simulation Fig. 6. Two interconnected LFC systems.
experiment is carried out on Simulink in Matlab 2019a, and
the details of the two interconnected LFC systems are provides
in Tab. I. The data is obtained by the simulation of the inter- as displayed in Tab. II. Among these samples, 7,000 were
connected two-area power system under the attack state and utilized for training, while 1,500 each were reserved for
the normal state, and three thousand system data are collected validation and testing purposes. Out of the gathered samples,
in this paper for the training samples of LM-BP NNs. This 4,000 were found to have FDIAs, with an equal distribution
paper presents the collection of 10,000 historical data samples, of 2,000 samples for each type of FDIA (i.e., the first and

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

TABLE I
PARAMETERS OF THE T WO I NTERCONNECTED LFC S YSTEMS

Fig. 8. 1 f under normal conditions.

TABLE II
T HE D IVISION OF THE DATA S ET

TABLE III
PARAMETERS C ONFIGURATION OF LM-BP N EURAL N ETWORK

Fig. 9. 1Ptie under normal conditions.

Fig. 10. 1 f under the frequency attack in the 1th area.

Fig. 7. Load disturbance of two-area power system.

second types). The parameters configuration of LM-BP NNs

are shown in Tab. III, according to the data size we set the
number of hidden layers to 20.
First, the first area on the frequency and tie-line power lines
are simultaneously attacked by FDI at 4s. We set up two
attack experiments: attacking 1 f on 1th area and attacking Fig. 11. 1Ptie under the frequency attack in the 1th area.
1 f and 1Ptie at the same time. The load disturbance are
presented in Fig. 7. Under the normal state, the LFC system is
operating normally, which can be verified by Fig. 8 and Fig. 9. shown in Fig. 10, and power deviation of contact line 1Ptie
Even though the frequency disturbance has an impact on the is presented in Fig. 11. Comparing with 1 f and 1Ptie in the
LFC system, the LFC system can still be modified to restore LFC system under normal conditions, it can be obtained that
stability. both of 1 f and 1Ptie have deviated from the set value, which
makes the system unstable for a long time.
In this experiment, the P I Dc adjustment calculated by the
A. The Attacker Only Attacks 1 f 1 LM-BP NNs and the system P I D adjustment are shown
In this attack experiment, the frequency of LFC system in in the Fig. 12. The comparison shows that the system is
1th area was attacked at 4s. The frequency deviation 1 f is attacked, then P I Dc is utilized to replace P I D to stabilize

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHANG et al.: FDIAs ON LFC SYSTEMS: AN AI-BASED DETECTION AND COUNTERMEASURE STRATEGY 7

Fig. 12. P I D adjustment under the frequency attack in the 1th area. Fig. 16. 1Ptie under the second FDIAs experiment in 1th area.

Fig. 13. 1 f with proposed mechanism under the frequency attack in

Fig. 17. P I D adjustment under the second FDIAs experiment in 1th area.
the 1th area.

Fig. 14. 1Ptie with proposed mechanism under the frequency attack in Fig. 18. 1 f with defense method under the second FDIAs experiment
the 1th area. in 1th area.

is shown in Fig. 16. One can demonstrate that the abso-

lute value of the frequency deviation remains above 0.02 pu
after 10s, and the power deviation of contact line remains
near 0.01 pu, which makes the system unstable for a long time.
The P I Dc adjustment calculated by the LM-BP NNs and the
system P I D adjustment are shown in the Fig. 17. After taking
defensive measures, the frequency deviation 1 f is shown in
the Fig. 18, and the tie-line power deviation 1Ptie is shown
in Fig. 19. It can be seen that when 1Ptie and 1 f are attacked
at the same time, the detection and defense measures are still
Fig. 15. 1 f under the second FDIAs on the frequency and tie-line power effective.
in 1th area.
Based on the two FDIAs experiment, the feasibility of the
proposed mechanism against FDIAs is verified. When the
the LFC system. From Figs. 13 and 14, it can be obtained that
proposed LM-BP NNs finds that the deviation between the sys-
the frequency deviation and tie-line power deviation return to
tem’s P I D adjustment and its calculated adjustment exceeds
normal values after utilizing the method proposed in this paper.
the range, it is determined that the system has been attacked,
and then P I Dc is used to send control signal to maintain the
B. The Attacker Attacks 1Ptie and 1 f 1 at the Same Time system stable. From the results of the two experiments, it can
In this attack experiment, the attacker attacks 1Ptie and 1 f 1 be concluded that the detection and defense methods proposed
at 4s. The frequency deviation 1 f in two area is presented in this article can cope with FDIAs. When the LM-BP NNs
in Fig. 15, and the power deviation of contact line 1Ptie detects that the system is subjected to FDIA, it can quickly

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS

[2] H. Luo, I. A. Hiskens, and Z. Hu, “Stability analysis of load frequency

control systems with sampling and transmission delay,” IEEE Trans.
Power Syst., vol. 35, no. 5, pp. 3603–3615, Sep. 2020.
[3] Z. Zhang, J. Hu, J. Lu, J. Cao, and F. E. Alsaadi, “Preventing false data
injection attacks in LFC system via the attack-detection evolutionary
game model and KF algorithm,” IEEE Trans. Netw. Sci. Eng., vol. 9,
no. 6, pp. 4349–4362, Nov. 2022.
[4] X. Li and D. Ye, “Security-based event-triggered fuzzy control for
multiarea power systems under cross-layer DoS attacks,” IEEE Trans.
Circuits Syst. I, Reg. Papers, vol. 70, no. 7, pp. 2995–3004, Jul. 2023,
doi: 10.1109/TCSI.2023.3268309.
[5] J. Hu, Z. Zhang, J. Lu, J. Yu, and J. Cao, “Demand response control of
smart buildings integrated with security interconnection,” IEEE Trans.
Fig. 19. 1Ptie with defense method under the second FDIAs experiment Cloud Comput., vol. 10, no. 1, pp. 43–55, Jan. 2022.
in 1th area. [6] Z. Li, M. Shahidehpour, A. Alabdulwahab, and A. Abusorrah, “Bilevel
model for analyzing coordinated cyber-physical attacks on power sys-
calculate the correct P I Dc adjustment of the system to keep tems,” IEEE Trans. Smart Grid, vol. 7, no. 5, pp. 2260–2272, Sep. 2016.
[7] Z. E. Mrabet, N. Kaabouch, H. E. Ghazi, and H. E. Ghazi, “Cyber-
the system stable. security in smart grid: Survey and challenges,” Comput. Electr. Eng.,
vol. 67, pp. 469–482, Apr. 2018.
V. D ISCUSSION [8] R. Tan et al., “Optimal false data injection attack against automatic
In the LFC system, the proposed AI-based method is generation control in power grids,” in Proc. ACM/IEEE 7th Int. Conf.
Cyber-Phys. Syst. (ICCPS), Apr. 2016, pp. 1–10.
capable of detecting and defending against FDIAs. Unlike the
[9] Y. Li, P. Zhang, and L. Ma, “Denial of service attack and defense method
general detection method, this method can take countermea- on load frequency control system,” J. Franklin Inst., vol. 356, no. 15,
sure after the FDIA is detected. The LM-BP NNs gives the pp. 8625–8645, Oct. 2019.
correct P I Dc adjustment in time according to the state of the [10] X. Liu and Z. Li, “Local topology attacks in smart grids,” IEEE Trans.
Smart Grid, vol. 8, no. 6, pp. 2617–2626, Nov. 2017.
system after the attack, so as to stabilize the LFC system. [11] A. Abbaspour, A. Sargolzaei, P. Forouzannezhad, K. K. Yen, and
Implementation: The AI-based method which utilize LM-BP A. I. Sarwat, “Resilient control design for load frequency control system
NNs proposed in this article is necessary to collect the data under false data injection attacks,” IEEE Trans. Ind. Electron., vol. 67,
no. 9, pp. 7951–7962, Sep. 2020.
[1 f, 1Ptie , 1PL ] during the operation of the system to train
[12] Q. Yang, J. Yang, W. Yu, D. An, N. Zhang, and W. Zhao, “On false
the neural network. After the training is completed, place it data-injection attacks against power system state estimation: Modeling
in parallel at the location of PID controller of the system, and and countermeasures,” IEEE Trans. Parallel Distrib. Syst., vol. 25, no. 3,
implement monitoring to ensure the safety of the system. pp. 717–729, Mar. 2014.
[13] Md. A. Rahman and H. Mohsenian-Rad, “False data injection attacks
Application: The LM-BP NNs can deal with the detection against nonlinear state estimation in smart power grids,” in Proc. IEEE
and countermeasure strategy against FDIAs. When an attacker Power Energy Soc. Gen. Meeting, Jul. 2013, pp. 1–5.
launches FDIAs, the frequency deviation and power deviation [14] Q. Yang, J. Yang, W. Yu, N. Zhang, and W. Zhao, “On a hierarchical
false data injection attack on power system state estimation,” in Proc.
of contact line will exceed the set value, and the LM-BP IEEE Global Telecommun. Conf. (GLOBECOM), Dec. 2011, pp. 1–5.
NNs will detect that the system is under attack and utilize the [15] O. Kosut, L. Jia, R. J. Thomas, and L. Tong, “Limiting false data attacks
calculated P I Dc adjustment to replace the system’s P I D. on power system state estimation,” in Proc. 44th Annu. Conf. Inf. Sci.
Syst. (CISS), Mar. 2010, pp. 1–6.
In summary, the proposed detection and countermeasure
[16] Y. Zhang, J. Wang, and B. Chen, “Detecting false data injection attacks
strategy treats FDIAs as a part of the LFC system, and the in smart grids: A semi-supervised deep learning approach,” IEEE Trans.
goal of the method is to detect FDIA in the LFC system and Smart Grid, vol. 12, no. 1, pp. 623–634, Jan. 2021.
then provide defensive measures to maintain the stability of [17] Y. Li, R. Huang, and L. Ma, “False data injection attack and defense
method on load frequency control,” IEEE Internet Things J., vol. 8,
the system. no. 4, pp. 2910–2919, Feb. 2021.
[18] T. T. Kim and H. V. Poor, “Strategic protection against data injection
VI. C ONCLUSION attacks on power grids,” IEEE Trans. Smart Grid, vol. 2, no. 2,
This paper focuses on the detection and countermeasure pp. 326–333, Jun. 2011.
strategy for the LFC system concerning FDIAs. By thoroughly [19] J. Jiang and Y. Qian, “Defense mechanisms against data injection
attacks in smart grid networks,” IEEE Commun. Mag., vol. 55, no. 10,
analyzing the LFC system and researching FDIAs, it has been pp. 76–82, Oct. 2017.
revealed that injecting FDI into the frequency and power [20] Y. He, G. J. Mendis, and J. Wei, “Real-time detection of false data injec-
lines can indeed affect the stability of the LFC system. tion attacks in smart grid: A deep learning-based intelligent mechanism,”
IEEE Trans. Smart Grid, vol. 8, no. 5, pp. 2505–2516, Sep. 2017.
To address this issue, we propose an AI-based detection and [21] A. Tabakhpour and M. M. A. Abdelaziz, “Neural network model for
countermeasure strategy for FDIAs, ensuring that the LFC false data detection in power system state estimation,” in Proc. IEEE
system can maintain its stability even when subjected to such Can. Conf. Electr. Comput. Eng. (CCECE), May 2019, pp. 1–5.
attacks. The efficacy of our proposed strategy is validated [22] Md. A. Rahman and H. Mohsenian-Rad, “False data injection attacks
with incomplete information against smart power grids,” in Proc. IEEE
through simulations on a power system interconnected in Global Commun. Conf. (GLOBECOM), Dec. 2012, pp. 3153–3158.
multiple areas. Future work will be dedicated to designing the [23] X. Liu, Z. Bao, D. Lu, and Z. Li, “Modeling of local false data injection
most cost-effective protection strategies for the LFC system, attacks with reduced network information,” IEEE Trans. Smart Grid,
vol. 6, no. 4, pp. 1686–1696, Jul. 2015.
enhancing its resilience against potential threats. [24] Y. Liu, P. Ning, and M. K. Reiter, “False data injection attacks against
state estimation in electric power grids,” ACM Trans. Inf. Syst. Secur.,
R EFERENCES vol. 14, no. 1, pp. 1–33, May 2011.
[1] J. Hu, J. Cao, J. M. Guerrero, T. Yong, and J. Yu, “Improving frequency [25] Y. Yuan, Z. Li, and K. Ren, “Modeling load redistribution attacks in
stability based on distributed control of multiple load aggregators,” IEEE power systems,” IEEE Trans. Smart Grid, vol. 2, no. 2, pp. 382–390,
Trans. Smart Grid, vol. 8, no. 4, pp. 1553–1567, Jul. 2017. Jun. 2011.

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZHANG et al.: FDIAs ON LFC SYSTEMS: AN AI-BASED DETECTION AND COUNTERMEASURE STRATEGY 9

[26] H. M. Merrill and F. C. Schweppe, “Bad data suppression in power sys- Jianqiang Hu (Member, IEEE) received the
tem static state estimation,” IEEE Trans. Power App. Syst., vol. PAS-90, B.S. degree in mathematics and applied mathematics
no. 6, pp. 2718–2725, Nov. 1971. from the North China University of Water Resources
[27] H. Liu, Y. Ni, L. Xie, and K. H. Johansson, “How vulnerable is and Electric Power, Zhengzhou, China, in 2010, and
innovation-based remote state estimation: Fundamental limits under the M.S. degree in applied mathematics and the
linear attacks,” Automatica, vol. 136, Feb. 2022, Art. no. 110079. Ph.D. degree in control theory and control engi-
[28] G. Liang, J. Zhao, F. Luo, S. R. Weller, and Z. Y. Dong, “A review of neering from Southeast University, Nanjing, China,
false data injection attacks against modern power systems,” IEEE Trans. in 2013 and 2016, respectively. Currently, he is an
Smart Grid, vol. 8, no. 4, pp. 1630–1638, Jul. 2017. Associate Professor with the Jiangsu Provincial Key
[29] H. Zhang, W. Meng, J. Qi, X. Wang, and W. X. Zheng, “Distributed Laboratory of Networked Collective Intelligence and
load sharing under false data injection attack in an inverter-based the Department of System Science, School of Math-
microgrid,” IEEE Trans. Ind. Electron., vol. 66, no. 2, pp. 1543–1551, ematics, Southeast University, China. His current research interests include the
Feb. 2019. distributed optimization and control of multiagent systems, and demand-side
[30] C. Chen, M. Cui, X. Wang, K. Zhang, and S. Yin, “An investigation control in smart grids.
of coordinated attack on load frequency control,” IEEE Access, vol. 6,
pp. 30414–30423, 2018.
[31] A. M. Mohan, N. Meskin, and H. Mehrjerdi, “Covert attack in load Jianquan Lu (Senior Member, IEEE) received
frequency control of power systems,” in Proc. 6th IEEE Int. Energy the B.S. degree in mathematics from Zhejiang
Conf. (ENERGYCon), Sep. 2020, pp. 802–807. Normal University, Zhejiang, China, in 2003, the
[32] A. Abbaspour, A. Sargolzaei, and K. Yen, “Detection of false data M.S. degree in mathematics from Southeast Univer-
injection attack on load frequency control in distributed power sys- sity, Nanjing, China, in 2006, and the Ph.D. degree
tems,” in Proc. North Amer. Power Symp. (NAPS), Sep. 2017, in applied mathematics from the City University of
pp. 1–6. Hong Kong, Hong Kong, in 2009.
[33] M. Khalaf, A. Youssef, and E. El-Saadany, “A particle filter-based From 2010 to 2012, he was an Alexander von
approach for the detection of false data injection attacks on automatic Humboldt Research Fellow of PIK, Germany. He is
generation control systems,” in Proc. IEEE Electr. Power Energy Conf. currently a Professor with the Department of Sys-
(EPEC), Oct. 2018, pp. 1–6. tems Science, School of Mathematics, Southeast
[34] Ö. Aydın, Ş. Sönmez, and S. Ayasun, “Determination of stability delay University. His current research interests include collective behavior in
margins for multi-area load frequency control systems with incommen- complex dynamical networks and multi-agent systems, logical networks, and
surate time delays through eigenvalue tracing method,” Int. J. Electr. hybrid systems. He has published over 100 articles in refereed international
Power Energy Syst., vol. 137, May 2022, Art. no. 107821. journals. He was named as a Highly Cited Researcher by Clarivate Analytics
in 2018 for three consecutive years, and he was elected Most Cited Chinese
[35] A. S. Bretas, N. G. Bretas, and B. E. B. Carvalho, “Further contributions
Researchers by Elsevier (2014–2019). He is an Associate Editor of Neural
to smart grids cyber-physical security as a malicious data attack: Proof
Processing Letters, Journal of Franklin Institute, and Neural Computing and
and properties of the parameter error spreading out to the measurements
Applications, and a Guest Editor of Science China: Information Sciences and
and a relaxed correction model,” Int. J. Electr. Power Energy Syst.,
IET Control Theory and Applications.
vol. 104, pp. 43–51, Jan. 2019.
[36] W. Jin, Z. J. Li, L. S. Wei, and H. Zhen, “The improvements of
BP neural network learning algorithm,” in Proc. 5th Int. Conf. Sig-
Jinde Cao (Fellow, IEEE) received the B.S. degree
nal Process. (WCC-ICSP), 16th World Comput. Congr., vol. 3, 2000,
in mathematics/applied mathematics from Anhui
pp. 1647–1649.
Normal University, Wuhu, China, in 1986, the
[37] Y.-J. Sun, S. Zhang, C.-X. Miao, and J.-M. Li, “Improved BP neural M.S. degree in mathematics/applied mathematics
network for transformer fault diagnosis,” J. China Univ. Mining Technol., from Yunnan University, Kunming, China, in 1989,
vol. 17, no. 1, pp. 138–142, Mar. 2007. and the Ph.D. degree in mathematics/applied math-
[38] J. J. Moré, “The Levenberg–Marquardt algorithm: Implementation ematics from Sichuan University, Chengdu, China,
and theory,” in Numerical Analysis. Berlin, Germany: Springer, 1978, in 1998.
pp. 105–116. He was a Postdoctoral Research Fellow at the
[39] S. Yang and G. Yang, “Emotion recognition of EMG based on Department of Automation and Computer-Aided
improved L-M BP neural network and SVM,” J. Softw., vol. 6, no. 8, Engineering, Chinese University of Hong Kong,
pp. 1529–1536, Aug. 2011. Hong Kong, from 2001 to 2002. He is an Endowed Chair Professor, the Dean
[40] X. Li and Y. Wang, “Prediction model of biogas production for anaerobic of the School of Mathematics and the Director of the Research Center for
digestion process of food waste based on LM-BP neural network and Complex Systems and Network Sciences at Southeast University (SEU). He
particle swarm algorithm optimization,” in Proc. Chin. Autom. Congr. is also the Director of the National Center for Applied Mathematics at SEU-
(CAC), Oct. 2017, pp. 7629–7633. Jiangsu of China and the Director of the Jiangsu Provincial Key Laboratory
of Networked Collective Intelligence of China.
Prof. Cao was a recipient of the National Innovation Award of China,
the Obada Prize, and the Highly Cited Researcher Award in Engineering,
Computer Science, and Mathematics by Thomson Reuters/Clarivate Analytics.
He is elected as a member of Russian Academy of Sciences, a member of
the Academy of Europe, a member of Russian Academy of Engineering, a
member of the European Academy of Sciences and Arts, a member of the
Lithuanian Academy of Sciences, a fellow of African Academy of Sciences,
and a fellow of Pakistan Academy of Sciences.

Jie Yu (Member, IEEE) received the B.S. and

M.S. degrees in electrical engineering and the
Zhixun Zhang received the B.S. degree in net-
Ph.D. degree in electricity engineering from South-
work engineering from North Minzu University,
east University, China, in July 1996, March 2000,
Yinchuan, China, in 2020. He is currently pursuing
and June 2009, respectively. She was a Power
the Ph.D. degree with the School of Cyber Science
System Monitoring Software Engineer with the
and Engineering, Southeast University, Nanjing,
State Grid Electric Power Research Institute, Nan-
China. His research interests include the security
jing, from April 2000 to March 2006. She
and protection of smart grid, and defense methods
was an Associate Professor with the School of
against false data injection attack.
Electrical Engineering, Southeast University. Her
research interests include power optimization dis-
patch, renewable generation, and power system monitoring technology.

Authorized licensed use limited to: National Institute of Technology Andhra Pradesh. Downloaded on September 05,2023 at 09:54:41 UTC from IEEE Xplore. Restrictions apply.