IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 68, NO.

2, FEBRUARY 2021 717

False Data Injection Cyber-Attacks Mitigation in

Parallel DC/DC Converters Based on Artificial
Neural Networks
Mohammad Reza Habibi , Graduate Student Member, IEEE, Hamid Reza Baghaee , Member, IEEE,
Tomislav Dragičević , Senior Member, IEEE, and Frede Blaabjerg , Fellow, IEEE

Abstract—Because of the existence of communication networks proposed to detect FDIAs in DC microgrids. The proposed
and control applications, DC microgrids can be attacked by method in [8] is based on artificial neural networks, and
cyber-attackers. False data injection attack (FDIA) is one type it can detect attacks on both current and voltage measure-
of cyber-attacks where attackers try to inject false data to the
target DC microgrid to destruct the control system. This brief
ments. Reference [9] proposes a method to detect two types of
discusses the effect of FDIAs in DC microgrids that are struc- cyber-attacks, i.e., FDIAs and denial of service (DoS) attacks
tured by parallel DC/DC converters and they are controlled by in the DC microgrid that is controlled based on cooperative
droop based control strategies to maintain the desired DC volt- distributed approaches. Also, in [10], a trust-based cooperative
age level. Also, an effective and proper strategy based on an controller is proposed in DC microgrids based on the dis-
artificial neural network-based reference tracking application is tributed approach to mitigate FDIAs on communication links
introduced to remove the FDIAs in the DC microgrid.
and controller hijacking. Furthermore, in [11], a framework is
Index Terms—Artificial neural networks, cyber-attack, DC proposed based on the detection of a change in the set of can-
microgrid, droop control, false data injection attack. didate invariants in the DC microgrid that is controlled based
on the consensus theory.
I. I NTRODUCTION Most of the mentioned works focus on the detection of
FDIAs. Also, the majority of them work on DC microgrids,
C MICROGRIDS need improved voltage regulation and
D load sharing to have a reliable and economical operation,
and droop control is a common application of load sharing in
which are controlled based on distributed controller and con-
sensus protocol. FDIAs try to inject the false data into the
system to convert a data with real value to a data with a false
microgrids [1]. To control the DC microgrids, a hierarchical
value and because of that, this brief has implemented an arti-
control structure can be implemented [2], [3]. To implement
ficial neural network as a type of data based-techniques to
the hierarchical control of a droop controlled DC microgrid,
prevent the data conversion from a real value to a false value.
measurement devices, and the communication link are needed,
Also, focuses are on DC microgrids, which are structured by
which makes the DC microgrids vulnerable to cyber-attacks.
parallel DC/DC converters, and they are controlled hierarchi-
There are different kinds of cyber-attacks, e.g., replay attacks,
cally and based on the droop control method. Furthermore, the
denial of service (DoS) attacks, and FDIAs [4]–[6]. FDIA tries
proposed strategy can detect and also mitigate FDIAs in the
to change the state of the system via injecting false data [7],
DC microgrids.
and this brief focuses on mitigation of the FDIAs in DC
microgrids, which are structured by parallel DC/DC converters
and controlled by hierarchical droop based control structure. II. C ONVENTIONAL C ONTROL OF PARALLEL
Recently, some works are done to analysis of FDIAs in DC C ONVERTERS IN DC M ICROGRIDS AND FDIA S
microgrids. In [6], a method is introduced to detect FDIAs Fig. 1 shows the implementation of secondary and primary
on current measurements in DC microgrids that are controlled controllers in parallel DC/DC converters in a DC microgrid
based on a distributed control strategy. In [8], a strategy is based on a droop control method that consists of n paral-
lel converters. The voltage references for the converters are
Manuscript received May 8, 2020; revised June 23, 2020; accepted adjusted as follows:
July 6, 2020. Date of publication July 22, 2020; date of current version
January 29, 2021. This brief was recommended by Associate Editor J. Wu.
(Corresponding author: Mohammad Reza Habibi.)
Vref _p = (Vref + V)J1,n − IC × RD . (1)
Mohammad Reza Habibi and Frede Blaabjerg are with the Department of
Energy Technology, Aalborg University, 9220 Aalborg East, Denmark (e-mail: Here, J1,n is an 1 × n matrix of ones and V is implemented
[email protected]). to restore the output voltage [2], [3] and it is determined by
Hamid Reza Baghaee is with the Department of Electrical Engineering, the secondary controller. In addition, Vref _p , IC and RD can be
Amirkabir University of Technology, Tehran 15875-4413, Iran (e-mail:
[email protected]).
defined as follows:
Tomislav Dragičević is with the Department of Electrical Engineering,  
Technical University of Denmark, 2800 Lyngby, Denmark. Vref _p = Vref1 Vref2 · · · Vrefn , (2)
 
Color versions of one or more of the figures in this article are available I C = i 1 i2 · · · i3 , (3)
online at https://ieeexplore.ieee.org.  
Digital Object Identifier 10.1109/TCSII.2020.3011324 RD = diag RD1 , RD2 , · · · , RDn . (4)
1549-7747 
c 2020 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: CLEMSON UNIVERSITY. Downloaded on May 22,2024 at 16:20:41 UTC from IEEE Xplore. Restrictions apply.
718 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 68, NO. 2, FEBRUARY 2021

strategy, the value of Vdc (t) is estimated by a neural network,

and the estimated value is used as a reference in the refer-
ence tracking application. The reference tracking application
consists of a neural network, a PI controller, and the output,
which is added to the gathered value of Vdc (t). This value is
replaced by Vdc a (t) in the case of an attack. Fig. 2 illustrates

the implementation of the proposed method. In Fig. 2, a neural

network is used to estimate the real value of Vdc (t). Also, the
PI controller is implemented to converge the input of the PI
controller to the output of the neural network. If the output of
the PI controller is named β(t), the input of the PI controller
is as follows:
θ (t) = Vdc (t) + α + β(t). (8)
Fig. 1. Control of n parallel DC/DC converters in a DC microgrid based on
secondary and primary control. In (8), θ (t) is the input of the PI controller and it should
converge to V̄dc (t) by the PI controller as follows:
lim θ (t) = lim Vdc (t) + α + β(t) = V̄dc (t). (9)
t→∞ t→∞
In addition, if the neural network operates properly, it can be
assumed that Vdc (t) = V̄dc (t) and as a result:
lim α + β(t) = 0, (10)
t→∞
therefore:
lim β(t) = −α. (11)
t→∞
Fig. 2. Implementation of the reference tracking application for removing In other words, the output of the PI controller is converging
the FDIA.
to −α to remove the false data from the system. Also, by
converging the β to a non-zero value, the existence of the
where, Vrefk and ik are the reference voltage and the output
FDIA can be detected. Based on the proposed attack mitigation
current of the kth converter, respectively. In addition, in Fig. 1,
strategy, the neural network as the estimator has an important
HMG , HV and HI are PI controllers.
role, and it should estimate Vdc (t) precisely with low error.
FDIAs try to inject the false data to the DC microgrid, and in
A feedforward structured neural network is implemented to
this brief, the attack is considered on Vdc . As a result, if α is the
estimate Vdc (t). The implemented neural network consists of
value of the false injected data, the implemented mathematical
one input layer, a hidden layer, and an output layer. The inputs
model of FDIAs is as follows:
of the neural network are considered the output current and
a
Vdc (t) = Vdc (t) + α. (5) the voltage of the DC/DC converters. Therefore, if the DC
a (t) is the sensor value after attack at t that goes to
microgrid has n converters, the input of the neural network is
In (5), Vdc a set that has 2 × n data. Also, to extract a proper dynamic
the secondary controller. The PI controller of the secondary behavior of the estimator, a history of input data is gathered
layer (HMG ) is implemented to converge the input of the PI to be used as the input to the neural network. For example,
controller to the reference value, which is Vdc∗ . In the case of
a (t), and as a
if the DC microgrid has 2 converters, the input of the neural
an attack, the input of the PI controller is Vdc network is as follows:
result:
∗
X(t) = [i1 (t − 1), . . . , i1 (t − D),
a
lim Vdc (t) = Vdc . (6)
t→∞ i2 (t − 1), . . . , i2 (t − D),
In addition, based on (5) and (6), if the value of α is v1 (t − 1), . . . , v1 (t − D),
considered constant, it can be concluded that: v2 (t − 1), . . . , v2 (t − D)], (12)
∗
lim Vdc (t) = Vdc − α. (7) where D is the input-memory order. Also, the hidden layer has
t→∞
one layer, and the output layer consists of one neuron as the
Based on (7), by adjusting the value of α by the attacker, output of the neural network. The estimated value of Vdc (t),
Vdc (t) can converge to a value out of the allowance bounds, which is the output of the neural network and named V̄dc (t),
and it can shut down the DC microgrid. is calculated by the neural network as follows:

III. P ROPOSED M ETHOD V̄dc (t) = f2 (f1 (X(t)W1 + b1 )W2 + b2 ). (13)

The proposed work is based on a reference tracking appli- In (13), b1 and b2 are the bias vectors of the hidden and output
cation to remove the α and use the correct value of Vdc (t) as layer, respectively. Furthermore, W1 , W2 , f1 , and f2 are the
the input to the secondary controllers. Based on the proposed hidden layer weight matrix, the output layer weight matrix, the

Authorized licensed use limited to: CLEMSON UNIVERSITY. Downloaded on May 22,2024 at 16:20:41 UTC from IEEE Xplore. Restrictions apply.
HABIBI et al.: FALSE DATA INJECTION CYBER-ATTACKS MITIGATION IN PARALLEL DC/DC CONVERTERS 719

activation function of the hidden layer and activation function

of the output layer, respectively. In this brief, for the online
implementation of the neural network, the neural network is
trained offline to obtain the optimized and proper values of
W1 , W2 , b1 and b2 to have a well-tuned neural network. It is
important to note that the main reason for the offline training
is that the operator can be sure that the gathered data for the
training are from non-attack conditions, and as a result, they
are true and real.

IV. S IMULATION R ESULTS

The proposed strategy is verified on a simulated DC
microgrid. The simulated DC microgrid is structured by two
DC/DC buck converters that connect two DC sources with a
value of 350 V to the load bus. In addition, the reference volt-
age is considered to be 125 V. The simulation parameters are Fig. 3. Case study 1 (without the implementation of the proposed attack
as follows: R1 = 0.5 , R2 = 0.6 , RD1 = RD2 = 4. Also, removing strategy): The values of a) a and V , and b) output currents of
Vdc dc
the PI controller of the reference tracking application is made converters.
by a proportional and also an integral term with a value of
P = 10, K I
KAM AM = 100, respectively. The effectiveness of the
proposed method is examined under 4 scenarios. The imple-
mented neural network has an input layer, one hidden layer
that is made by 10 neurons, and an output layer with one neu-
ron. It is important to note that the number of neurons of the
input layer is equal to 2×n, which is equal to 4 for Case study
1, Case study 2, and Case study 3. Also, in Case study 4, the
number of converters is increased to 6 to have a larger system
and in this case, the number of neurons in the input layer is
12. To train the neural network, simulations were run to gather
data. The duration of the simulation for the training was 7s,
and the sampling time was considered to be 10μs, and as a
result, 7 × 105 samples of inputs were gathered. Also, for the
training, several load changing was considered to have more
dynamic data sets. Furthermore, the input-memory order (D)
is considered to be 2.
Fig. 4. Case study 1 (with the implementation of the proposed attack
a and V , and b) output currents of
removing strategy): The values of a) Vdc
A. Case Study 1: Load Change and FDIA dc
converters.
In this part, one scenario is simulated without the imple-
mentation of the proposed attack mitigation strategy to show shows the output current of the converters, and in the case
the effect of the attack and after that, the simulated scenario is of injecting the false data, the output currents do not change
simulated again while the proposed reference tracking appli- and just a little disturbance which is removed after a transient
cation is used. In the proposed scenario, at t = 0.5s a load is time and it can be seen at t=1.5 s when the attack starts in
added to the DC microgrid and at t = 1.5s a false data with the system. It is important to note that the change of current
the value of α = −25 is injected to the system and finally at after t=2.5 s is because of the load change. Furthermore, Fig. 5
t = 2.5s a load is added to the DC microgrid and the value of shows the output of the neural network that is V̄dc (t) and the
the false data is changed to 25, simultaneously. Fig. 3(a) shows real voltage of the load bus, which is Vdc (t). As can be seen
the Vdca (t) and V (t) for the tested scenario without the imple-
dc from Fig. 5, the value of V̄dc (t) and Vdc (t) are so close to each
mentation of the proposed mitigation method. As it can be seen other and it proves the effectiveness of the neural network to
in Fig. 3(a), the value of the Vdc a (t) is converged to the refer-
estimate Vdc (t).
ence voltage in the case of the attack. Furthermore, Fig. 3(b)
shows the output DC current of both converters. Based on
Fig. 3(b), the behavior of the output currents of the converters B. Case Study 2: Time-Varying FDIA
is seen as a load changing during the planned attacks. Fig. 4(a) In this tested case study, the injected false data is considered
a (t) and V (t) while the attack miti-
illustrates the value of Vdc as a sine wave with an amplitude of 10 and frequency 1 Hz,
dc
gation strategy is used in the DC microgrid. The real value of and the attack starts to inject the false data at t = 0.5s. The
the voltage of the load bus, which is Vdc (t), is converging to values of the injected false data (α) and also estimated Fig. 6
the reference voltage, even when it is under the attack and the shows false data (-β). Based on Fig. 6, the estimated value and
false data is removed from the DC microgrid. Also, Fig. 4(b) the real value of the false data match each other. Also, Fig. 7

Authorized licensed use limited to: CLEMSON UNIVERSITY. Downloaded on May 22,2024 at 16:20:41 UTC from IEEE Xplore. Restrictions apply.
720 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—II: EXPRESS BRIEFS, VOL. 68, NO. 2, FEBRUARY 2021

Fig. 5. The outputs of the neural network and the real value of the load Fig. 9. The output of the neural network-based estimator and the real value
bus voltage during Case study 1 and with the implementation of the proposed of the load bus voltage during Case study 2 while the DC microgrid is under
attack removing strategy. a time-varying FDIA.

Fig. 6. The real value of the false injected data and the estimated value of Fig. 10. The real value of the injected false data (α) and the estimation of
that by the neural network estimator in Case study 2. the false data (−β) that is determined by the neural network based application
in Case study 3.

Fig. 7. The value of the monitored dc voltage of the load bus (Vdc a ) during
the time-varying attack and also the real value of the dc voltage of the load
bus (Vdc ) with the mitigation strategy and also without that in Case study 2.

Fig. 8. The current outputs of the converters in Case study 2.

a
Fig. 11. Case study 3 (with a time delay of 2 ms.): The values of a) Vdc
and Vdc , and b) output currents of converters.
a and also V during this case study
illustrates the values of Vdc dc
with the mitigation method and also without the implementa-
tion of the proposed method. As can be seen from Fig. 7, a time delay that is 2 ms, and it is the same for both of the
without the implementation of the proposed reference track- units. In this case, the false data with a value of 40 is injected
ing application, the domain of the voltage is oscillating with a into the DC microgrid at t = 0.5s. Fig. 10 illustrates the real
high domain that is closed to the domain of the sine wave false and estimated value of the false data, which is determined by
data and the DC voltage seem as an ac voltage with a non-zero the neural network-based application. As shown in Fig. 10,
average value. Furthermore, Fig. 8 is related to output currents bypassing the time, the estimator application can estimate the
of converters. Based on Fig. 8, by the implementation of the value of the false data precisely though after a short transient
proposed attack removing strategy, the oscillation is removed time. Fig. 11(a) shows the attacked and real value of the DC
from the current waves. Besides, Fig. 9 plots the real value of bus voltage. Based on Fig. 11(a), the effect of the FDIA is
the DC bus and the estimated value by the neural network. As removed after a transient time, and the real value of the bus
can be shown in Fig. 9, the neural network-based estimator voltage is converged to the reference voltage. Also, Fig. 11(b)
can work properly even in the case of a time-varying attack. illustrates the output DC current of converters. Furthermore,
Fig. 12 shows the real value of the bus voltage and the esti-
mated value of that, which is calculated by the neural network.
C. Case Study 3: Delay and FDIA It can be seen in Fig. 12 that the estimator operates properly,
In this case study, a time delay is considered for the con- and it can determine the DC bus voltage precisely even in the
trol signal, which is sent by the secondary controller to the case of a time delay. It is important to note that the dc bus volt-
primary control layer. Here, it is considered that the output of age has an oscillation with a very low domain that is caused
the secondary controller is sent to the primary controllers by by the time delay effect. As can be observed from Fig. 12, the

Authorized licensed use limited to: CLEMSON UNIVERSITY. Downloaded on May 22,2024 at 16:20:41 UTC from IEEE Xplore. Restrictions apply.
HABIBI et al.: FALSE DATA INJECTION CYBER-ATTACKS MITIGATION IN PARALLEL DC/DC CONVERTERS 721

Fig. 12. The outputs of the neural network and the real value of the load Fig. 15. The real value of the injected false data (α) and the estimation of
bus voltage in Case study 3. the false data (−β) that is determined by the neural network based application
in Case study 4.

the false injected data. Based on Fig. 15, it can be concluded

that the proposed method can estimate the value of false data.

V. C ONCLUSION ( AND F UTURE W ORK )

In this brief, a method to mitigate the FDIA as a cyber-attack
in the DC microgrids, which is made by the parallel DC/DC
converters was proposed. The proposed method is based on
a reference tracking application, and it consists of a neural
network, which is implemented to produce a reference for a
PI controller. The output of the PI controller mitigates the
false injected data with success. This brief mostly focused
on FDIAs with a constant value of the false data. However,
in a scenario, the proposed method was examined under a
time-varying attack. It is important to note that time-varying
false data can affect the system, and the effect of that can be
a and V , and b) output currents considered in more depth in the DC microgrids in the future
Fig. 13. Case study 4: The values of a) Vdc dc
of converters. work.

R EFERENCES
[1] P. Prabhakaran, Y. Goyal, and V. Agarwal, “A novel communication-
based average voltage regulation scheme for a droop controlled DC
microgrid,” IEEE Trans. Smart Grid, vol. 10, no. 2, pp. 1250–1258,
Mar. 2019.
[2] T. Dragičević, X. Lu, J. C. Vasquez, and J. M. Guerrero, “DC
microgrids—Part I: A review of control strategies and stabilization tech-
niques,” IEEE Trans. Power Electron., vol. 31, no. 7, pp. 4876–4891,
Fig. 14. The outputs of the neural network and the real value of the load Jul. 2016.
bus voltage during Case study 4 and with the implementation of the proposed [3] L. Meng, T. Dragicevic, J. C. Vasquez, and J. M. Guerrero, “Tertiary and
attack removing strategy. secondary control levels for efficiency optimization and system damping
in droop controlled DC–DC converters,” IEEE Trans. Smart Grid, vol. 6,
no. 6, pp. 2615–2626, Nov. 2015.
[4] H. Wang, J. Ruan, G. Cao, Z. Ma, B. Zhou, and X. Fu, “Deep learning
domain of the oscillation is about 0.4 V, and considering that aided interval state prediction for improving cyber security in energy
the reference voltage bus is 125 V, the domain of oscillation Internet,” Energy, vol. 99, pp. 540–554, May 2019.
is about 0.32%, and it can be neglected. [5] Y. Mo and B. Sinopoli, “Secure control against replay attacks,” in
Proc. 47th Annu. Allerton Conf. Commun. Control Comput. (Allerton),
Sep. 2009, pp. 911–918.
[6] S. Sahoo, J. C. Peng, D. Annavaram, S. Mishra, and T. Dragicevic,
D. Case Study 4: More Units and FDIA “On detection of false data in cooperative DC microgrids—A discor-
In this scenario, the number of parallel converters is dant element approach,” IEEE Trans. Ind. Electron., vol. 67, no. 8,
increased to 6 to have a larger DC microgrid. Also, to evaluate pp. 6562–6571, Aug. 2020.
[7] S. Sahoo, S. Mishra, J. C. Peng, and T. Dragičević, “A stealth cyber-
the performance of the proposed method under load remov- attack detection strategy for DC microgrids,” IEEE Trans. Power
ing situation, a load is removed after the injection of the false Electron., vol. 34, no. 8, pp. 8162–8174, Aug. 2019.
data. Firstly, at t = 1s a false data with the value of α = 80 [8] M. R. Habibi, H. R. Baghaee, T. Dragicevic, and F. Blaabjerg, “Detection
of false data injection cyber-attacks in DC microgrids based on recurrent
is injected to the DC microgrid and after that, at t = 2s, a neural networks,” IEEE J. Emerg. Sel. Topics Power Electron., early
load is removed from the DC microgrid. Fig. 13(a) shows the access, Jan. 20, 2020, doi: 10.1109/JESTPE.2020.2968243.
DC bus voltage of the DC microgrid and Fig. 13(b) illustrates [9] O. A. Beg, L. V. Nguyen, T. T. Johnson, and A. Davoudi, “Signal tem-
poral logic-based attack detection in DC microgrids,” IEEE Trans. Smart
the currents of the converters. As can be seen from Fig. 13(a), Grid, vol. 10, no. 4, pp. 3585–3595, Jul. 2019.
after the injection of the false data, the DC bus voltage is [10] S. Abhinav, H. Modares, F. L. Lewis, and A. Davoudi, “Resilient coop-
restored to the reference value, successfully. Also, the real erative control of DC microgrids,” IEEE Trans. Smart Grid, vol. 10,
and estimated values of the DC bus voltage are depicted by no. 1, pp. 1083–1085, Jan. 2019.
[11] O. A. Beg, T. T. Johnson, and A. Davoudi, “Detection of false-data
Fig. 14, and as it can be seen, they are fitted to each other. injection attacks in cyber-physical DC microgrids,” IEEE Trans. Ind.
Furthermore, Fig. 15 shows the real and estimated values of Informat., vol. 13, no. 5, pp. 2693–2703, Oct. 2017.

Authorized licensed use limited to: CLEMSON UNIVERSITY. Downloaded on May 22,2024 at 16:20:41 UTC from IEEE Xplore. Restrictions apply.