7
Fig.3. ExamplesofIoTapplications.
drives,toaddresstheseconcernsandenhancethereliabilityof in a smart grid. In [176], an artificial intelligence-based
intelligent power electronics systems. For IoT enabled power method is proposed for detection of DIAs in DC microgrids.
electronicssystems,wefocusontwosub-applications:power- By exploiting the nonlinear mapping capability of nonlinear
electronics-based smart grid and intelligent motor drives. autoregressive exogenous (NARX) neural networks, cyber-
Power electronics converters play a crucial role in inte- attacks in DC microgrids could be identified. Although a
grating distributed energy resources (DERs) with the power supervised learning method could generate a good estimation
grid. The increasing use of industrial Internet of Things (IoT) or classification result to distinguish cyber-physical threats, it
edge devices has led to the incorporation of various built- needs a large amount of data from the plant, either simulated
in functionalities in power converters, such as remote control or actual, during the training phase. However, it is impossible
and wireless communication with a central plant controller. to emulate all of the potential cyber-physical threats in real
Recent literature highlights the use of such functionalities, in- applications. Specifically for a PV system, [177] proposed a
cludingconverterremotecontrolandwirelesscommunication, multilayer LSTM-based diagnosis solution for DIAs in a two-
to enhance power converter operation [166], [167], [168]. stagePVconverter.Byusingthecollectedelectricalwaveform,
However,thelatestrevisionstoIEEE1547standardmandatea a supervised-learning-based classifier is trained for cyber-
setofcontrolparametersforgrid-tiedDERconvertersthatre- attackdetection.In[178],anattemptismadetodetecttypical
quireremotecontrolandadjustmentbyaSupervisoryControl cyber-attacks in PV systems by using micro phasor measure-
and Data Acquisition (SCADA) system via a communication ment units (µPMU) data. Notice that the above two studies
network [169], [170], [171]. With the increasing number of donotconsiderphysicalfaultsbecauseelectricwaveformsand
DER equipment connected to the IoT infrastructure, power µPMUundercyber-attacksandphysicalfaultsshowverysim-
electronics based smart grids are more susceptible to cyber ilarpatterns.Toovercomethedrawbackofsupervisedlearning
attacks. methods, [179] proposed a binary matrix factorization-based
cyber-attack diagnosis without a training process. The results
Detection of cyber-attacks in smart grids has become an
have shown that most cyber-attacks are clustered at different
essential topic because more internet technologies are em-
locations.Physics-informedneuralnetworksofferapromising
ployedincommunicationbetweendistributedenergyresources
avenue for addressing these concerns. This is achieved by
and central locations [166], [167]. Smart grid cyber-physical
seamlessly integrating principles rooted in physics (referred
security studies have proposed many cyber-attack detection
to as physics-informed) into the cutting-edge framework of
and diagnosis methods using AI techniques. In [172], RNN is
deep learning. These paradigms encompass diverse aspects,
usedforfalsedatainjectionattackdetectioninaDCmicrogrid
including the formulation of physics-informed loss functions,
with dynamic loads. In [173], a novel cooperative mecha-
the initialization of neural networks with physics-driven in-
nism based on a secondary voltage controller is proposed to
sights, the architectural design guided by physical principles,
facilitate the detection and mitigation of stealth attacks on
andtheamalgamationofphysics-derivenknowledgewithdeep
DC microgrid. In [174], a measurement data authentication
learningmodelsinahybridfashion[180].In[181],innovative
algorithmbasedonFastFourierTransform(FFT)andmachine
physics-guidedfeaturessuchasfrequency-domainmagnitude-
learning methods is developed for smart grids to protect
basedresidualstime-domainmeancurrentvector-basedfeature
against data spoofing attacks. In [175], a novel optimization
wereproposedtoaddressnovelcyber-attacksthatareexcluded
technique is developed to train an Artificial Neural Network
from the machine learning training process.
(ANN) to classify and detect cyber-attacks and intrusions