Page 64 - Read Online

Page 407 Yang et al. Intell Robot 2024;4(4):406-21 I http://dx.doi.org/10.20517/ir.2024.24

1. INTRODUCTION
The rapid development of the energy internet in industrial parks is transforming the operational model of tra-
ditional power systems. State estimation, as a key link in ensuring the stability and security of power systems,
[1]
has emerged as a key research focus . The application of phasor measurement units (PMUs) in state estima-
tion has significantly enhanced their efficiency and accuracy. PMUs are widely used in smart grids to enhance
[2]
or replace traditional sensors in supervisory control and data acquisition (SCADA) systems . Compared
to traditional sensors, PMUs have higher sampling rates, significantly enhancing the real-time performance
[3]
of wide-area measurement systems . Micro PMUs (µPMUs) offer even higher sampling rates, finer mea-
surement accuracy, and more flexible deployment, demonstrating distinct advantages in broader power grid
[4]
monitoring applications . However, accurate time synchronization is critical for µPMUs to capture precise
state data from the power grid.

Accurate time synchronization serves as a cornerstone for effective power grid monitoring, enabling precise
[5]
coordination and data reliability . It ensures that each µPMU device can collect data under the same time
reference, allowing precise monitoring of the state of the grid. µPMUs can accurately capture the voltage and
current phasor information of the grid, providing real-time monitoring of grid operations, thus enhancing the
stability and reliability of the power system [6,7] . However, with the continued advancement of information
[8]
technology, grid time synchronization technologies face new challenges , one of which is the time synchro-
nization attack (TSA).

TSA is a malicious behavior that disrupts time synchronization in the power grid by tampering with time sig-
nals. This type of attack exploits system vulnerabilities, using techniques such as injecting false time signals
or interfering with the transmission of time signals [9,10] , causing µPMUs to receive incorrect time signals. To
effectively mitigate these threats, the implementation of intelligent detection technologies is essential. Recent
advances in robust consensus mechanisms have shown that resilient consensus can be achieved in multihop
communication with path-dependent delays, even within random dynamic networks under mobile malicious
attacks [11,12] . These studies provide essential theoretical support for building robust frameworks capable of
detecting TSA in complex networked environments. Such technologies employ sophisticated algorithms to
continuously monitor synchronization data in real time, facilitating automatic identification of anomalies in-
dicative of TSA [13] . This proactive approach facilitates early warnings, enhancing system resilience and robust-
ness against potential attacks. As a result, µPMUs can report incorrect voltage or current phasor data, lead-
ing to misjudgments about the state of the grid at the control center, and potentially resulting in operational
scheduling errors. Moreover, during long-term grid operation, repeated errors may lead to the accumulation
of synchronization data discrepancies [14] , potentially triggering more severe grid safety incidents. Detecting
and preventing TSAs and ensuring the authenticity of grid monitoring data have become urgent issues that
need to be addressed.

Scholarsfromvariousregionshaveconducted relevantresearchonattackdetection. Existingdetectionmodels
can be categorized into fully supervised learning models [15–17] , semi-supervised learning models [18–20] , and
unsupervised learning models [21,22] based on their learning approaches. Fully supervised learning models rely
onlargeamountsoflabeleddata, usingknownattacksandnormalbehaviortotrainthemodelforclassification.
Semi-supervised learning models combine a small amount of labeled data with a large amount of unlabeled
data, improving detection performance through pseudo-labeling or other techniques. Unsupervised learning
models, on the other hand, do not require labeled data; they rely on the intrinsic structure or patterns in the
data to identify potential anomalies and attack behaviors.

While fully supervised models are dependent on a substantial quantity of high-quality labeled data, in the
context of TSA issues within park-level energy internet environments involving µPMU measurement data,
it is often difficult to obtain a large volume of labeled data. Although unsupervised learning models do not