Page 65 - Read Online

Yang et al. Intell Robot 2024;4(4):406-21 I http://dx.doi.org/10.20517/ir.2024.24 Page 408

require labeled data, they are generally incapable of precise classification or prediction and are commonly used
to detect distributed denial of service (DDoS) attacks [23] . Semi-supervised learning models can be effectively
trained with a small amount of labeled data and a large amount of unlabeled data; however, these models are
oftencomplexinstructure, accompaniedbyhighcomputationalcostsandunstabletrainingprocesses. Against
this backdrop, the vector neural network (VNN) has emerged as an advanced model offering greater flexibility.
VNNs possess strong adaptability and can be tailored to various supervised learning modes based on actual
application needs. In wide-area power grids, VNNs have demonstrated the capability to detect attacks [24] ,
thereby enhancing the security and stability of the system.

Therefore, this paper analyzes the application of µPMUs in park-level energy internet systems, the principles
of TSAs, and their impact on grid monitoring. To address the TSA issue, a VNN-based detection model
(VNN-DM) is proposed. The model incorporates the design concepts of both VNNs and convolutional neural
networks (CNNs), enabling the analysis of features in data collected by µPMUs, and achieving detection of
TSA data in smart grids. VNNs offer considerable flexibility, adapting to various supervised learning modes
based on actual usage scenarios, and possess the ability to recognize complex patterns and attack behaviors.
In complex attack scenarios, they exhibit high accuracy and robustness. The contributions of this paper are as
follows:

1. Clarification of the characteristics of µPMU measurement data under TSA: This paper conducts an indepth
study of the impact of TSAs on µPMU data, revealing how attacks manipulate time signals to cause devi-
ations in power grid state estimation. This provides a theoretical foundation for the design of effective
detection methods.
2. Proposal of a TSA detection model based on VNNs (VNN-DM): Building on the VNN architecture, this
paper introduces convolutional layers for multi-scale temporal feature extraction and enhances the feature
classification capability of capsule networks by optimizing the dynamic routing mechanism. The VNN-
DM model achieves precise analysis of power grid states and attack detection in response to TSA issues in
park-level energy internet, offering strong detection accuracy and robustness.

Theremainderofthepaperisstructuredasfollows: Chapter2explorestheapplicationofµPMUinpowergrids
and the threat of TSAs to grid security. Chapter 3 introduces the improved VNN-DM model for detecting
TSAs in µPMU measurement data within park-level energy internet. Chapter 3 validates the performance
of the VNN-DM model through simulation experiments, demonstrating its advantages in detection accuracy
and robustness. Finally, Chapter 4 concludes the research findings and proposes directions for future work.

2. TSA IN PARK-LEVEL ENERGY INTERNET
In the park-level energy internet, µPMUs, with their high sampling rate and fine measurement accuracy, have
become key components in power system state estimation. The measurement data from µPMUs is transmitted
through the power communication network to the control center, where it is analyzed to estimate the current
stateofthepowersystem, detectpotentialfaults, andsendcontrolsignalsbacktoremoteterminalunits(RTUs)
to execute corresponding operations. This ensures the reliability of the system, as illustrated in Figure 1. The
accuracy of µPMU data is crucial for the precision of state estimation and forms the foundation for the stable
operation of power systems.

However, TSA poses a real threat to the accuracy of µPMU data. As shown in Figure 1, TSA attackers can send
GPSspoofingsignalsaroundthetargetµPMUwithoutneedingtoinfiltratethemonitoringsystemorphysically
access the µPMU. By tampering with the time synchronization signals, TSA causes measurement data with
incorrect timestamps to be transmitted to the control center, leading to errors in the system state estimation.
These erroneous data may bypass simple algorithmic detection and affect the operational scheduling of the
power system. While some data processing programs can handle measurement errors, most of them only