.png){kind=icon}
.png){kind=icon}
Return
Domain Adaptive Representation Learning for Attack Detection in Smart Grids
.png){kind=icon}
PRE
AI
K
M
B
X
DOI：10.1109/TSG.2025.3642949.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The proliferation of time-synchronized measurements in smart grids has accelerated the development of machine learning-based data-driven attack detection methods, which have been widely adopted due to their powerful feature extraction capabilities. However, these methods typically assume that the data in the training and deployment phases share an identical distribution, which can be violated when the configurations or components of the deployment system change. Undoubtedly, such domain shift scenarios can harm the attack detection performance. In this work, we tackle this issue by performing domain adaptive and representation learning in one coherent framework termed Domain Adaptive Representation Learning (DARL). First, the nonlinear oscillation modes in sensor measurements during system transients are captured by Koopman mode decomposition. Second, DARL incorporates a hierarchical alignment strategy to reduce the cross-domain discrepancy, which is achieved by performing global distribution and instance-wise feature alignment. Then, DARL learns the target intrinsic structure and representation in a self-supervised manner to avoid over-reliance on source supervision. Extensive case studies on IEEE test systems demonstrate that DARL significantly outperforms baselines in various domain shift scenarios, achieving the smallest detection accuracy drops in target domains.
Keywords:
Smart grids
Training
Phasor measurement units
Transient analysis
Detectors
Feature extraction
Cyberattack
Vectors
System dynamics
Representation learning
Cyber-physical security
attack detection
domain adaptation
machine learning
smart grid
Journal
IF:
9.8
Papers:
5.6K
Citations:
4.3W
Organization
Cited Papers