This project introduces a semi-supervised learning framework for classifying disturbance events using high-dimensional PMU data when labels are sparse. Real power-system PMU archives contain thousands of unlabeled snapshots and only a limited number of trusted labeled events.
▣ Problem
- PMU archives contain very limited labeled data, making supervised training unreliable
- Grid conditions shift across time, seasons, and topology changes
- Physics-based features alone cannot fully capture dynamic variability
▣ Approach
- Combine spectral/modal signatures with learned temporal embeddings
- Apply self-training, pseudo-label refinement, and consistency constraints
- Leverage small labeled sets to guide structure in large unlabeled PMU corpora
▣ Validation
- Strong generalization under low-label regimes
- Robust identification across multiple utilities and operating conditions
- Outperforms traditional supervised PMU classifiers in label-scarce settings
▣ Publication
arXiv preprint (2024):
https://arxiv.org/abs/2309.10095