To enable fast and accurate analysis for modern power systems, Machine Learning (ML) models have achieved great success with the emerging synchrophasor measurements. However, it is hard to train an ML model directly for data-limited scenarios like a newly constructed power system. Thus, we propose a transfer learning framework to transfer knowledge from a source grid to a target grid. We focus on event identification and study how knowledge can be efficiently transferred with or without event labels. While existing work mainly focuses on regular data like image datasets, knowledge transfer for power systems is challenging since the size and distribution of the two systems are different. Thus, we propose methods to align the dimensionality and maintain close data distributions in the latent space. For generality, we propose transfer learning to both supervised and unsupervised approaches. For supervised learning methods, we develop efficient data processing and common knowledge voting scheme to enable efficient transfer. For unsupervised learning methods, we introduce dimensionality-reduction and clustering algorithms to find common features with small distribution distances. Finally, we introduce proper theorems to validate our model and conduct both supervised and unsupervised transfer learning over diversified power system datasets.
