A Multi-stage Hybrid Active Transfer Learning Scheme for Power System Transient Stability Assessment Adaptive to the Topological Variability

A great challenge to machine learning-based power system transient stability assessment (TSA) is the adaptability of a data-driven model to topological variations of the grid. To quickly improve the model performance in a new system, this paper proposes a multi-stage hybrid active transfer learning (MSH-ATL) scheme, requiring fewer new labeled samples to update the model. MSH-ATL first aligns the distributions between the source and target domain in the model's hidden space to learn an initial target model without any new labeled sample. Afterward, the supplementary samples in the source domain are selected and combined with incremental new labeled samples to train the target model. Meanwhile, diversity and uncertainty-based active learning strategies continuously query labels of valuable samples in the new scenario and, thus, significantly reduce the computational cost of labeling. Experiments on the IEEE 39-bus system show the superiority of our method and scheme.