Uncertainty-aware frequency stability prediction with label-free model adaptation to grid topology changes

Data-driven frequency stability prediction (FSP) methods generally lack uncertainty quantification (UQ) and require labeled samples when adapting to topological changes in power grids. To address these limitations, this paper proposes an FSP model based on evidential deep learning (EDL) and designs an unsupervised domain adaptation (UDA) framework, multi-scale feature alignment (MSFA), to enable label-free adaptation. Specifically, EDL outputs Dirichlet parameters in a single forward pass, and the FSP result and predictive uncertainty are then computed analytically from these parameters. MSFA integrates multi-kernel maximum mean discrepancy (MK-MMD) and cross-domain contrastive learning (CDCL): MK-MMD reduces the global distribution discrepancy of low-level features across domains, while CDCL enhances class discriminability by promoting similarity among high-level features within the same class across domains. Together, the two components enable MSFA to achieve synergistic alignment in terms of global distribution and class consistency. Case studies in the IEEE 39-bus system and Illinois 200-bus system show that the proposed FSP model maintains high prediction accuracy while effectively quantifying predictive uncertainty, significantly improving the reliability of FSP results. Moreover, for unseen topologies, MSFA enables the FSP model to be updated without labeling samples from those topologies. Ablation studies further confirm the necessity of the synergistic alignment strategy.