A survey of machine unlearning for electric power systems: From privacy compliance to resilient grid operations

The increasing reliance on machine learning (ML) in electric power systems for forecasting, control, and security raises challenges in data governance, privacy, and resilience. Machine unlearning (MU), the selective erasure of training data influence without full retraining, offers a promising solution but remains largely unexplored in the power sector. We bridge this gap with a domain-focused, comprehensive survey that translates MU advances from healthcare, finance, social media, and IoT into actionable guidance for grid operators, addressing constraints in power system data streams, regulatory frameworks, and operational requirements not examined in prior MU literature. After distilling a taxonomy of exact and approximate unlearning techniques, verification metrics, and differential privacy extensions, we critically analyze trade-offs and failure modes, including the demonstrated inability of approximate methods to remove data poisoning, and examine MU challenges specific to large language models increasingly deployed for grid intelligence. A critical appraisal of cross-domain transferability reveals barriers stemming from latency requirements, failure consequences, and regulatory mismatches. We map MU to four power system use cases: GDPR/CCPA-compliant smart meter erasure, removal of poisoned SCADA/PMU data, federated unlearning across utilities, and real-time model adaptation, supported by operational scenarios, concrete specifications, and quantified performance trade-offs. To bridge theory and practice, we identify algorithmic families for online MU, formalize speed–stability trade-offs, define a tiered certification framework with compliance workflows, and specify a blockchain-based audit architecture. The intended outcome is to convert MU from a theoretical safeguard into a verifiable, standards-ready tool that strengthens privacy, cybersecurity, and operational resilience in next-generation grids.