Structure-generalizable large models for power system operation under topology uncertainty: A graph-based foundation learning framework

This paper addresses a fundamental limitation of existing data-driven power system models, namely their inability to generalize under frequent topology changes. As modern distribution networks undergo dynamic reconfiguration due to switching operations and high penetration of distributed energy resources, models trained on fixed structures often exhibit significant performance degradation when applied to unseen configurations. To overcome this issue, we propose a structure-generalizable large model framework that reformulates power system operation as a topology-aware learning problem. The approach leverages graph-based topology encoding and multi-topology training to construct a unified latent representation space, enabling the model to capture invariant physical and operational patterns across diverse network structures. A structure-aware learning mechanism is further introduced to enhance cross-topology adaptability, while a confidence-aware prediction module quantifies uncertainty and supports reliable deployment under distribution shifts. Extensive experiments on benchmark distribution systems demonstrate that the proposed framework improves generalization accuracy by over 18.7%, reduces prediction error by 22.4%, and achieves 35.2% faster inference compared to conventional topology-specific models. These results indicate that the proposed large model paradigm provides a scalable, robust, and practical solution for real-time operation of future adaptive energy systems under inherent topology uncertainty.