GNNs' Generalization Improvement for Large-Scale Power System Analysis Based on Physics-Informed Self-Supervised Pre-Training
Yuhong Zhu, Yongzhi Zhou, Wei Wei, Peng Li, Wenqi Huang
Abstract
Efficient and informative representation of system topologies is critical in AI-driven power system analysis (PSA). Despite a major breakthrough, recent approaches employing Graph Neural Networks (GNNs) face significant challenges in large-scale PSA, including high computational demands for sufficient labeled data and poor generalization to unseen disturbed topologies. To tackle these issues, we propose a self-supervised strategy for pre-training GNNs that enhances their expressiveness at both the individual node feature level and the whole graph structure. Integrating physics-informed techniques, our strategy allows GNNs to internalize fundamental principles applicable to multiple downstream tasks. We demonstrate that our method enables the efficient training of GNNs on extensive topology datasets without supervision, effectively addressing the noted challenges. By pre-training GNNs with 145 million parameters on 20 million unlabeled topologies and subsequently fine-tuning them, we observe a significant performance improvement, averaging over 13%, compared to existing state-of-the-art (SOTA) methods across four challenging tasks.