검색 상세

전력계통 특화 물리정보 주입 그래프 파운데이션 모델 연구

Physics-Informed Graph Foundation Model for Power Systems

  • 송근주 (서강대학교 일반대학원)

초록(요약문)

With the increasing integration of distributed energy resources （DER） and data centers, modern power systems are becoming increasingly complex and uncertain, making their operation more challenging. In particular, alternating current optimal power flow （AC-OPF） has been a challenging problem for decades due to its nonconvexity, but fast and efficient solutions are even more needed because of high penetration of large scale renewable generation and load growth. Recently, neural networks （NN） have gained attention in solving AC-OPF and DER forecasting, but it is still in an early stage to be applicable for real and large-scale power system operation with topology-changing characteristics. To overcome these challenges, this dissertation proposes graph neural network-based methods for renewable energy forecasting and AC-OPF. First, this dissertation proposes a scalable and missing-insensitive framework for probabilistic multi-site renewable energy forecasting called AnyCast, specifically focused on large-scale renewable energy sites and space-time missing data. The proposed scalable graph learning mechanism with random coarse graph attention and probabilistic spatio-temporal learning performs efficiently for large-scale renewable energy sites in terms of forecasting accuracy and model training complexity. At the same time, the proposed framework adaptively imputes the missing energy data in the space and time domain, respectively. Second, this dissertation proposes a GraphOPF, a graph neural network framework for solving AC-OPF in large-scale power systems. GraphOPF consists of edge-aided graph neural networks to combat against topology changes and a hard-constrained embedded layer for satisfying power flow equations. By doing this, the proposed framework jointly considers topology adaptability, scalability, NN training time, self-supervision, and feasibility altogether. In particular, GraphOPF exhibits high sample efficiency, enabling fast NN training with a small amount of training data, e.g., one tenth samples only. Furthermore, graph transfer learning is shown to further accelerate AC-OPF computation under topology changes, allowing NN training to be completed within 1 minute even for the Korea national grid （the fifth largest power system in the world）. Finally, this dissertation extends the proposed GraphOPF into a FedOPF-FM, a graph foundation model for AC-OPF designed to address diverse operating scenarios, including load variability and N − 1 contingencies. FedOPF-FM leverages hierarchical federated learning to account for the heterogeneous nature of power system data, data privacy, and improved generalization capability. The resulting foundation model demonstrates fast adaptability through fine-tuning and zero-shot transfer, enabling flexible and efficient derivation of AC-OPF solutions across various network topologies and operating conditions. Overall, the proposed AnyCast, GraphOPF, and FedOPF-FM enable fast and reliable solutions under diverse conditions in large-scale power systems. These contributions pave the way for the practical adoption of artificial intelligence in future power system operations.

more

목차

I Introduction 1
1.1 Graph-based Large-Scale Renewable Energy Forecasting 3
1.2 Physics-Informed Graph Learning for Large-Scale AC Opti-mal Power Flow 6
1.3 Federated Graph Foundation Model for Optimal Power Flow 9
1.4 Contributions 10
II Graph-based Large-Scale Renewable Energy Forecasting 13
2.1 Proposed Methodologies 13
2.2 Experimental Settings 25
2.3 Experimental Results 32
III . Physics-Informed Graph Learning for Large-Scale AC Optimal Power Flow 45
3.1 Problem Formulation 45
3.2 Proposed Methodologies 46
3.3 Experimental Settings 58
3.4 Experimental Results 62
IV Federated Graph Foundation Model for Optimal Power Flow 75
4.1 Proposed Methodologies 75
4.2 Experimental Settings 80
4.3 Experimental Results 82
V Conclusion 87
Reference 91

more
반출 Meta View 목록