Graph Learning for Power Flow Analysis: A Global-Receptive Graph Iteration Network Method

• Published in: IEEE transactions on network science and engineering Vol. 12; no. 2; pp. 599 - 609
• Main Authors: Huang, Junyan, Li, Yuanzheng, He, Shangyang, Hao, Guokai, Zhou, Chunjie, Zeng, Zhigang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.03.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Computer architecture; Convolution; Data-driven approach; graph iteration network; Graph neural networks; Image segmentation; Learning; Metalearning; Newton method; Newton methods; Power flow; power system; Power system stability; Power systems; Topology; Training
• ISSN: 2327-4697; 2334-329X
• DOI: 10.1109/TNSE.2024.3506012

The data-driven methods based on the graph convolution architecture provide a promising direction for accelerating power flow (PF) calculation. These methods directly predict operational states of power systems according to given conditions, such as loads, states of buses, topology, etc. However, we find that the neighborhood aggregation of the graph convolution architecture violates operational constraints of power systems. In this paper, a global-receptive graph iteration architecture that overcomes this shortcoming is designed to replace the graph convolution architecture. Specifically, Newton's method, one of the most classical methods for PF, is embedded into the graph iteration network (GIN) to form an implicit residual learning architecture. To retain the interpretability, the GIN follows a non-activation paradigm, in which the ability of non-linear representation stems from the iterative architecture rather than the activation function. Finally, without the demand to reclaim global information, the GIN allows shallower network structure by eliminating fully connected layers. Extensive numerical experiments are conducted on IEEE 30-bus, 57-bus, 118-bus, and 300-bus power systems. The results validate the higher computational efficiency and the better prediction performance of the proposed method, compared with both classical approaches and precedent data-driven approaches.