Graph Learning Method for Voltage Area Partition in Inaccurate Multiphase Distribution Networks

Power distribution networks are rapidly evolving into large-scale, highly complex, and topologically variable active distribution networks (ADNs). In such ADNs, voltage issues are serious and thus the distributed or decentralized voltage control tends to be called for, whose effectiveness is critically dependent on the proper partitioning of the network into multiple sub-networks, a.k.a., voltage area partition (VAP). Existing partitioning methods are focused on three-phase balanced systems with known line parameters, making them less applicable to real-world distribution networks. To address this limitation, this paper proposes a data-driven VAP method, which can perceive voltage sensitivity from the historical operation data without need of precise distribution line parameters while simultaneously assigning buses into clusters by designing a physics-informed graph neural network architecture. In particular, a phase-completing layer is developed to convert a phase-missing bus to a three-phase bus, which can fairly evaluate the voltage cohesiveness of different buses. Simulation in the modified IEEE 123 bus system demonstrates that the proposed method yields partitioning results with higher voltage cohesiveness compared to benchmark methods.