Graduation Term

Spring 2026

Degree Name

Master of Science (MS)

Department

Department of Mathematics

Committee Chair

Mehdi Karimi

Committee Member

Maochao Xu

Committee Member

Xiaotian Dai

Abstract

The convergence behavior of distributed optimal power flow (OPF) depends strongly on how the power network is partitioned into regions. Classical graph-based methods such as METIS are widely used, but they rely mainly on static topological criteria and do not explicitly incorporate operating-point-dependent information that may affect distributed optimization performance. This thesis develops a data-driven partitioning framework for distributed OPF using graph neural networks (GNNs). Each OPF scenario is represented as a graph in which buses are nodes and transmission lines are edges. Node and edge features capture both structural and operational characteristics of the network. Partition prediction is formulated as a node classification problem. A two-stage learning strategy is proposed. In Stage 1, a GNN is pretrained to imitate METIS-generated partitions. In Stage 2, the pretrained model is refined using iteration-aware labels obtained from a local-search procedure guided by distributed OPF performance. The framework is evaluated on perturbed benchmark scenarios and selected standard MATPOWER systems. Results show that the pretrained GNN closely reproduces the METIS baseline and achieves modest improvements in several cases, particularly in iteration count and runtime. On selected standard systems, the refined Stage 2 GNN improves upon the pretrained model in several settings and is often competitive with METIS in convergence speed, while METIS frequently retains an advantage in final optimality GAP. On the large case1354pegase instance, the learned partitioners remain competitive in iteration count, although broader large-scale experiments reveal robustness limitations. Overall, the results demonstrate that GNN-based partitioning is a viable and competitive approach for distributed OPF.

Access Type

Thesis-Open Access

Download

Share

COinS