Transferable Graph Learning for Transmission Congestion Management via Busbar Splitting
Ali Rajaei, Peter Palensky, Jochen L. Cremer · 2025
Read Transferable Graph Learning for Transmission Congestion Management via Busbar Splitting by Ali Rajaei, Peter Palensky, Jochen L. Cremer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-20591.txt.
Introduction
Ali Rajaei, Peter Palensky, and Jochen L. Cremer use graph learning to speed up transmission congestion management through busbar splitting, a topology change that can cut redispatch costs. Solving the underlying mixed-integer nonlinear problem for large grids in near-real time is beyond current solvers, and earlier machine learning approaches generalize poorly to unseen topologies and systems. Their formulation uses linearized AC power flow.
A heterogeneous, edge-aware message passing graph neural network predicts which nodes are effective candidates for busbar splitting actions. The network captures local flow patterns and transfers across systems and topology changes. Case studies show speedups of up to four orders of magnitude, delivering AC-feasible solutions within one minute with a small optimality gap on a two-thousand-bus system.
Catalog record
| CREATOR | Ali Rajaei, Peter Palensky, Jochen L. Cremer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-20591
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Transferable Graph Learning for Transmission Congestion Management via Busbar Splitting Ali Rajaei, Peter Palensky, Jochen L. Cremer Abstract Network topology optimization (NTO) via busbar splitting can mitigate transmission grid congestion and reduce redispatch costs. However, solving this mixed-integer non-linear problem for large-scale systems in near-real-time is currently intractable with existing solvers. Machine learning (ML) approaches have emerged as a promising alternative, but they have limited generalization to unseen topologies, varying operating conditions, and different systems, which limits their practical applicability. This paper formulates NTO for congestion management problem considering linearized AC PF, and proposes a graph neural network (GNN)-accelerated approach. We develop a heterogeneous edge-aware message passing NN to predict effective busbar splitting actions as candidate NTO solutions. The proposed GNN captures local flow patterns, achieves generalization to unseen topology changes, and improves transferability across systems. Case studies show up to 10 4 10^{4} speed-up, delivering AC-feasible solutions within one minute and a 2.3% optimality gap on the GOC 2000-bus system. These results demonstrate a significant step toward near-real-time NTO for large-scale systems with topology and cross-system generalization. Nomenclature Indices and Sets i\in\mathcal{V} Index for substations (nodes). b\in\mathcal{B} Index for busbars \mathcal{B}=\{b1,b2\} . l_{ij}\in\mathcal{L} Index for transmission lines (from i to j ). g\in\mathcal{G} Index for generators. d\in\mathcal{D} Index for load demands. \mathcal{L}_{i},\mathcal{G}_{i},\mathcal{D}_{i} Elements e\in E_{i} connected to substation i . Parameters \hat{P}_{d},\hat{Q}_{d} Active and reactive power demand. \hat{P}_{g},\hat{Q}_{g} Active and reactive power generation. \underline{V}_{i},\overline{V}_{i} Minimum and maximum squared voltage magnitude. \overline{Z}^{s} Maximum number of busbar splitting. b_{l},g_{l} Susceptance and conductance of line l . \overline{S}_{l} Maximum apparent power flow of line l . \phi_{d} Tangent of the power factor angle of loads. \gamma Congestion cost parameter. s^{th} Line loading penalization threshold. M^{\theta},M^{v} Big-M parameters. Variables P_{g,b},Q_{g,b} Axillary variables for power generation. P_{d,b},Q_{d,b} Axillary variables for load demand. P_{l,i},Q_{l,i} Active and reactive power flow of line l from i side. P_{l,i,b},Q_{l,i,b} Active and reactive power flow of line l through busbar b at substation i . P^{L}_{l,i},Q^{L}_{l,i} Active and reactive power loss of line l . \theta_{i,b},V_{i,b} Voltage angle and squared voltage magnitude of substation i busbar b . \theta_{l,i},V_{l,i} Voltage angle and squared voltage magnitude at the end of line l on the side of i . z_{i} Connection status of coupler in substation i (0: open, 1: close). z_{g},z_{d} Connection status of generator and demand to busbars ( 0:b1,1:b2 ). z_{l,i} Connection status of line l to substation i busbars ( 0:b1,1:b2 ). \mathcal{O}^{CM} Congestion management objective. I Introduction Congestion management (CM) in transmission networks is becoming increasingly challenging due to the growing integration of distributed energy resources and electrification across sectors. NTO, particularly substation reconfiguration via busbar splitting, is a promising approach to relieve congestion and reduce re-dispatch costs. However, solving this mixed-integer non-linear optimization problem for large-scale systems in near-real-time remains computationally challenging. As a result, operators often rely on expert knowledge or predefined manuals, which can lead to sub-optimal system performance [ 1 ] . Recent approaches have explored ML to identify effective topological actions in near-real-time, but most models lack scalability, remain system-specific and fail to generalize to new topologies, operating conditions, or unseen systems. This lack of generalization and transferability limits their practical use due to: 1) the need for retraining and costly data collection for each new system, 2) difficulty for operators to detect when models become outdated, and 3) performance drops under rapidly changing grid conditions. In this context, this paper proposes a GNN-based approach that exploits the locality of CM to scale to large-scale systems, generalize to topology changes, and transfer to unseen grids. Optimal substation reconfiguration (OSR) extends the optimal power flow (OPF) problem by introducing binary variables for breaker statuses, allowing changes to the network topology. Several works [ 2 , 3 , 4 , 5 ] apply sensitivity analysis as a heuristic to identify candidate busbar splitting actions. These heuristic approaches assume a fixed substation configuration and focus solely on selecting splitting actions, but they do not provide an optimal substation configuration. [ 6 , 7 ] formulate DC-OSR as a MILP using node-breaker modeling and shift factors, respectively. [ 8 ] develops a heuristic approach to solve DC-OSR by focusing on local congested regions. Building on [ 6 ] , [ 9 ] propose a sparse tableau formulation of AC PF, while [ 10 ] develops a MISOCP formulation. However, these approaches either rely on heuristics, are limited to DC power flow, or become computationally intractable for large-scale systems. Recent research has investigated supervised and reinforcement learning (RL) approaches for NTO [ 11 ] . For instance, [ 12 ] develops an RL agent based on proximal policy optimization, while [ 13 , 14 ] propose curriculum-based RL to improve learning performance. Other works such as [ 15 , 16 ] mitigate unforeseen action consequences by simulating future outcomes, but at the cost of a computationally intensive Monte-Carlo tree search. [ 17 ] proposes a GNN-based imitation learning approach to predict the probability of each action’s effectiveness (as a soft label) in reducing maximum overloading, and [ 18 ] trains a heterogeneous GNN to imitate a rule-based expert agent for day-ahead operation and improve generalization to N-1 conditions. [ 19 ] investigates a different application of NTO, aiming to maximize cross-zonal capability rather than to relieve congestion. In [ 19 ] , two separate GNNs predict breaker statuses and DC flows for a given topology, followed by a repair optimization to enforce DC PF equations. Nevertheless, these approaches rely on reduced action spaces, cannot guarantee operational feasibility, and struggle with generalization and transferability. Here, we refer to generalization as the ability to perform well on unseen topologies or operating conditions within the same grid, and transferability as the ability of a trained model to be efficiently applied or adapted to different grids. This paper formulates NTO for CM with linearized AC PF constraints (AC-NTO-CM) as a MIP. We propose an edge-aware GNN that quickly predicts effective splitting actions, which are used as candidate solutions for AC-NTO-CM and also prioritized as branching variables in the branch-and-bound MIP solver. The proposed GNN is fully localized to learn local PF patterns, while a proximity filter improves learning performance by focusing on the most relevant regions. Thus, the proposed approach addresses fundamental challenges in ML-based NTO: scalability, generalization to unseen topologies, and transferability to different systems. The main contributions are: • The MIP formulation of NTO for CM considering linearized AC equations. • The proposed edge-aware GNN to predict busbar splitting candidates. The GNN’s expressiveness, the learning task design and the proximity filter, enable generalization to unseen N-k topologies and transferability across different power systems. • Significant computational speed-ups while maintaining constraint feasibility and minor optimality gaps. • A novel regression index that estimates the effectiveness of splitting actions, improving interpretability. II Congestion Management via Busbar Splitting Fig. 1: Schematic of the substation model for AC power flow. Substation reconfiguration via busbar splitting can effectively reduce transmission congestion by reconfiguring the grid to reroute flows, providing economic advantages over re-dispatch or load shedding measures. This paper focuses on two commonly used substation arrangements: the double-bus double-breaker and the breaker-and-a-half arrangements [ 20 , 6 , 10 ] . We extend