A self-supervised neural network paired with a conditional repair step for constraint violations reported low violations across five simulated IEEE feeders, while one comparison approach failed to converge. Called Penalty+SLFS, the method had the lowest violations among the learning-based methods on feasible cases, meaning cases that met the constraints. Direct Penalty failed to converge on every feeder in the base tests, while the other learning-based methods succeeded.
The network learns from the equations
Penalty+SLFS combines a neural network that predicts control settings for distributed energy resources, or DERs, with a differentiable fixed-point power-flow solver. It trains directly from the objective and constraints of AC optimal power flow, or AC-OPF, rather than from labeled optimal solutions. In practical terms, the network is taught to make settings that fit the electrical equations and limits built into the problem.
SLFS is the repair stage used when the network makes a new prediction. It corrects the predicted DER setpoints only when the total network constraint violation exceeds a threshold. The paper states a monotonic-violation guarantee for SLFS and describes its implementation as mainly GPU-friendly matrix-vector operations. That guarantee is conditional on stated assumptions about contraction, the linear model and damping.
A test that changed scale and topology
The paper targets a combination of multiphase, unbalanced physics, validity across discrete switch configurations and very large distribution feeders. Here, topology, the way the network is connected, is part of the problem rather than a fixed background condition.
The evaluation used five IEEE feeder models: 13-bus, 123-bus, 240-bus, 906-bus and 8500-node systems. They contained, respectively, 3, 13, 22, 12 and 100 PV units, and 2, 8, 9, 8 and 45 switches. This gave the method a test range that extended from a small feeder to the paper's largest network.
Each feeder contributed 200,000 training samples, including feasible and infeasible cases, while each test set contained 2,000 samples. The sampled conditions varied loads and DER availability. Switch statuses were sampled independently, with each switch having a 0.5 probability of being closed, producing both meshed and radial topologies.
Strong benchmark numbers, narrow reach
On feasible cases in the base tests, Penalty+SLFS's maximum reported constraint violation was no more than 3.0 × 10^-4 on the 8500-node feeder. On the 906-bus feeder, the reported violation was as low as 3.2 × 10^-8. These are benchmark measurements for the sampled scenarios, not a result covering every possible grid condition.
Against the study's IPOPT reference, the optimality gap, a percentage measure of distance from the reference objective, stayed below 0.1% on the 13-bus, 123-bus and 906-bus feeders and below 0.2% on the 240-bus feeder. On the 8500-node case, the gap reached 1.5%, although the absolute cost error was $0.06. The paper says the relative gap there was inflated by small objective values.
Those percentages should not be read as proof of global optimality. IPOPT is the stated reference, and the paper warns that gaps on infeasible instances require careful interpretation. The evaluation is bounded by IEEE models and sampled conditions, so it does not settle how the method would perform on other feeders, loads, DER availabilities or switch configurations. Transfer to real utility feeders, and stability when the stated assumptions fail or many switches change at once, remain open questions.
Paper data and sources
Original title: Scalable Self-Supervised Learning for Multiphase AC-OPF in Distribution Systems with Topology Reconfiguration
Authors: Hoang T. Nguyen, Shaohui Liu, Reetam Sen Biswas et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text