Learning Topology-Transferable ADMM Penalties for Real-Time Power Flow

Power grid operators must re-solve the optimal power flow problem every time the network changes: a line trips, a feeder is reconfigured, a storm takes out a branch. Optimal power flow (OPF) means choosing generator outputs and voltages so that demand is met at least cost without violating line or voltage limits. A standard way to split this large, nonlinear problem across regions or buses is the alternating direction method of multipliers (ADMM). In ADMM, each piece solves its own small problem and then negotiates with its neighbors over shared boundary quantities until they agree. On real grids that negotiation is slow. On French transmission networks with up to 6,705 buses, cold-started ADMM needs about 3,000 rounds, and still about 500 when warm-started by learned predictors (Learning Regionally Decentralized AC Optimal Power Flows with ADMM). Most of that slowness comes from one choice that is usually made by hand: how strongly each boundary is penalized for disagreeing.

We propose to learn that choice with a small graph neural network (GNN), a model that computes each node's output by repeatedly exchanging messages with its neighbors and therefore runs on graphs of any size or shape. The GNN reads only local electrical features: line impedance, loading, the size of the current disagreement, and how it is changing. From these it outputs a separate penalty and step size for every boundary line at every ADMM round. Because nothing in the model is tied to a particular bus count or line list, the same trained model can be applied unchanged to a grid with a line outage, or to a much larger grid it never saw. The claim is that a fixed budget of 10–25 ADMM rounds with these learned, locally conditioned penalties can reach near-optimal, near-feasible AC solutions on unseen grid topologies. That is the regime where today's learned OPF solvers must be retrained and classical ADMM is still far from converged.

Core Claim and What Is New

Three bodies of work each cover part of this. Deep unfolding treats each iteration of an algorithm as one layer of a network and learns per-iteration parameters. Its topology-transferable versions all live in wireless interference networks. UWMMSE parameterizes the WMMSE power-control iterations with GNNs and transfers from 20 to 10 and 30 nodes (Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation). USCA does the same for successive concave approximation and transfers from 8 to 100 users (Graph-based Algorithm Unfolding for Energy-aware Power Allocation in Wireless Networks). In both, users interact only through aggregate interference, with no conservation law linking them. These papers supply the method, GNN-parameterized unfolding with progressive layer training. They do not test whether it survives Kirchhoff coupling on a meshed grid.

Two papers unfold ADMM specifically. Unfolded distributed ADMM shows a clear tradeoff on lasso and linear regression with 5 agents (Distributed Learn-to-Optimize): per-agent hyperparameters are expressive but cannot move to new graphs, while globally shared ones transfer but cannot adapt to locally hard subproblems. GNN-ADMM predicts per-node step sizes and per-edge weights (Learning to accelerate distributed ADMM using graph neural networks) and cuts consensus error from 9.17 to 1.96 at 10 iterations, but only on 8-node random graphs with convex objectives. These two papers motivate the proposal by exposing the tradeoff. We resolve it by computing per-edge parameters from local features with shared GNN weights. GNN-ADMM's architecture could in principle do this, but it has never been tested for transfer, at grid scale, or on a non-convex problem.

Finally, the learned OPF solvers that do reach realistic grids are tied to one topology. DeepOPF maps loads to generation with a fixed-size network (DeepOPF). ML-ADMM uses regional MLPs tied to fixed boundary variables. Both explicitly cannot handle reconfiguration. They constrain our design in two ways: any learned component must be size-agnostic, and feasibility must be checked, not assumed.

The new element is therefore narrow and testable. We learn only the coordination parameters of an otherwise classical, physics-exact ADMM for AC-OPF. Every local subproblem still enforces the true nonlinear power-flow equations. We then ask whether those parameters transfer across topology at a fixed, real-time iteration budget.

Research Program

We decompose AC-OPF by region, following ML-ADMM, and also at finer bus-cluster granularity. Each line $(i,j)$ that crosses a partition boundary carries duplicated voltage variables $v_{ij}$ and $v_{ji}$ held by its two endpoint subproblems. At round $k$, each subproblem minimizes its local cost $f_i$ plus a penalized disagreement term:

$$ \min_{x_i \in \mathcal{X}_i} \; f_i(x_i) + \sum_{j \in \mathcal{N}(i)} \Big[ \lambda_{ij}^{k\top}(v_{ij} - z_{ij}^k) + \tfrac{\rho_{ij}^k}{2}\lVert v_{ij} - z_{ij}^k\rVert^2 \Big]. $$

Here $x_i$ are the local decisions, $\mathcal{X}_i$ is the exact local AC feasible set, $\mathcal{N}(i)$ is the set of neighboring subproblems, $z_{ij}^k$ is the current consensus value for the shared boundary voltages, $\lambda_{ij}^k$ is the dual price on disagreement, and $\rho_{ij}^k$ is the penalty. The dual update becomes $\lambda_{ij}^{k+1} = \lambda_{ij}^k + \eta_{ij}^k \rho_{ij}^k (v_{ij}-z_{ij}^{k+1})$, where $\eta_{ij}^k$ is a learned dual step multiplier. One GNN $g_\theta$, with weights shared across all edges and all rounds, outputs $(\rho_{ij}^k, \eta_{ij}^k)$. Its inputs are line admittance, thermal limit, current flow, primal residual $\lVert v_{ij}-z_{ij}^k\rVert$, dual residual $\lVert z_{ij}^k - z_{ij}^{k-1}\rVert$, and their recent history. Sharing weights across rounds means one trained model serves any budget $K$.

Training uses the gap to a reference AC solution plus a penalty on power-flow mismatch, summed over rounds, so early rounds are rewarded too. Backpropagating through all subproblem solves is the memory bottleneck flagged by GNN-ADMM and FlowOptimizer. We therefore train progressively, a few rounds at a time, as in USCA. Gradients flow through the subproblems by implicit differentiation of their optimality conditions, rather than by unrolling the inner solver.

Training data are N-1 and N-2 outage variants of IEEE 30, 57, 118, and 300-bus systems, with load variations, following DeepOPF's sampling. Testing is zero-shot on unseen outage patterns and on larger grids: the RTE France_EHV (1,737 buses), LYON (3,411), and France (6,705) networks used by ML-ADMM, and selected ARPA-E GO Challenge 2 synthetic grids (ARPA-E Grid Optimization (GO) Competition Challenge 2). Reference solutions come from a standard nonlinear solver.

Discrete line switching is deliberately excluded. The model must generalize over topologies that are given to it, not choose them. Choosing topology is a later step, and the method only earns it if it works on given topologies first.

Baselines

The naive reference point is cold-started ADMM with a single hand-tuned global penalty, run for the same budget $K$. ML-ADMM reports this reaching errors as high as 24% at iteration 5 on its grids.

We also run a classical adaptive-penalty rule, residual balancing: raise $\rho$ when the primal residual dominates and lower it when the dual residual dominates. It is the obvious training-free way to get per-edge adaptivity, and it is tuned on the same training grids.

The shared-scalar unfolded variant learns one $\rho^k$ and one $\eta^k$ per round, the transferable scheme of Distributed Learn-to-Optimize. It isolates whether per-edge conditioning matters.

The strongest existing method is ML-ADMM (Learning Regionally Decentralized AC Optimal Power Flows with ADMM). It reports reaching optimality in about 500 iterations versus 3,000 for cold-started ADMM, and an objective within 1–3% of optimal within 5 iterations. We retrain it per network on the RTE grids with its own data filtering, so it receives in-distribution training while ours does not. We also give it the same training-sample budget per network in a second comparison.

Two further baselines complete the set. A size-agnostic GNN warm-start, which predicts initial primal and dual values followed by fixed-penalty ADMM, is the topology-transferable version of ML-ADMM. A direct GNN predictor of the OPF solution, with a power-flow correction step, is the single-shot analogue of DeepOPF. Every learned method gets the same training grids, the same hyperparameter search budget, and the same hardware. Latency is reported as both communication rounds and wall-clock time.

Beating the field means three things on unseen-outage RTE grids at $K \le 10$. Our zero-shot model must match ML-ADMM's in-distribution 1–3% optimality gap at comparable constraint violation. It must clearly beat residual balancing and the shared-scalar variant. And on topologies where ML-ADMM cannot run without retraining, it must stay within that band. No reported result was found on Lacuna for any GNN-unfolded ADMM on AC-OPF.

Why not just…?

Why not just use a GNN to warm-start classical ADMM? This is the real threat. If a good starting point is what matters, learning penalties adds complexity for nothing. ML-ADMM's results show that warm-starting alone buys a 6x speedup, so the threat is serious. We settle it with a factorial ablation: warm-start only, learned penalties only, and both together, all evaluated on unseen topologies. If warm-start-only matches the combined model at $K \le 10$, we will report that and recast the contribution as a transferable warm-starter. Our prior expectation is that a warm start helps at round zero, but a bad penalty on a stressed or outaged line keeps the iterations from converging no matter where they start.

Why not just predict the solution directly with a GNN? Single-pass prediction is the fastest option; DeepOPF reports 0.37 ms on the 300-bus case with under 0.2% cost error. But it reports this on a fixed topology and only with a dedicated feasibility projection. Distributed Learn-to-Optimize and USCA both found that black-box networks generalize worse than unfolded ones under limited iterations. We test this directly by measuring power-flow mismatch and line-limit violations of the direct predictor on unseen outages and larger grids.

Why not just use a classical adaptive penalty rule? It needs no training and transfers trivially to any grid. It is included as a baseline. If it closes most of the gap, the learned model's value would be limited to the hardest stressed-outage cases, and we will report it that way.

Why not just linearize and solve centrally? Rectangular linearization solves the IEEE 123-node feeder in 2–2.5 seconds (Scalable Optimization Methods for Distribution Networks With High PV Integration). That is too slow for sub-second use, and the linearization discards the reactive and loss behavior that matters most under outages. We include it on the 123-node and 2,500-node feeders as a latency and accuracy reference point.

Decisive Checks

Two outcomes would sink the proposal. The first is that the learned per-edge penalties fail to beat residual balancing on held-out outages of the training grids. The second is that their advantage disappears when moving from 300-bus training grids to the 6,705-bus France network. We will also check whether the learned $\rho_{ij}$ correlates with line admittance and loading in a way that is consistent across grids, since a consistent relationship is the mechanism that would explain transfer.