Large power grids are increasingly optimized by several regional operators at once. Each operator solves its own piece of the problem and repeatedly exchanges boundary values (voltages and power flows on the lines between regions) with its neighbours until everyone agrees. The standard algorithm for this is ADMM, the alternating direction method of multipliers, which alternates local solves with price-like corrections on disagreements. Its behaviour depends on one knob, the penalty $\rho$, which sets how strongly disagreement is punished. A large penalty forces agreement quickly but makes each local problem badly conditioned. A small penalty keeps local solves easy but lets disagreement linger. Neural warm starts help a great deal. Learning Regionally Decentralized AC Optimal Power Flows with ADMM trains per-region networks to predict both boundary states and multipliers, and reaches the optimum in about 500 iterations against 3,000 for a cold start on the French grid. That work keeps $\rho$ fixed, though, and trains and tests only on uniformly scaled loads.
This note proposes to set the penalty on each boundary from how much the warm start should be trusted there. Boundaries where the predictor is confident get a stiff penalty, so they lock in fast. Boundaries where it is uncertain get a soft penalty, so ADMM can move them without ill-conditioning the neighbouring solves. Uncertainty comes from disagreement within a small ensemble of predictors. The point is the setting where the warm start is wrong: training on synthetic loads and deploying on real, spatially correlated demand. In that setting a fixed penalty either wastes the good boundaries or gets trapped by the bad ones.
Core Claim
The trade-off between speed and accuracy that the penalty controls should not be resolved once per algorithm. It should be resolved per boundary and per instance, using a calibrated estimate of warm-start error. Concretely, each region holds an ensemble of $K$ predictors. For each boundary variable $e$ (an edge in the consensus graph), we compute the ensemble variance $\sigma_e^2$ of its predicted value and set the initial penalty as
$$\rho_e^{(0)} = \bar\rho \cdot \mathrm{clip}\!\left(\frac{\bar\sigma^2}{\sigma_e^2},\, a,\, b\right),$$
where $\bar\rho$ is a reference penalty tuned on cold-start ADMM, $\bar\sigma^2$ is the median variance over all boundaries, and $[a,b]$ bounds the ratio so that no subproblem becomes arbitrarily ill-conditioned. The inverse-variance form comes from reading the augmented Lagrangian's quadratic term as a precision weight on the consensus value. After a short warm phase, the penalties relax geometrically toward residual balancing.
We claim three things. On in-distribution loads, this rule matches the best fixed-$\rho$ warm start. Under load-distribution shift, it reduces iterations to tolerance and the number of runs that stall. Both gains vanish when the variances are shuffled across edges. That last property is what separates the claim from "any adaptive penalty helps."
The closest prior work plays three different roles here. ML-ADMM motivates the proposal: it shows that warm starts matter and leaves the penalty fixed. Two adaptive-penalty papers supply the mechanism we compete with. Fast ADMM Algorithm for Distributed Optimization with Adaptive Penalty adapts edge-specific penalties from residuals and local loss fit, and ADMM without a Fixed Penalty Parameter grows the penalty stage-wise according to local sharpness. Both start from cold primal states and react to residuals after the fact. Our rule instead sets penalties before the first iteration, from information that only exists because a learned warm start exists. Learning-based AC-OPF Solvers on Realistic Network and Realistic Loads constrains the design. It shows that feasibility of a centralized neural solver falls to about 90% when trained on uniform loads and tested on real Tasmanian demand. Warm-start error under shift is therefore large and spatially structured, not small noise.
Research Program
We first reproduce ML-ADMM's regional setup, with per-region feed-forward predictors of boundary primals and multipliers and the convergence and standard-deviation filters on training data. We replace the single predictor with an ensemble of five networks trained with different seeds and bootstrap resamples. All ensemble members have the same architecture, so inference cost grows only linearly and stays negligible next to ADMM iterations.
The first scientific question is a precondition. Does ensemble variance actually track per-boundary warm-start error when the load distribution shifts? Ensembles are often overconfident out of distribution. If the rank correlation between $\sigma_e$ and realized error is weak under shift, the method has no basis and we stop. A cheaper fallback is to set penalties from the predicted local residual at iteration zero, computed with one local power-flow evaluation.
The second question is whether the trap described in the file's Gap 1 actually occurs. That hypothesis is that residual balancing misreads the skewed primal and dual residuals of a good warm start and erases its advantage. It is a hypothesis rather than a measured result, so we measure it directly.
The third question is the main comparison under shift, described below. Theory is secondary. Using a sharpness-based analysis in the style of LA-ADMM, we will try to bound how a mismatch between $\rho_e$ and true warm-start error inflates iteration count in the locally convex region around an AC solution. We make no global claims for the non-convex problem.
Baselines
The naive reference is cold-started regional ADMM with a single penalty tuned on validation instances. ML-ADMM reports it needs about 3,000 iterations on the French grid and shows objective errors up to 24% at iteration 5.
The strongest existing method is ML-ADMM itself, warm-started with fixed $\rho$. It reports optimality in about 500 iterations, a 6x speedup, and is within 1–3% of the optimal objective by iteration 5 on the French grid (Learning Regionally Decentralized AC Optimal Power Flows with ADMM).
We also run the same warm start combined with each of three adaptive rules: residual balancing, the edge-specific trust penalty ADMM-AP from Fast ADMM Algorithm for Distributed Optimization with Adaptive Penalty, and the stage-wise doubling with restarts from ADMM without a Fixed Penalty Parameter. The adaptive-penalty paper reports 37–40% speedups over standard ADMM on Hopkins 155, but that is a vision problem; for AC power flow there is no reported result found on Lacuna for either adaptive rule.
The ablation that isolates the key ingredient keeps the exact penalty distribution but randomly permutes the $\sigma_e$ across boundaries. This preserves heterogeneity and removes only the alignment between penalty and warm-start error. A second ablation replaces the ensemble with a single network that regresses $\rho_e$ directly.
All methods share the same predictors, data splits, ADMM implementation, local solver (Ipopt), and stopping tolerance. Each method gets the same number of validation instances for tuning its own hyperparameters: $\bar\rho$, $a$, $b$, the balancing ratio, and the doubling schedule. Beating ML-ADMM means a statistically significant reduction in median and 90th-percentile iterations to tolerance under shift, together with fewer runs that fail to converge within budget, and no loss in final objective gap or feasibility. Matching it in distribution is enough there.
Why not just…?
Why not just keep ML-ADMM's fixed penalty and tune it well? In distribution this may be nearly optimal, and we expect at most a small gain there. The proposal stands or falls on shifted loads. A single scalar cannot be stiff on boundaries the predictor gets right and soft on boundaries it gets wrong at the same time. The per-instance-tuned fixed $\rho$ baseline is an oracle that is unavailable in deployment. If our rule does not beat even the honestly tuned version under shift, the idea is wrong.
Why not just retrain the warm-starter on realistic correlated loads, as Cheung et al. do? This is the most serious threat. Training on a correlated load model restored feasibility from about 90% to 99.73% on TAS-97 in Learning-based AC-OPF Solvers on Realistic Network and Realistic Loads. If that removes the shift, confidence weighting has little left to fix. We therefore run both regimes. In the first, predictors are trained on the correlated model fitted to the early months of the 2016 logs and tested on later months, a real seasonal shift. In the second, predictors are trained on uniform loads, as most practitioners still do. If the correlated-trained warm start with fixed $\rho$ closes the gap in both, we will report that retraining is sufficient. In that case the contribution narrows to the uniform-trained or unknown-shift case.
Why not just apply off-the-shelf residual balancing to the warm start? It costs nothing and might work. The Gap 1 concern that it destabilizes good warm starts is unverified. The residual-balancing-plus-warm-start baseline answers this directly. If it matches our rule, the uncertainty estimate adds nothing and we drop it. We also log the penalty trajectories to see whether residual balancing makes large early jumps.
Why not just have the network predict the penalty directly? It needs labels for "good penalty," which come from expensive per-instance search, and it is itself a learned output that will be miscalibrated under the same shift. The direct-regression ablation tests this.
Proposed Experiments
The French transmission grids from ML-ADMM (France_EHV with 1,737 buses, LYON with 3,411, full France with 6,705) are the scale benchmark. We use the original uniform 80–122% scenarios for training and in-distribution testing. Shift is created by sampling test loads from a spatially correlated Gaussian whose covariance is built from bus geography. We label this plainly as a synthetic stress test, not real-world validation.
TAS-97 is the only real-load benchmark in the file. We partition it into three or four regions and run regional ADMM on the real 2016 demand. No regional-ADMM result exists for this grid, so all its numbers are new. Its small size limits claims about scale, and we will say so. ARPA-E GO confidential networks would be the natural real-world test at scale, but we do not assume access to them.
Metrics are iterations and communication rounds to a primal-dual tolerance, objective gap relative to centralized Ipopt, maximum boundary consensus violation, and the fraction of runs that converge within a 3,000-iteration budget. We report full distributions, not only means.
There are three decisive checks. First, the rank correlation between $\sigma_e$ and realized warm-start error under shift must be clearly positive, or we stop. Second, the shuffled-variance ablation must lose most of the gain, which shows that alignment between penalty and error, not heterogeneity alone, is doing the work. Third, the gain over the strongest warm-started adaptive baseline must persist on TAS-97's real later-month loads, not only on the synthetic French stress test.