When Does Learned Constraint Selection Pay in Security-Constrained Optimisation? A Boundary Map from SCOPF to Unit Commitment
Authors: Milad Kazemi, Behzad Kazemtabrizi
Organizations: M. Kazemi is with the Department of Computer Science, Durham University, Durham, U.K. · B. Kazemtabrizi is with the Department of Engineering, Durham University, Durham, U.K.
Learning-to-optimise methods for security-constrained optimal power flow (SCOPF) predict the security constraints that bind on each instance and solve only the reduced problem, with reported speed-ups of two to four orders of magnitude over the full problem. We ask whether learned selection beats the strongest learning-free alternative: a fixed umbrella set enforcing every constraint observed to bind in training. Our criterion, the price of a spare, answers this in advance: an enforced but inactive constraint is free, so the umbrella is exactly optimal wherever feasible, and learning can pay only where a spare decision carries a price. We test it systematically, including every PGLib-OPF network up to 10,000 buses. For SCOPF on a fixed network the umbrella is never beaten: its median size is five constraints across 43 networks, with no size trend, and our selector, while reproducing speed-ups of published magnitude (about x80,000 at case500), gains at most 0.25 ms over the umbrella and violates constraints on 4-5% of instances. Configuration diversity can break the umbrella's coverage on a small, densely meshed network but not at 2,000 buses, and we certify coverage with conformal risk control. In unit commitment, where a spare has a price, a static commitment loses 4-12% out of sample, and a predict-repair-reduce pipeline recovers it (cost gap 0.2-2.9%, fully feasible, 16-161x faster than the MILP). In our experiments, whether learning pays depends not on problem size but on whether a spare carries a price.
Figures & tables
Model
trained on labelled
trained on cost and
reduced QP at
active set (BCE)
violation (through QP)
deployment
Hybrid (full model)
yes
yes
yes
SSL
no
yes
yes
SL-active-set
yes
no
yes
SL-direct
regresses p directly
no
no
Umbrella
no training; fixed U
n/a
yes
TABLE I: Models compared. All learned modes share the same MLP.
Fig. 1: The two pipelines. (a) SCOPF: the learned selector (shaded) and the umbrella share the screen and the reduced QP; they differ only in who chooses the enforced constraints. (b) Unit commitment: predict, repair, reduced solve; no integer variables at deployment.
Fig. 2: The criterion as a map. Fixed-network SCOPF sits at the origin (free spares, closed coverage), whatever ρ looks like. Unit commitment moves right and learning pays. Configuration diversity moves up on a small dense network, where coverage must be certified, and returns to the origin at 2000 buses.
Model
viol. rate
worst (p.u.)
P / R
speed-up
Hybrid
4.10%
0.161
0.930 / 0.991
×241
SSL
3.20%
0.296
0.329 / 0.993
×195
SL-active-set
2.73%
0.161
0.988 / 0.994
×243
SL-direct
100%
0.454
—
×5560
Umbrella
0.00%
∼0
—
×235
TABLE II: Case118, 3000 held-out instances. Viol. rate: share of test instances whose reduced solution, before restoration, violates a screened constraint, the power balance or a generator limit by more than 10−5 p.u.; P/R: precision and recall of the selected set against the true active set. Speed-up is the median exact solve ( 286 ms) over the median per-query time, i.e., the network’s single-query latency plus the reduced solve (learned modes with a QP 1.2 – 1.5 ms, SL-direct 0.05 ms, umbrella with U1.2 ms), all timed single-threaded on exclusive nodes of one cluster.
Fig. 3: Left: reduced-solve time against the number of enforced constraints (median over the same test instances, one compiled solver per k , single-threaded on one node; dashed: the exact solve timed the same way); at every umbrella observed on a fixed network ( 6 – 19 constraints) it is a few milliseconds. Right: mean umbrella size over random training subsets of size n ; it saturates within the label budget.
Fig. 4: Tightening the emergency rating at fixed load. Left axis: umbrella size ∣U∣ (solid) and per-instance active set E∣S∣ (dashed). Right axis: fraction of the 100 instances that remain feasible (dotted). Case118’s window closes within one rung with ρ≥0.71 ; case500’s stays open longer and ρ dips to 0.33 , but ∣U∣ never exceeds 12 constraints.
Lever
∣U∣
E∣S∣
ρ
Outcome
Grid size: case500_goc (pre-registered)
6
2.61
0.44
umbrella shrank ; hypothesis falsified
Load spread: ±40% envelope, case118
16
6.18
0.39
umbrella exactly feasible at every load scale up to 1.1× ; selectors degrade
Contingency order: full N-2, case118
13
8.3
0.64
327 k candidates; saturates by n≈30
Corrective SCOPF, case2000_goc
161
54
0.34
ramp constraints scale, but umbrella = per-instance max
Library sweep, 43 networks ( 3 – 9591 buses)
med. 5
—
med. 0.67
max 279 at ρ=0.58 ; no size trend ( r=0.15 )
Rating tightness: rC to the boundary
≤12
≤6
0.33 – 0.80
∣U∣ stays tens; window closes by infeasibility
TABLE III: Falsification attempts on a fixed network. ρ=E∣S∣/∣U∣ . None opens a regime for learned selection.
Fig. 5: Umbrella size across the PGLib-OPF library ( 43 measured networks, 200 instances each, most congested securely servable operating point).
case118
case500_goc
case2000_goc
configurations measured
221 ( 169 branch, 52 gen.)
150 ( 30 branch, 120 gen.)
120 (gen.; +101 synthetic branch)
within-configuration ∣Ut∣
8.43±0.86
4.01±0.57
7.82±0.43
pooled ∣U∣ (pooled / within)
57 ( 6.8× )
9 ( 2.2× )
8 ( 1.02× ); 10 under branch outages
ρ pooled
0.11
0.27
0.87
median configurations per pooled constraint
2/221
14/150
119/120
CRC α=0.20 : achievable / test miss
74% / 0.202
100% / 0.096
100% / 0.000
TABLE IV: Configuration diversity and certified coverage. CRC rows give the fraction of 500 ( 300 for case118) family/calibration/test splits on which level α is achievable and the mean test miss rate; a level achievable on only a sliver of splits (case118, α=0.10 ) is conditioned on a rare calibration event and should not be read as reliably achievable.
Fig. 6: Size of the pooled umbrella against the number of configurations pooled. Case118 does not saturate; case500 slows; case2000 is flat under generator outages and creeps by six constraints under 101 adversarial and random branch outages.
Setting
static union gap
best-feasible static gap
static test-infeasible
1-NN gap (infeas.)
P&M, K=50
single-period, α=0.4
13.96%
+3.73%
0/150
1.2% ( 38% )
0.05% ( 0% )
single-period, α=0.8
23.08%
+6.90%
0/150
1.6% ( 41% )
0.05% ( 0% )
single-period, ±40% envelope
13.59%
+12.00%
0/143
1.3% ( 43% )
0.07% ( 0% )
multi-period, 1% MILP gap
22.7%
+9.6%
1/20
0.8% ( 25% )
—
multi-period, case500_goc ( 5% MILP gap)
13.2%
+9.7%
1/20
0.8% ( 25% )
—
Pipeline (multi-period)
L2O cost gap (infeas.)
bare predictor infeas.
best static schedule
speed-up vs. MILP
P&M, K=50
TABLE V: Unit commitment, out of sample. Top: static commitment gap on case118 (single-period 300/150 ; multi-period 30/20 daily profiles). Bottom: the predict-repair-reduce pipeline. P&M: the K-candidate method of [ 14 ] , not run on the 30/20 multi-period splits (see the bottom block). P&M times are per day with the K candidates priced in parallel.
AC Optimal Power Flow (ACOPF) and Security-Constrained Unit Commitment (SCUC) are fundamental optimization problems in power system operations. ACOPF serves as the physical backbone of grid simulation and real-time operation, enforcing nonlinear power flow feasibility and network limits, while SCUC represents a core market-level decision process that schedules generation under operational and security constraints. Although these problems share the same underlying transmission network and physical laws, they differ in decision variables and temporal coupling, and prior learning-based approaches address them in isolation, resulting in disjoint models and representations.We propose a learning framework that jointly models ACOPF and SCUC through a shared graph-based backbone that captures grid topology and physical interactions, coupled with task-specific decoders for static and temporal decision-making. Training includes solver supervision with physics-informed objectives to enforce AC feasibility and inter-temporal operational constraints. To evaluate generalization, we assess cross-case transfer on unseen grid topologies for ACOPF and SCUC without retraining, and systematic generalization on the UC-ACOPF problem using unsupervised, physics-based objectives and a power-dispatch consensus mechanism. Experiments across multiple grid scales demonstrate improved performance and transferability relative to existing learning-based baselines, indicating that the model can support learning across heterogeneous power system optimization problems.
Zeeshan Memon, Yijiang Li, Hongwei Jin +2
Emory University Atlanta, GA, USA · Argonne National Laboratory Lemont, IL, USA
Learning-based surrogates for AC optimal power flow (ACOPF) promise large speedups over classical solvers, but their operational value depends on physical feasibility as much as predictive accuracy. Physics-aware objectives such as the augmented Lagrangian (AL) improve constraint satisfaction at additional per-step cost, yet how this trade-off behaves with scale is uncharacterized. We sweep model and dataset sizes under MSE and AL training, and measure how violation changes with network size. Both objectives improve as power laws: MSE prediction loss depends more on model size than on data, while the AL composite of prediction loss and violation improves comparably along both axes. Violation grows markedly more slowly with network size under AL. On matched hardware and equal training samples, AL reduces violation about 19× with 2.4× the training time and negligible added memory. The training objective shapes not only where a surrogate lands but how its quality evolves with scale.
Yijiang Li, Emon Dey, Stefano Fenu +3
Argonne National Laboratory Lemont, USA · Oak Ridge National Laboratory Oak Ridge, USA
AC optimal power flow (ACOPF) is foundational yet computationally expensive in power grid operations, driving learning-based surrogates for large-scale grid analysis. These surrogates, however, often fail to generalize across network topologies, a critical gap for deployment on grids not seen during training and for routine operational what-if studies. We introduce LUMINA-Bench, a comprehensive benchmark suite for ACOPF surrogate learning covering multi-topology pretraining, transfer, and adaptation. The benchmark evaluates homogeneous and heterogeneous architectures under single- and multi-topology learning settings using unified metrics that capture both predictive accuracy and physics-informed constraint violations. We additionally compare constraint-aware training objectives, including MSE, augmented Lagrangian, and violation-based Lagrangian losses, to characterize accuracy-robustness trade-offs across settings. Data processing, training, and evaluation frameworks are open-sourced as the LUMINA suite to support reproducibility and accelerate future research on feasibility-aware OPF surrogates.
Hongwei Jin, Keunju Song, Zeeshan Memon +5
Mathematics and Computer Science Division, Argonne National Laboratory · Department of Electronic Engineering, Sogang University · Department of Computer Science, Emory University +1