When a power line trips, the electricity it carried moves onto other lines, sometimes distant ones, and those lines may trip in turn. Machine-learning models now speed up the search for such chains of failures. A graph neural network (GCN, a model that passes information along the grid's wiring diagram) can rank which branch trips will lead to load loss. Causal models can predict the next line to fail. These models output rankings or scores, not probabilities that anyone has checked against how often failures actually happen. Operators need those probabilities. Whether to island a region or shed load is a bet that weighs the cost of acting against the chance of collapse. A threshold rule for that bet is cost-optimal only if the probability is right in absolute terms, a point proven for model-routing systems by Forecasting Model Success at Inference Time.
The tempting fix is to take each line's predicted trip score and recalibrate it with a standard method such as isotonic regression. Our central claim is that this is not enough, for a specific reason. Line failures in a cascade are strongly dependent, because flow redistribution couples them. Per-line probabilities can therefore be well calibrated on average while the probabilities that matter for decisions come out badly wrong. Those are the probability that total load shed exceeds some level, and the probability that a cascade reaches a given depth. We propose to measure calibration at the level of the cascade outcome. We will build a stage-wise surrogate trained to be calibrated at that level, and we will test whether this changes operating decisions under asymmetric costs.
Core Claim
Let $x$ be the pre-contingency grid state together with an initiating $N\!-\!k$ outage, and let $S$ be the resulting total load shed. Here $S$ is random because loads fluctuate and protection relays trip stochastically. A predictor supplies a cumulative distribution $\hat F(s \mid x)$. It is cascade-calibrated if the probability integral transform is uniform:
$$ \hat F(S \mid x) \sim \mathrm{Uniform}(0,1), $$
and we will also require this uniformity separately within strata of grid stress (load multiplier) and cascade stage. This is the survival-calibration criterion of Kamran and Wiens, which states that predicted cumulative probabilities evaluated at true event times must be uniform. We transplant it from patient event times to cascade outcome size.
The claim we will test is that a surrogate whose per-stage line-trip probabilities $p_\ell^{(k)}$ (line $\ell$, stage $k$) are each marginally calibrated can still violate this condition. The gap is largest in the upper tail, which is the part an operator cares about. Composing stage-wise marginals into an outcome distribution treats failures as conditionally independent, and that assumption fails for non-local redistribution. Calibration error therefore compounds with cascade depth. The research file names this gap without filling it. Cascade predictors "treat line failure likelihoods as deterministic physics triggers or uncalibrated classification scores," and no work addresses recalibration as the topology mutates during collapse.
Research Program
The first component is a ground truth with real frequencies. Existing cascade benchmarks are mostly deterministic given a scenario. We will build a stochastic cascade simulator on top of DC power flow in MATPOWER. It perturbs bus loads by a random factor in [0.9, 1.1], the protocol used in Guiding Cascading Failure Search with Interpretable Graph Convolutional Network. It also gives each overloaded line a loading-dependent trip probability, so that relay behaviour is stochastic rather than a hard flow-versus-rating threshold. Rerunning the same $x$ hundreds of times yields an empirical distribution of $S$ and of stage-wise trip frequencies, which is what calibration is checked against. We acknowledge the DC-flow limitation the file flags for most of this literature: no voltage collapse and no reactive-power deficits. The claims are scoped to that physics.
The second component is the surrogate itself. It is a graph network applied repeatedly, once per cascade stage. At each stage it reads the current topology and flows and outputs trip probabilities. Sampling from these probabilities, updating the topology, and repeating gives a fast Monte Carlo estimate of $\hat F(s \mid x)$.
Two training signals sit on top of ordinary per-stage cross-entropy. The first is an outcome-level calibration loss built in the style of X-CAL. That paper made the D-calibration test differentiable using soft bin memberships and a mini-batch upper bound. We apply the same construction to the PIT values of $S$ computed from surrogate rollouts, with a straight-through relaxation of the sampled trips. Its weight $\lambda$ trades calibration against ranking quality, exactly as X-CAL trades calibration against concordance. The second signal is a cheap post-hoc option: an isotonic recalibrator conditioned on stage index and a stress summary, so that calibrators fitted on stage-1 states are not reused on stage-5 states. This targets the distribution shift that invalidates static calibrators, a failure noted by both Aligning Evaluation with Clinical Priorities and Kotte.
The third component is decision evaluation. We use the Schervish representation from Aligning Evaluation with Clinical Priorities, which reads a proper score as an average of cost-weighted decisions. With it we compute expected regret of a threshold load-shed or islanding policy across a sweep of cost ratios between unmitigated collapse and unnecessary shedding. We also confront the "calibration trap" of Rethinking Confidence Calibration for Failure Prediction. That paper showed calibration regularizers such as label smoothing and mixup cut failure-detection AUROC on vision benchmarks, and it preferred flat-minima training (FMFP) for separability. We will report whether outcome-level calibration costs AUROC at ranking dangerous contingencies. If it does, we will report whether a flat-minima-trained surrogate with post-hoc outcome calibration recovers both properties.
Baselines
The naive reference is the deterministic surrogate used as practice does today. A GCN of the interpretable GCN search type is trained to predict load shedding, and its sigmoid scores are read directly as probabilities. Its ranking quality is strong: on IEEE RTS-79 it found all 22 load-shedding cascade paths in 50 search attempts, against over 1000 for LODF and PFW. No reported result was found on Lacuna for calibration error of this or any cascade predictor. Beating it means lower outcome-level PIT deviation and lower decision regret, while staying within a few points of its contingency-ranking AUROC.
The strongest existing next-failure predictor is C-Path from Cascading Failure Prediction via Causal Inference. It learns a causal graph among lines with Cyclic-LiNGAM and simulates outages as interventions. It reports 45.7% next-failure precision on the IEEE 39-bus system. We will convert its intervention outputs into trip probabilities via a softmax over predicted flow excess, fitted on the same calibration split, and score it with the same metrics. Matching its 45.7% precision while being calibrated would show calibration is not bought with accuracy.
The baseline we consider the real threat is per-line isotonic recalibration of the stage-wise surrogate. This is the method that cut ECE from 0.12 to 0.03 in Kotte's language-model routing setting, though that figure comes from a different domain. It receives the same data, the same calibration split and the same hyperparameter budget. The proposal beats it only if the outcome-level PIT test and tail-probability error are clearly better while per-line ECE is comparable.
Two ablations isolate the key ingredient. One removes the outcome-level loss but keeps stage-conditioned isotonic recalibration. The other keeps the outcome loss but strips stage conditioning. The full physics Monte Carlo serves as the accuracy ceiling at its own cost. All learned methods share the same simulator draws, architectures, and roughly equal tuning budgets, measured in GPU hours.
Why not just…?
Why not just recalibrate each line's probability with isotonic regression? This is the most likely way the proposal fails, so it is the first experiment. If per-line isotonic calibration already yields a uniform PIT for load shed on IEEE 118-bus at high stress, the compounding argument is wrong. In that case the contribution shrinks to a benchmark and the method should be dropped. Our prior that it will not suffice rests on the non-local redistribution documented by Zocca et al. and the causal paper, but the experiment decides.
Why not just run the physics simulator many times? Monte Carlo over a stochastic DC simulator is the ground truth and needs no calibration. On small grids it may well be fast enough, and we will report the compute at which it matches the surrogate's error. The case for a surrogate rests on scale: high-fidelity models "take days to run" according to HotSpots, and traditional search needed 34,822 to 38,608 attempts on the Henan grid. If equal-compute Monte Carlo wins on 300-bus, we will say so.
Why not just put a conformal interval on cascade size? Adaptive Robust Optimization with Data-Driven Uncertainty uses spatio-temporal conformal sets for weather outages. Split-conformal on a direct load-shed regressor would give marginal coverage at a chosen level cheaply, and that is a fair competitor. It does not provide a full distribution usable at every cost ratio. Its guarantee is also marginal over contingencies, not conditional on stress level. We include it, run at matched coverage levels, and test conditional coverage within stress strata, where we expect it to break.
Why not just regress cascade size directly and skip stages? A one-shot distributional regressor avoids compounding entirely. It is included as an ablation. It loses stage-wise trip probabilities, which mid-cascade control needs. If it matches on outcome calibration, the stage-wise design is justified only for in-cascade decisions.
Proposed Experiments
Data come from our stochastic simulator on IEEE 39-bus, RTS-79, 118-bus and 300-bus. Initiating events are $N\!-\!2$ contingencies, loads are scaled from nominal up to 150% (the stress range the causal paper tested), and each initial state is repeated 200 times for empirical frequencies. Splits are by initiating contingency, so test cascades start from unseen outages, with an extra shifted test at stress levels absent from training.
Metrics are per-line ECE and Brier score, stage-stratified ECE, PIT uniformity of load shed (D-calibration statistic), absolute error on $P(S > 10\%\text{ of load})$, contingency-ranking AUROC, and expected regret of threshold shedding across cost ratios from 1 to 1000.
The decisive check is the compounding test. We plot outcome-level calibration error against cascade depth for the per-line-isotonic baseline and for our model. The hypothesis is supported only if the baseline's error grows with depth and stress while ours stays flat, at equal per-line ECE, and if that gap translates into lower regret at high cost ratios. A second check settles the calibration-trap question for this domain: does outcome calibration reduce contingency AUROC, as label smoothing did on vision benchmarks in Zhu et al., and does combining flat-minima training with post-hoc calibration remove the loss? A null on the first check is itself publishable. It would tell grid operators that cheap per-line recalibration is enough.
References
- Forecasting Model Success at Inference Time: Calibrated Probabilistic Forecasts for Cost-Optimal LLM Cascades
- Estimating Calibrated Individualized Survival Curves with Deep Learning
- Guiding Cascading Failure Search with Interpretable Graph Convolutional Network
- X-CAL: Explicit Calibration for Survival Analysis
- Aligning Evaluation with Clinical Priorities: Calibration, Label Shift, and Error Costs
- Rethinking Confidence Calibration for Failure Prediction
- Cascading Failure Prediction via Causal Inference
- A Spectral Representation of Power Systems with Applications to Adaptive Grid Partitioning and Cascading Failure Localization
- HotSpots: Failure Cascades on Heterogeneous Critical Infrastructure Networks
- Adaptive Robust Optimization with Data-Driven Uncertainty for Enhancing Distribution System Resilience