Scalable and Reliable State-Aware Inference of High-Impact N-k Contingencies

Increasing penetration of inverter-based resources, flexible loads, and rapidly changing operating conditions make higher-order $N!-!k$ contingency assessment increasingly important but computationally prohibitive. Exhaustive evaluation of all outage combinations using AC power-flow or ACOPF is infeasible in routine operation. This fact forces operators to rely on heuristic screening methods whose ability to consistently retain all critical contingencies is not formally established. This paper proposes a scalable, state-aware contingency inference framework designed to directly generate high-impact $N!-!k$ outage scenarios without enumerating the combinatorial contingency space. The framework employs a conditional diffusion model to produce candidate contingencies tailored to the current operating state, while a topology-aware graph neural network trained only on base and $N!-!1$ cases efficiently constructs high-risk training samples offline. Finally, the framework is developed to provide controllable coverage guarantees for severe contingencies, allowing operators to explicitly manage the risk of missing critical events under limited AC power-flow evaluation budgets. Experiments on IEEE benchmark systems show that, for a given evaluation budget, the proposed approach consistently evaluates higher-severity contingencies than uniform sampling. This allows critical outages to be identified more reliably with reduced computational effort.

💡 Research Summary

The paper addresses the growing need for higher‑order N‑k contingency analysis in modern power systems, where the proliferation of inverter‑based resources, flexible loads, and rapid renewable fluctuations makes traditional N‑1 security insufficient. Exhaustively evaluating every possible N‑k outage combination with AC power‑flow (ACPF) or AC optimal power flow (ACOPF) is computationally infeasible because the number of combinations grows combinatorially with the size of the network and the order k. Existing approaches—reduced‑order models, sensitivity‑based screening (e.g., LODF), and supervised severity‑prediction—still require evaluating or ranking a huge set of candidates and provide no formal guarantee that the most critical contingencies are retained.

The authors propose a fundamentally different paradigm: instead of enumerating all contingencies, directly generate a compact set of high‑impact N‑k outage scenarios conditioned on the current operating state. The framework consists of three tightly coupled components:

1. Topology‑aware Edge‑Varying Graph Neural Network (EVGNN) – Trained only on the base case and all N‑1 outages, EVGNN learns to map a contingency mask (which lines are out) together with the current system state (loads, generation, voltages) to a risk score (S(x,c)). By operating on the bus‑branch graph, it captures how stress propagates through network topology, enabling rapid scoring of multi‑line outages without any AC‑flow simulation.

2. Conditional Diffusion Model for Contingency Generation – A diffusion‑based generative model is conditioned on the same operating state (x) and learns the conditional distribution (p(c|x)). During training, the high‑risk samples identified by EVGNN are used as the target “tail” of the distribution, so the diffusion process is encouraged to concentrate probability mass on severe contingencies. Sampling from the diffusion model therefore directly yields candidate N‑k outage patterns that are statistically likely to be high‑severity, eliminating the need to loop over the combinatorial space.

3. Probabilistic Coverage Guarantee and Budget‑aware Evaluation – Operators specify an evaluation budget (B) (the number of ACPF/ACOPF solves they can afford) and a risk tolerance (\epsilon). The diffusion model generates (B) candidates, EVGNN instantly scores them, and the top‑(B) are forwarded for detailed AC analysis. The authors prove that, by appropriately setting the diffusion sampling parameters and a quantile threshold (e.g., the 95‑percentile of the severity distribution), the resulting shortlist contains the true top‑(m) most severe contingencies with probability at least (1-\epsilon). This provides a formal, user‑controllable trade‑off between computational effort and the probability of missing a critical event.

The methodology is validated on IEEE 14‑, 39‑, 57‑, and 118‑bus test systems across a range of operating points, including high‑load and high‑renewable scenarios. For each point, the authors compare their approach against a uniform random sampling baseline under the same ACPF budget. Results show:

• Near‑100 % success rate of ACPF solves for generated candidates (no infeasibilities caused by unrealistic masks).
• A significantly larger fraction of evaluated contingencies fall into a high‑severity band (e.g., the top 5 % of severity scores), often 2–3 times higher than the baseline.
• Top‑(m) screening curves that rise much faster, indicating that severe contingencies appear earlier as the budget increases.
• Particularly strong performance in stressed conditions where multi‑line interactions cause voltage collapse or overloads that heuristic methods frequently miss.

Key insights include:

• Data efficiency: Training only on base and N‑1 cases suffices because EVGNN can extrapolate to multi‑line outages by leveraging graph structure.
• Rare‑event focus: Diffusion models excel at learning and sampling from the tail of a distribution, which aligns perfectly with the need to capture rare but catastrophic contingencies.
• Operational flexibility: The coverage guarantee lets system operators explicitly set how much risk they are willing to accept given limited computational resources.

In conclusion, the paper delivers a scalable, reliability‑aware pipeline that transforms N‑k contingency analysis from an exhaustive or heuristic screening problem into a conditional generation problem with provable risk bounds. It dramatically reduces the number of AC simulations required while improving the likelihood of detecting the most dangerous outages. Future work suggested includes online adaptation with streaming measurements, extension to larger real‑world grids, and integration of additional risk metrics such as restoration cost or cascading failure probability.