EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Adversarial Self‐Improvement for Resilient Smart Grid Dispatch Control Under Extreme N‐2 Contingencies

Mohamed Shamseldein

IET Smart Grid · 2026-01 · 9 (1)

DOI: 10.1049/stg2.70090 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IET Smart Grid, online erschienen am 19.11.2026. Verfasst von Mohamed Shamseldein. Die Arbeit nennt 31 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Deploying AI‐enabled dispatch controllers in smart grids demands quantifiable resilience against contingencies beyond the conventional N‐1 design basis. We present AG2CDT + , an adversarial self‐improvement framework that hardens offline‐trained dispatch controllers through iterative stress testing and corrective fine‐tuning. A grid adversary generates targeted N‐1, N‐2 and compound contingencies from critical transmission corridors. A safety oracle combining PTDF‐based preaction thermal screening, an advisory AC Jacobian voltage‐sensitivity predictor and duration‐aware poststep voltage enforcement identifies violations and generates corrected actions via line search; a special protection scheme (SPS) controller invokes targeted load shedding and line‐bridge switching whenever AC power flow diverges turning previously‐fatal scenarios into recoverable ones. Empirical measurement shows that the agent's active‐power dispatch moves monitored bus voltages by p.u./MW—two orders of magnitude less than postcontingency voltage sags. This reframes the original ‘thermal pre‐action, voltage post‐step’ asymmetry as a physically correct consequence of the agent's control authority, not a modelling oversight. Failed trajectories are corrected and used for behavioural cloning closing the loop between vulnerability discovery and policy hardening. On the NPCC 140‐bus benchmark with 4‐s AGC‐aligned control steps, AG2CDT + improves stress‐suite compliance from a ∼20% baseline to a peak of across three seeds (81% at Iter 4 on seed 7, 79% at Iter 2 on seed 42 and 83% at Iter 3 on seed 123), with the SPS recovering a high fraction of the PF‐divergent scenarios it attempts (10/10 at peak on seed 7) and reducing the unrecoverable‐divergence failure count from ∼30 to 7–13 per 100 scenarios (and from 76 to 7 for the previously‐degenerate seed 123). Ablation attributes the gain as follows: adversarial exposure alone provides the primary lift (to 56%); oracle‐guided correction adds a further 11 pp (to 67% at the no‐SPS Iter 2 peak); the SPS contributes ∼10 pp by converting PF‐divergent N‐2 cases into successes and one additional self‐improvement iteration enabled by the cleaner SPS‐augmented training signal (Iter 4 rather than the original Iter 2 peak) adds the final ∼4 pp. The advisory predictor contributes 0 pp to compliance by design, since it does not alter dispatch decisions; its value is methodological (symmetric CVU capability) and as an extension point for a future reactive‐power action channel. We further report, as an ablation diagnostic, that wiring as a hard preaction voltage gate collapses peak compliance to 37%–43% on the same seed—an empirical confirmation of the control‐authority argument. We quantify the catastrophic forgetting tradeoff: N‐0 compliance degrades from 100% to 70% at the peak‐robustness checkpoint (Iter 2) but recovers to 88% at the Pareto‐preferred Iter 4, establishing a Pareto frontier between normal‐operation reliability and contingency robustness that operators can navigate by selecting the appropriate iteration checkpoint—Iter 4 dominates Iter 2 on both axes. The framework requires ∼3.5 h on a single RTX 4070 GPU with 12 ms inference latency compatible with real‐time EMS/SCADA advisory deployment.

Zitieren

APA

Shamseldein, M. (2026). Adversarial Self‐Improvement for Resilient Smart Grid Dispatch Control Under Extreme N‐2 Contingencies. IET Smart Grid, 9(1). https://doi.org/10.1049/stg2.70090

Vancouver

Shamseldein M. Adversarial Self‐Improvement for Resilient Smart Grid Dispatch Control Under Extreme N‐2 Contingencies. IET Smart Grid. 2026;9(1). doi:10.1049/stg2.70090

BibTeX
@article{Shamseldein2026,
  author = {Shamseldein, Mohamed},
  title = {Adversarial Self‐Improvement for Resilient Smart Grid Dispatch Control Under Extreme N‐2 Contingencies},
  journal = {IET Smart Grid},
  year = {2026},
  volume = {9},
  number = {1},
  doi = {10.1049/stg2.70090}
}
Kostenlos starten →