• Citer

Topology-aware feature engineering for data-driven Dynamic Security Assessment using electricity market data

par Paldino, Gian Marco ;Petit, Raphael;Iuliano, Silvia;Vaccaro, Alfredo A.;Bontempi, Gianluca
Référence International journal of electrical power & energy systems, 177, 111865
Publication Publié, 2026-04

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Data-driven surrogate models for Dynamic Security Assessment increasingly rely on machine learning to classify the security state of power systems from operational measurements. However, most existing approaches treat input features as flat, unstructured vectors, discarding the spatial relationships encoded in the underlying grid topology. This paper proposes a systematic feature engineering framework that explicitly encodes the power system graph structure into the input space of a classification model, using publicly available electricity market data aggregated at the zonal level. Three categories of topology-aware features are developed: local neighborhood context, global spatial distributions, and formal graph-theoretic properties. The framework is validated on a real-world dataset from the Italian transmission power system, comprising more than 51,000 observations at 15-minute resolution. Through four experimental scenarios evaluated with time-series cross-validation, we show that naively adding all spatial features degrades performance due to noise, but a surgical selection of the most consistently predictive spatial features, combined with the removal of redundant raw inputs, yields a focused model that achieves the best Precision Recall Area Under Curve (0.1458 vs. 0.1406 for the baseline) and the lowest asymmetric operational cost. Sensitivity analyses confirm that these conclusions are robust to hyperparameter choice, imbalance-handling strategy, and cost ratio assumptions across the 5:1 to 100:1 range.