arXiv · Jul 17, 2026
Revisiting data-driven dynamic security assessment with a tabular foundation model
A single tabular foundation model (TFM) achieves an average Macro F1 score of about 90% for pre-fault dynamic security assessment on the IEEE 68-bus system using only 120 labelled samples per contingency, eliminating the need for separate models per contingency and enabling in-context learning without retraining.
What happened
Data-driven pre-fault dynamic security assessment (DSA) traditionally requires a large labelled database and a separate model for each contingency, with poor generalization to unseen contingencies. This work introduces a tabular foundation model (TFM) that uses in-context learning to assess stability across many contingencies with a single model, requiring no retraining or hyperparameter optimization. The study also characterizes the use of electrical distance coordinates (EDC) as continuous features to improve generalization to unseen contingencies, showing that a few labelled samples can reliably enhance performance. Comprehensive case studies on the IEEE 68-bus system demonstrate that the TFM attains an average Macro F1 score of about 90% with only 120 labelled samples per contingency, roughly two orders of magnitude fewer than traditional methods.
Technical significance
The TFM leverages in-context learning to perform dynamic security assessment without task-specific training, using electrical distance coordinates as continuous features to encode topological information. This enables generalization to unseen contingencies when EDC captures relevant structural similarities, but the paper also identifies conditions where EDC alone is insufficient, suggesting that a small number of labelled examples can bridge the gap.
Industry impact
This approach could significantly reduce the computational and data burden for power system operators by replacing many specialized models with a single foundation model, potentially lowering barriers to real-time security assessment and enabling faster adaptation to grid changes.
What to watch
Next signals include validation on larger, more realistic power systems, integration with real-time SCADA data, and exploration of TFM's applicability to other power system tasks such as optimal power flow or fault diagnosis. The need for few-shot generalization to truly unseen topologies remains an open challenge.
Decision value
Reducing the need for extensive labelled datasets and per-contingency model maintenance can lower operational costs for utilities and grid operators, while improving reliability and speed of security assessments. This could accelerate adoption of AI in critical infrastructure monitoring.