Revisiting Data-Driven Dynamic Security Assessment with a Tabular Foundation Model

This paper applies a tabular foundation model (TFM) to pre-fault dynamic security assessment in power systems. The model uses in-context learning to evaluate multiple credible contingencies without retraining or hyperparameter optimization for each classifier. On the IEEE 68-bus system, one TFM reaches approximately 90% average Macro F1 with only 120 labeled samples per contingency, about two orders of magnitude fewer than conventional assumptions. For unseen contingencies, electrical distance coordinate encoding plus 10 labeled samples reaches the performance of the best achievable transfer-learning oracle reported in the study.