Total: 1
This paper presents a hybrid model-AI framework for real-time dynamic security assessment of frequency stability in power systems. The proposed method rapidly estimates key frequency parameters under a dynamic set of disturbances, which are continuously updated based on operating conditions and unit commitment. To achieve this, the framework builds on a modal-based formulation of the system frequency response (SFR), which leverages the system's eigenstructure to predict key frequency stability metrics. A Deep Sets-inspired neural network is employed to estimate the complex modal coefficients required by the modal-based SFR approach, formulated as a permutation-equivariant learning problem. This enables fast and accurate prediction of the frequency nadir and its timing across different operating conditions and disturbances. The framework achieves scalability by reusing precomputed modal structures and updating only the disturbance-specific coefficients. It demonstrates strong generalization capabilities without requiring an extensive set of operating scenarios during training or the widespread deployment of phasor measurement units (PMUs). The method is validated on the IEEE 39-bus and 118-bus systems, showing superior accuracy, robustness, and computational efficiency compared to purely data-driven approaches.