Generator Frequency Trajectory Prediction After Large Disturbances Using Conditional Invertible Models

High penetration of converter-interfaced renewable generation is reducing system inertia and making post-disturbance frequency behavior more nonlinear, operating-condition dependent, and difficult to predict. This challenge is particularly critical for severe events such as short-circuit faults and generator outages, where fast assessment of short-term dynamic trajectories is relevant to stability analysis and security evaluation. Existing data-driven models often struggle to generalize across operating conditions and disturbance scenarios, while non-invertible latent representations may lose structure during multi-step rollout. To address this issue, this paper proposes an Effective Information-Guided Conditional Invertible Dynamics Model (EI-CIDM) for data-driven prediction of large-disturbance frequency dynamics in power systems. The model combines a conditional invertible neural network, a compact latent dynamics module, and an Effective Information regularizer to capture operating-condition dependence while improving latent transition structure. The method is evaluated on disturbance trajectories generated in ANDES for the IEEE 14-bus system. ... mehrComparative studies across three settings show that condition inputs improve fitting performance, while EI regularization substantially enhances generalization on held-out trajectories. In particular, the test-set NRMSE is reduced from 47.08% without condition inputs or EI regularization to 23.97% when both are used. These results indicate that structured invertible latent modeling can provide a promising data-driven surrogate for post-disturbance frequency-response analysis.