Fourier Neural Operator (FNO)

Unlike standard CNNs or finite element solvers, FNOs learn operators that are independent of the computational mesh resolution. The model is trained on a specific grid but can be evaluated on arbitrary discretizations without retraining.

• Zero-shot super-resolution: A model trained on low-resolution sensor data can infer dynamics on a high-resolution grid.
• Transferable learning: The same trained operator applies to different sensor densities across a utility's heterogeneous PMU network.
• Mechanism: Parameterizes the integral kernel directly in Fourier space, bypassing the spatial grid dependency.

The core innovation of the FNO is replacing standard spatial convolutions with global integral operators parameterized in the Fourier domain. This efficiently captures long-range dependencies critical for inter-area oscillation modes.

• Global receptive field: A single Fourier layer connects every node in the grid, modeling wide-area dynamics without deep stacking.
• Low-pass filtering: Truncating high-frequency Fourier modes acts as a built-in regularizer, preventing overfitting to measurement noise.
• Complex multiplication: Learns weights directly on complex-valued Fourier coefficients, preserving phase information essential for rotor angle coherence.

FNOs function as learned propagators that map the current system state directly to a future time window in a single forward pass, eliminating the sequential time-stepping bottleneck of traditional numerical integrators.

• Autoregressive rollouts: The model predicts the state at t + Δt and feeds its own output back as input for long-horizon stability assessment.
• Fixed compute budget: Inference time is constant regardless of the simulation horizon, unlike RK45 methods where cost scales linearly with stiffness.
• Batch parallelism: Simultaneously evaluates thousands of contingency scenarios on a GPU, enabling real-time N-k ranking.

The FNO learns the underlying Green's function of the dynamical system purely from simulation or PMU data, without requiring explicit formulation of the differential-algebraic equations.

• Black-box modeling: Applicable to inverter-based resources where precise control parameters are proprietary and unknown to the grid operator.
• Hybrid augmentation: Can be trained on synthetic data from a digital twin and fine-tuned on sparse real-world PMU measurements.
• Uncertainty propagation: Naturally handles stochastic inputs like wind forecast ensembles by mapping the entire distribution of initial conditions to a distribution of stability margins.

FNOs provide a deep learning realization of Koopman operator theory, lifting nonlinear rotor angle dynamics into an infinite-dimensional linear space where global stability analysis becomes tractable.

• Linear spectral analysis: The learned Fourier layers approximate the Koopman eigenfunctions, enabling extraction of damping ratios and mode shapes directly from the latent representation.
• Global linearization: Unlike local linearization around an operating point, the FNO captures the true nonlinear manifold valid for large disturbances.
• Stability certificates: The eigenvalues of the learned operator in Fourier space provide a quantitative measure of the transient energy margin.

FNOs employ a hierarchical learning curriculum where the model first learns coarse-grained dynamics on downsampled data before progressively refining high-frequency details, dramatically accelerating convergence.

• Frequency curriculum: Lower Fourier modes (capturing inter-area swings) are learned first; higher modes (capturing local oscillations) are added incrementally.
• Transfer learning: A base FNO trained on a standard IEEE bus system can be fine-tuned for a specific utility's topology with minimal additional data.
• Memory efficiency: The truncated Fourier representation compresses the state space, allowing the model to process large-scale interconnections with thousands of buses on a single GPU.