Surrogate Model

A surrogate model reduces execution time from minutes or hours to milliseconds. While a full AC optimal power flow might require iterative Newton-Raphson solves, a trained surrogate evaluates a simple algebraic expression or kernel function. This speed enables Monte Carlo simulation with tens of thousands of samples, which would be computationally prohibitive using the original high-fidelity model. In grid operations, this allows operators to run contingency analyses and probabilistic forecasts in near real-time rather than relying on offline batch studies.

A well-constructed surrogate does not merely interpolate training data; it preserves the statistical moments of the original model's output distribution. Techniques like Gaussian Process Regression provide a posterior mean prediction and a variance estimate, quantifying the surrogate's own uncertainty about its approximation. This is critical for risk-averse grid planning: the surrogate must accurately reproduce tail behaviors so that metrics like Conditional Value at Risk (CVaR) remain reliable, ensuring that rare voltage violation events are not smoothed away by the approximation.

Surrogate models are typically built using a non-intrusive or black-box approach. The high-fidelity simulator is treated as an oracle: the surrogate is trained solely on input-output pairs without requiring access to the simulator's internal Jacobian matrices or source code. This is essential in utility environments where commercial power flow solvers are proprietary. Methods include:

• Polynomial Chaos Expansion: Projects the model response onto orthogonal polynomial bases.
• Stochastic Collocation: Evaluates the simulator at specific quadrature points.
• Neural Networks: Learn the mapping directly from data.

The curse of dimensionality is the primary constraint on surrogate model applicability. As the number of uncertain input parameters grows—such as hundreds of individual wind farm outputs—the number of training samples required to cover the input space grows exponentially. Standard surrogates become impractical beyond roughly 10-20 dimensions. Mitigation strategies include:

• Sensitivity analysis using Sobol indices to screen out unimportant variables.
• Dimensionality reduction via Principal Component Analysis on correlated inputs.
• Sparse grid techniques that use hierarchical basis functions to reduce the required collocation points.

Rather than pre-computing a static training set, advanced surrogates use active learning to sequentially query the high-fidelity model where it matters most. A Gaussian Process surrogate identifies regions of high prediction variance or regions near a critical limit state (e.g., voltage collapse boundary). The algorithm then evaluates the expensive simulator at these points, iteratively refining the surrogate. This exploration-exploitation trade-off minimizes the total number of expensive simulations while maximizing the surrogate's accuracy in decision-critical regions of the parameter space.

The choice of surrogate architecture depends on the smoothness of the underlying response and the required output:

• Gaussian Process (Kriging): Best for smooth, nonlinear functions; provides built-in uncertainty quantification.
• Polynomial Chaos Expansion: Highly efficient when the response is smooth in the random parameters; converges spectrally fast.
• Radial Basis Functions: A mesh-free interpolation method suitable for scattered data in moderate dimensions.
• Deep Neural Networks: Excel at highly non-smooth or discontinuous responses but require large training sets and lack native uncertainty estimates without Bayesian extensions.

A non-parametric Bayesian regression method that defines a distribution over functions. It provides both a mean prediction and a variance-based uncertainty estimate, making it ideal as a surrogate for expensive power flow models where knowing the confidence of the approximation is critical.

• Captures epistemic uncertainty naturally
• Uses kernel functions to encode smoothness assumptions
• Excels with small to medium training datasets

A spectral method that represents a stochastic system's response as a series of orthogonal polynomials in the random input variables. When used as a surrogate, PCE enables efficient calculation of output statistics like mean, variance, and Sobol sensitivity indices directly from the polynomial coefficients.

• Hermite polynomials for Gaussian inputs
• Legendre polynomials for uniform inputs
• Exponential convergence for smooth functions

A stratified sampling method that divides the cumulative distribution of each random variable into equal intervals, ensuring full coverage of the input space with fewer samples than simple random sampling. Essential for generating the training data used to fit surrogate models.

• Reduces variance compared to Monte Carlo
• Prevents sample clustering
• Works with arbitrary marginal distributions

Variance-based global sensitivity measures that decompose the total output variance of a model into fractions attributable to individual input variables and their interactions. Once a surrogate is built, Sobol indices can be computed analytically from PCE coefficients or via Monte Carlo on the cheap surrogate.

• First-order indices: main effect of each input
• Total-effect indices: includes all interactions
• Identifies which uncertainties drive output variability

A non-intrusive uncertainty quantification method that computes the coefficients of a polynomial chaos expansion by evaluating the deterministic model at specific collocation points in the random parameter space. It treats the expensive simulator as a black box, requiring no modification to legacy power flow code.

• Uses sparse grids (Smolyak algorithm) for efficiency
• Avoids the curse of dimensionality
• Directly constructs a surrogate from evaluations

The science of identifying, characterizing, and reducing uncertainties in computational models. It distinguishes between aleatoric uncertainty (inherent randomness in wind speed or load) and epistemic uncertainty (knowledge gaps from limited data). Surrogate models are a core UQ tool for making probabilistic analysis tractable.

• Forward UQ: propagate input uncertainty to outputs
• Inverse UQ: calibrate model parameters from data
• Enables risk-informed grid planning decisions