Gaussian Process Regression (Kriging)

GPR defines a distribution over functions rather than fitting parameters to a fixed form. This Bayesian foundation allows the model to update its beliefs as new simulation data arrives.

• Prior Specification: Encoded via a mean function (often zero) and a covariance kernel.
• Posterior Update: Conditioning the prior on observed training data yields a predictive distribution.
• Key Benefit: Avoids strong assumptions about the global shape of the power flow response surface.

The kernel function is the engine of GPR, defining the similarity between any two input points and thus the smoothness and structure of the learned function.

• Squared Exponential (RBF): Assumes infinitely differentiable, smooth functions.
• Matérn Family: Offers flexibility for modeling rougher, less smooth physical processes.
• Automatic Relevance Determination (ARD): Extends kernels to learn the sensitivity of the output to each input dimension, performing built-in feature selection.

Unlike neural networks or splines, GPR natively outputs a full predictive distribution N(μ, σ²) at any test point, not just a point estimate.

• Epistemic Uncertainty: The predictive variance σ² grows in regions far from training data, honestly signaling where the surrogate is unreliable.
• Confidence Intervals: 95% confidence bands are trivially computed as μ ± 1.96σ.
• Active Learning: This variance directly guides adaptive sampling strategies to optimally refine the surrogate where it matters most.

Kernel parameters (length-scales, signal variance) are learned from data by maximizing the log marginal likelihood, which automatically balances model fit against complexity.

• Closed-Form Gradient: The marginal likelihood and its gradients have analytical forms, enabling efficient gradient-based optimization.
• Occam's Razor: The marginal likelihood inherently penalizes overly complex models, mitigating overfitting without requiring a separate validation set.
• Practical Impact: This provides an objective, data-driven method for tuning the surrogate's behavior.

Standard GPR scales O(N³) in computation and O(N²) in memory due to the inversion of the N x N covariance matrix, making it challenging for very large datasets.

• Sparse Approximations: Methods like inducing points (e.g., FITC, SVGP) reduce complexity to O(NM²) for M inducing points.
• Kronecker Structure: Exploiting grid-structured inputs allows fast, separable covariance computations.
• Application: For surrogate modeling, training sets are often deliberately small (hundreds of expensive simulations), keeping standard GPR tractable.

GPR is deployed as a drop-in replacement for computationally heavy Newton-Raphson power flow solvers within Monte Carlo uncertainty quantification frameworks.

• Inputs: Stochastic nodal power injections (wind, solar, load).
• Outputs: Voltage magnitudes, line flows, and their associated uncertainty.
• Advantage: A trained GPR can evaluate millions of probabilistic scenarios in seconds, enabling real-time risk assessment that is impossible with the full physics model.

A computationally cheap approximation of a complex, high-fidelity simulation. In power systems, a Gaussian Process surrogate replaces the iterative Newton-Raphson power flow solver, reducing computation time from seconds to milliseconds. This enables real-time uncertainty quantification where thousands of Monte Carlo evaluations would otherwise be prohibitive. The surrogate learns the mapping from stochastic inputs—like renewable generation and load—to outputs such as bus voltages and line flows, preserving both the mean prediction and a variance-based confidence interval.

A statistical paradigm that updates the probability for a hypothesis as new evidence is acquired. Gaussian Process Regression is fundamentally a Bayesian non-parametric method that places a prior distribution over functions and updates it with observed data to form a posterior. This framework naturally provides epistemic uncertainty quantification, distinguishing between uncertainty due to lack of data (which decreases near training points) and inherent noise. In grid applications, this allows operators to trust predictions near known operating conditions while recognizing high uncertainty in unexplored regions.

The science of identifying, characterizing, and reducing uncertainties in computational models. GPR excels at UQ by providing a closed-form predictive variance alongside every mean prediction. This distinguishes between:

• Aleatoric uncertainty: Irreducible randomness from wind speed or load variability
• Epistemic uncertainty: Reducible model-form uncertainty from sparse training data

Grid planners use these uncertainty bounds to identify network regions requiring additional sensors or more exhaustive simulation sampling.

A spectral method representing a stochastic system's response as a series of orthogonal polynomials. While GPR is a data-driven surrogate, PCE is a physics-informed expansion that projects model outputs onto a polynomial basis. The two methods are often compared for probabilistic power flow:

• GPR: Flexible, handles arbitrary nonlinearities, provides native uncertainty
• PCE: Converges exponentially for smooth problems, requires careful basis selection

Hybrid approaches use GPR to model residuals between a low-order PCE and the true model.

A stratified sampling method that divides each input distribution into equal-probability intervals and samples once per interval. LHS is the standard method for generating the training dataset for a Gaussian Process surrogate. Compared to simple random sampling, LHS ensures:

• Full coverage of the parameter space with fewer samples
• Reduced clustering that would leave regions of the input space unexplored

A well-designed LHS design of experiments is critical for building an accurate GPR model that generalizes across the entire operating envelope of the grid.

Variance-based global sensitivity measures that decompose total output variance into contributions from individual inputs and their interactions. Once a GPR surrogate is trained, Sobol indices can be computed analytically from the GP's covariance structure without additional model evaluations. This reveals which uncertain inputs—such as specific wind farms or load centers—dominate voltage variability. Grid planners use this to prioritize sensor placement and identify the most impactful sources of forecast uncertainty.