Polynomial Chaos Expansion (PCE)

Unlike Monte Carlo Simulation which converges at a rate of 1/√N, PCE achieves exponential convergence for smooth functions. This means a PCE with only tens or hundreds of terms can match the accuracy of millions of random samples.

• Error decay: O(e^(-αN)) for analytic functions vs. O(1/√N) for Monte Carlo
• A 10-dimensional problem often requires only ~100-1000 model evaluations
• Exploits the inherent smoothness of power flow equations under normal operating conditions

PCE selects polynomial families that are orthogonal with respect to the probability density function of each input random variable, maximizing approximation accuracy.

• Hermite polynomials for Gaussian distributions (load forecast errors)
• Legendre polynomials for uniform distributions (solar irradiance bounds)
• Laguerre polynomials for Gamma distributions (wind speed squared)
• Jacobi polynomials for Beta distributions (cloud cover fraction)

The Askey scheme provides the optimal polynomial-to-distribution mapping, ensuring the expansion captures the statistical moments with minimal terms.

PCE can be applied as a black-box wrapper around existing deterministic power flow solvers without modifying their internal code. This non-intrusive approach treats the solver as an oracle.

• Spectral projection: Compute expansion coefficients via numerical integration using quadrature rules
• Stochastic collocation: Evaluate the solver at carefully chosen collocation points and solve a least-squares regression
• Smolyak sparse grids: Reduce the curse of dimensionality by using a subset of the full tensor-product grid
• Compatible with any commercial power system software (PSS/E, PowerFactory, OpenDSS)

Once the PCE coefficients are computed, all statistical moments of the output are available analytically without further sampling. The mean, variance, skewness, and kurtosis are derived directly from the polynomial coefficients.

• Mean: Equal to the zeroth-order coefficient (μ = c₀)
• Variance: Sum of squares of all non-zero-order coefficients (σ² = Σ c²ᵢ ||ψᵢ||²)
• Sobol sensitivity indices: Computed by grouping coefficients belonging to each input variable
• PDF reconstruction: Evaluate the polynomial surrogate millions of times at negligible cost to build histograms or kernel density estimates

Full tensor-product expansions suffer from the curse of dimensionality — the number of terms grows exponentially with input dimension. Sparse truncation schemes retain only the most significant polynomial interactions.

• Total-order truncation: Keep all polynomials where the sum of degrees ≤ p (reduces terms from (p+1)^d to (d+p)!/(d!p!))
• Hyperbolic truncation: Use q-quasinorm (q < 1) to penalize high-order interactions, favoring main effects
• Low-rank Smolyak: Exploit anisotropic importance where some input variables dominate the output variance
• For a 20-dimensional problem with p=3, total-order truncation reduces terms from ~3.5 billion to 1,771

Modern implementations use adaptive basis selection to iteratively enrich the polynomial basis only in directions that reduce the approximation error, avoiding wasted terms on irrelevant dimensions.

• Least Angle Regression (LARS): Greedily selects basis functions most correlated with the residual
• Basis-adaptive methods: Start with a low-order expansion and add higher-degree terms only where needed
• Leave-one-out cross-validation error: Used as a stopping criterion without requiring a separate test set
• Particularly effective for high-dimensional power grids where only a few uncertain injections dominate the voltage variance

The foundational numerical integration method that estimates output statistics by repeatedly evaluating a deterministic model with random input samples drawn from prescribed probability distributions.

• Convergence Rate: Proportional to 1/√N, requiring thousands of simulations for high accuracy.
• Role vs. PCE: Often used as a benchmark to validate PCE results, though PCE is significantly faster for smooth models.
• Application: Calculating the probability of voltage violations in a distribution feeder with stochastic solar generation.

A non-intrusive method for computing PCE coefficients by running the deterministic model only at specific, optimally chosen points in the random parameter space.

• Key Advantage: Treats the simulation as a black box, requiring no modification to legacy power flow code.
• Node Selection: Uses Gaussian quadrature points or sparse grids (Smolyak algorithm) to minimize the number of evaluations.
• Comparison: Intrusive methods require rewriting governing equations; collocation is purely sampling-based.

Variance-based global sensitivity measures that decompose total output uncertainty into contributions from individual inputs and their interactions.

• First-Order Index (S_i): Fraction of variance due to input X_i alone.
• Total-Effect Index (S_Ti): Includes all interaction effects involving X_i.
• PCE Advantage: Sobol indices are computed analytically from PCE coefficients with negligible extra cost, avoiding nested Monte Carlo loops.

A non-parametric Bayesian method that defines a distribution over functions, providing both a mean prediction and a variance-based uncertainty estimate at any point.

• Surrogate Role: Replaces expensive power flow solvers when the input-output relationship is highly nonlinear and PCE may require too many terms.
• Kernel Choice: The covariance kernel (e.g., Matérn, squared exponential) encodes assumptions about smoothness.
• Synergy: Can be combined with PCE in multi-fidelity frameworks where GPR models the low-fidelity correction.

A statistical framework for modeling the joint dependence structure between random variables—such as wind speeds at geographically dispersed farms—separately from their individual marginal distributions.

• Sklar's Theorem: Any multivariate joint distribution can be expressed in terms of its marginals and a copula.
• Common Types: Gaussian copula, t-copula (tail dependence), and Archimedean copulas (Clayton, Gumbel).
• PCE Integration: Copulas generate correlated input samples that feed into PCE, enabling accurate modeling of spatial correlation in renewable generation.

An optimization formulation where constraints containing random variables must be satisfied with a specified probability threshold (e.g., 95%).

• Formulation: P(Line flow ≤ Thermal limit) ≥ 1 - ε, where ε is the acceptable violation risk.
• PCE Role: PCE provides an analytical surrogate for the constraint's probability distribution, converting the chance constraint into a deterministic equivalent.
• Application: Optimal power flow ensuring voltage limits are met despite uncertain renewable injections.