Stochastic Collocation

The defining advantage of stochastic collocation is its non-intrusive nature. Unlike intrusive methods like the stochastic Galerkin approach, it requires no modification to the underlying deterministic solver source code.

• Treats the power flow or simulation engine as a black box
• Executes the deterministic model at specific collocation points in the random space
• Constructs the polynomial chaos expansion using the outputs from these runs
• Enables UQ on legacy or commercial software where source code is inaccessible

The accuracy and efficiency of the method depend critically on the selection of collocation nodes. These points are chosen to satisfy specific mathematical optimality criteria in the random parameter space.

• Gauss quadrature points: Roots of orthogonal polynomials, providing exact integration for polynomials up to degree 2n-1
• Clenshaw-Curtis points: Extrema of Chebyshev polynomials, offering nested sets for adaptive refinement
• Sparse grids (Smolyak algorithm): Mitigate the curse of dimensionality by using a subset of the full tensor product grid
• For a problem with d uncertain parameters and polynomial order p, a full tensor grid requires (p+1)^d points, while sparse grids dramatically reduce this count

Once the deterministic model is evaluated at all collocation points, the polynomial chaos expansion (PCE) coefficients are computed. This surrogate model represents the stochastic output as a series of orthogonal polynomials.

• The output Y(ξ) is approximated as: Y(ξ) ≈ Σ cᵢ Ψᵢ(ξ)
• Ψᵢ are multivariate orthogonal polynomials (Hermite, Legendre, Laguerre) chosen to match the input distributions
• Coefficients cᵢ are computed via discrete projection using quadrature weights at the collocation points
• Once constructed, the PCE surrogate provides analytical expressions for mean, variance, and Sobol sensitivity indices at negligible computational cost

The choice of grid structure represents a fundamental trade-off between accuracy and computational cost, especially as the number of uncertain parameters grows.

• Tensor product grids: Full Cartesian product of 1D node sets. Exponential growth: N = (p+1)^d. Exact for high-order interactions but computationally prohibitive beyond 4-5 dimensions
• Sparse grids: Select a subset of tensor product nodes where the sum of 1D polynomial orders is below a threshold. Growth is approximately N ∝ 2^d d^{p-1}/(p-1)!
• Adaptive sparse grids: Refine only in dimensions with high local variance, further concentrating computational effort where the model exhibits strong nonlinearity
• For a 10-dimensional wind farm correlation problem, a sparse grid may require 10^3 evaluations versus 10^6 for a full tensor grid

Stochastic collocation exhibits spectral convergence for smooth functions: the error decreases exponentially with increasing polynomial order, not merely polynomially as in Monte Carlo.

• For analytic functions, the L² error satisfies: ||Y - Y_PCE|| ∝ exp(-α p) where p is the polynomial order
• Convergence rate depends on the regularity of the model response in the random parameter space
• Discontinuities or sharp gradients (e.g., near voltage collapse points) degrade convergence to algebraic rates
• In practice, exponential convergence means that 10-20 collocation points per dimension can achieve accuracy equivalent to 10^4-10^6 Monte Carlo samples for smooth power flow problems

Stochastic collocation occupies a middle ground between the robustness of Monte Carlo and the efficiency of intrusive spectral methods, making it ideal for grid planning workflows.

• vs. Monte Carlo: Orders of magnitude fewer model evaluations for moderate-dimensional problems (d < 20). Provides analytical moments and sensitivities directly from the PCE coefficients
• vs. Latin Hypercube Sampling: LHS improves coverage but still converges at the O(1/√N) Monte Carlo rate. Collocation achieves exponential convergence for smooth responses
• vs. Stochastic Galerkin: Galerkin requires reformulating the governing equations, which is intrusive and impractical for commercial power flow solvers. Collocation reuses existing deterministic code
• Limitation: For very high-dimensional problems (d > 50), the curse of dimensionality favors Monte Carlo methods unless strong anisotropy or low effective dimension is present