Latin Hypercube Sampling (LHS)

The foundational mechanism of LHS is the division of the cumulative distribution function (CDF) of each random variable into N non-overlapping, equal-probability intervals. This ensures that the entire range of the distribution is represented, preventing the clustering of samples that occurs in simple random sampling.

• Mechanism: A sample is randomly selected from within each interval.
• Result: Forces coverage of low-probability tails, which is critical for assessing extreme grid conditions like voltage collapse.

After generating stratified samples for each variable independently, LHS pairs them to form input vectors. A naive random pairing can induce spurious statistical correlations. Advanced LHS implementations use restricted pairing algorithms, such as the Iman & Conover method, to induce a target rank correlation matrix.

• Purpose: Accurately models the dependence between correlated inputs like wind speeds at adjacent farms.
• Method: Minimizes the discrepancy between the sample's Spearman rank correlation and the specified correlation structure.

LHS is a variance reduction technique. By ensuring full marginal stratification, it produces an estimator of the mean with lower variance than simple random sampling for the same number of samples. The convergence rate is approximately O(N⁻¹), which is theoretically faster than the O(N⁻¹/²) rate of standard Monte Carlo.

• Efficiency: Achieves a target statistical accuracy with 10 to 100 times fewer power flow simulations.
• Trade-off: The variance reduction is most significant for monotonic functions and diminishes for highly non-linear, non-monotonic models.

LHS possesses the space-filling property, meaning its projection onto any single dimension is perfectly uniform. However, a standard LHS does not guarantee uniformity in higher dimensions. To address this, optimized LHS algorithms maximize the minimum distance between points in the full multivariate space.

• Maximin Criterion: Iteratively permutes sample coordinates to maximize the minimum Euclidean distance between any two sample points.
• Application: Ensures that the surrogate models built from LHS data, such as Gaussian Process Regression, have high predictive accuracy across the entire input domain.

Unlike methods that require a closed-form probability density function, LHS operates directly on the empirical cumulative distribution function. This makes it uniquely suited for power systems where historical load or solar irradiance data does not fit a standard parametric distribution.

• Process: The empirical CDF is constructed from historical data, and the vertical axis is divided into equal-probability strata.
• Advantage: Preserves the exact shape of the empirical data, including multi-modal behaviors, without assuming a Gaussian or Weibull model.

LHS is a hybrid method situated between standard pseudo-random Monte Carlo and fully deterministic Quasi-Monte Carlo (QMC) sequences. While QMC uses low-discrepancy sequences like Sobol for a deterministic, faster convergence rate, LHS retains a stochastic element within each stratum.

• Key Difference: LHS provides an unbiased estimator of variance, which is difficult to obtain with deterministic QMC.
• Practical Use: LHS is preferred when an estimate of the simulation error is required, while QMC is used for pure numerical integration.

The foundational random sampling technique that LHS improves upon. Monte Carlo draws independent samples from input probability distributions to numerically estimate output statistics like the mean and variance of bus voltages.

• Convergence Rate: Proportional to 1/√N, requiring massive sample sizes for high accuracy.
• Variance Reduction: LHS is a variance reduction technique applied within the Monte Carlo framework.
• Use Case: The baseline benchmark against which the efficiency of LHS is measured in grid reliability studies.

A deterministic alternative to pseudo-random sampling that uses low-discrepancy sequences like Sobol or Halton to fill the input space more uniformly than random numbers.

• Convergence: Achieves a rate close to 1/N, theoretically faster than standard Monte Carlo.
• Key Difference: QMC points are deterministic and lack the statistical error bounds of LHS, making error estimation more complex.
• Grid Application: Often compared directly with LHS for probabilistic load flow where uniform coverage of the hypercube is critical.

The parent category of variance reduction techniques to which LHS belongs. Stratified sampling partitions the sample space into non-overlapping strata and draws samples from each.

• LHS Specificity: LHS is a multi-dimensional extension that stratifies each marginal distribution individually, ensuring full coverage of each variable's range.
• Benefit: Guarantees that no region of any single input variable is missed, unlike simple random sampling.
• Contrast: Traditional stratified sampling requires partitioning the full joint space, which becomes computationally intractable in high dimensions—a problem LHS solves.

A variance reduction technique that concentrates computational effort on rare but critical regions of the input space, such as the tail of a load distribution that causes a voltage violation.

• Mechanism: Draws samples from a biased proposal distribution and reweights outputs by the likelihood ratio.
• Synergy with LHS: LHS can be applied to the proposal distribution in importance sampling to ensure stratified coverage of the biased region.
• Grid Use: Essential for estimating small failure probabilities like Loss of Load Probability (LOLP) where standard LHS would still require many samples to hit the tail.

A matrix factorization critical for generating correlated random samples in LHS. The covariance matrix Σ of input variables like spatially correlated wind farms is decomposed into a lower triangular matrix L such that Σ = LLᵀ.

• Process: Independent standard normal samples are multiplied by L to induce the correct correlation structure before transformation to the target marginals.
• Rank Correlation: LHS often uses Spearman rank correlation to maintain the dependence structure through the stratified sampling process.
• Implementation: The Iman-Conover method is a standard technique for imposing a target correlation matrix on an existing LHS design.

A mathematical method for transforming correlated non-normal variables into independent standard normal variables, enabling LHS to handle arbitrary marginal distributions with dependencies.

• Two-Step Process: First transforms each variable to standard normal via its marginal CDF, then applies the inverse of the Cholesky factor of the normal-space correlation matrix.
• Limitation: Assumes a Gaussian copula dependence structure, which may not capture tail dependence in extreme events.
• Alternative: Copula theory provides a more general framework for modeling joint dependence separately from marginals.