Sobol Indices

The Sobol method relies on the Analysis of Variance (ANOVA) representation, which uniquely decomposes a square-integrable model function $f(x)$ into a sum of orthogonal terms of increasing dimensionality:

• $f_0$: The mean response.
• $f_i(x_i)$: First-order effects of a single input $x_i$.
• $f_{ij}(x_i, x_j)$: Second-order interaction effects between inputs $x_i$ and $x_j$.
• Higher-order terms up to $f_{1,2,...,k}$.

This decomposition is unique and ensures that the variance contributions from each term are additive, provided the inputs are independent.

All Sobol indices are normalized by the total output variance $D$, ensuring they sum to one:

$1 = \sum_i S_i + \sum_{i<j} S_{ij} + ... + S_{1,2,...,k}$

This property allows for direct comparison of the relative importance of different inputs and their interactions. An index of $S_i = 0.6$ means that input $i$ alone accounts for 60% of the output variance, providing an intuitive, unitless measure of sensitivity.

The framework defines two principal indices for each input $x_i$:

• First-Order Index ($S_i$): Measures the main effect contribution of $x_i$ alone, without considering interactions. $S_i = \frac{V_{x_i}[E_{\mathbf{x}_{\sim i}}(Y|x_i)]}{D}$.
• Total-Effect Index ($S_{Ti}$): Measures the total contribution of $x_i$, including all interactions with any other variables. $S_{Ti} = \frac{E_{\mathbf{x}{\sim i}}[V{x_i}(Y|\mathbf{x}_{\sim i})]}{D}$.

The difference $S_{Ti} - S_i$ quantifies the impact of all interactions involving $x_i$. A large difference indicates strong interaction effects.

Unlike local sensitivity methods (e.g., partial derivatives at a nominal point), Sobol indices are global. They explore the entire input parameter space by averaging over the joint probability distribution of all inputs.

• Model-Free: The method makes no assumptions about the linearity, additivity, or monotonicity of the model. It works equally well on complex, non-linear, and non-monotonic power system simulations.
• Distribution-Aware: The sensitivity measure is a function of both the model structure and the specified uncertainty distributions of the inputs, reflecting realistic operational variability.

A key strength is the explicit quantification of interaction effects. The total variance can be partitioned into:

• Main effects: $\sum_i S_i$
• Interaction effects: $\sum_{i<j} S_{ij} + \sum_{i<j<k} S_{ijk} + ...$

If the sum of all first-order indices is significantly less than 1, the model exhibits strong interactions. For example, in a power flow model, the combined effect of high load and low wind generation on a voltage violation may be greater than the sum of their individual effects.

Sobol indices directly support two critical engineering decisions:

• Factor Prioritization: Inputs with a high total-effect index ($S_{Ti}$) are the primary drivers of output uncertainty. Reducing their uncertainty yields the greatest reduction in output variance.
• Factor Fixing: Inputs with a negligible total-effect index ($S_{Ti} \approx 0$) can be fixed at any nominal value within their range without significantly affecting the output distribution. This simplifies models by reducing dimensionality.

This is invaluable for grid planners identifying which renewable forecast errors most critically impact system risk.

In probabilistic power flow studies, Sobol indices rank the influence of uncertain inputs on output variance. A first-order index quantifies the direct contribution of a single wind farm's forecast error to the variance of a congested transmission line's flow.

• Example: A wind farm with a first-order index of 0.65 accounts for 65% of the line flow variance, making it the primary target for improved forecasting or curtailment strategies.
• Total-effect indices capture both direct effects and interactions with other farms, preventing underestimation of a variable's importance in complex, correlated renewable portfolios.

Sobol sensitivity analysis acts as a screening tool for distribution system operators. By computing total-effect indices for hundreds of nodal loads, planners can fix non-influential variables at their mean values, dramatically reducing the dimensionality of the stochastic problem.

• Mechanism: A total-effect index below 0.01 indicates the variable is non-influential and can be treated as deterministic without losing accuracy in the uncertainty quantification.
• Benefit: This factor fixing reduces the computational burden of Monte Carlo simulation on large feeder models by 80-90%, enabling real-time risk assessment.

A unique advantage of Sobol indices is the ability to isolate interaction effects between inputs, which linear sensitivity methods miss. In transient stability, the combined effect of a fault location and a low-inertia condition may cause instability even if neither factor alone is critical.

• Second-order index S_{ij} measures the variance due solely to the interaction between inputs X_i and X_j.
• Application: If the sum of first-order indices for fault clearing time and renewable penetration is 0.4, but their total combined effect is 0.7, the 0.3 difference signals a dangerous synergistic effect requiring specific mitigation.

Sobol indices guide phasor measurement unit (PMU) placement by identifying which bus injections most influence the uncertainty in unobserved voltage states. A high total-effect index for a load bus on a critical voltage magnitude indicates a high-value location for a new sensor.

• Process: Build a surrogate model of the distribution system state estimation; compute Sobol indices for each unmeasured load's impact on voltage estimate variance.
• Outcome: Sensors are placed where they maximally reduce the output variance of the state estimator, optimizing the observability-per-dollar ratio.

Before a Gaussian process regression or polynomial chaos expansion surrogate replaces a full power flow solver, its fidelity must be validated. Sobol indices provide a robust validation metric.

• Method: Compute Sobol indices for the top 5 influential variables using both the expensive original model and the cheap surrogate.
• Acceptance Criterion: If the surrogate's first-order indices match the original model's within a 5% tolerance, the surrogate is deemed structurally accurate for variance decomposition, not just point prediction.

In chance-constrained optimal power flow, a constraint like P(voltage violation) ≤ 5% must be satisfied. Sobol indices decompose the probability of violation into contributions from each random source.

• Conditional variance: Compute indices on a failure indicator function to see which uncertainties drive the risk of constraint violation.
• Actionable insight: If a specific solar farm's intermittency has a high Sobol index on the violation probability, the operator can procure a targeted reserve or reactive power contract to mitigate that specific risk driver.