Stochastic AC-OPF: Methods & Applications

• Stochastic AC optimal power flow is a computational framework that integrates nonlinear AC power equations with probabilistic constraints to manage grid reliability amid renewable variability.
• Key methodologies include adaptive surrogate models like stochastic spectral embedding, polynomial chaos expansions, and data-driven approaches to reduce computational costs while maintaining accuracy.
• Scalability is achieved through decomposition and risk-aware formulations, enabling near-real-time operation and reliable performance for power systems with high renewable penetration.

Stochastic AC Optimal Power Flow (AC-OPF) is a rigorous computational framework that integrates the nonconvex nonlinear AC power flow equations with explicit models of uncertainty, typically arising from renewable generation variability and stochastic loads. By replacing deterministic constraints with probabilistic (chance) constraints or risk measures, stochastic AC-OPF quantifies and manages the risk of operational violations, balancing optimality with reliability. This formalism underpins modern approaches to the operation and planning of power systems exposed to high penetrations of uncertain renewable resources.

1. Core Problem Formulation and Probabilistic Extension

The deterministic AC-OPF problem, defined on a network with buses $i=1,\dots,N$ and generator subset $\mathbb G$, seeks setpoints $u_i=(P_{G_i},Q_{G_i},V_i)$ for each generator $i$ so as to minimize total cost subject to power flow feasibility and operational limits:

$$\begin{aligned} \min_{u} \quad & \sum_{i\in\mathbb G} C_i(P_{G_i}) \ \text{s.t.}\quad & f_i^{\rm P}(x,u)=0, \quad f_i^{\rm Q}(x,u)=0, \quad \forall i=1,\ldots,N \ & V_i^{\min} \le V_i \le V_i^{\max},\quad P_{G_i}^{\min} \le P_{G_i} \le P_{G_i}^{\max},\quad Q_{G_i}^{\min}\le Q_{G_i} \le Q_{G_i}^{\max} \ & |S_{ij}(x,u)| \le S_{ij}^{\max},\quad \forall ij\in\mathcal E \end{aligned}$$

In stochastic AC-OPF, model data (e.g., loads, renewable injections) are represented as random variables $\xi\in\mathbb R^M$ with known PDF $\rho(\xi)$. Decision and state variables become measurable functions of $\xi$. Chance constraints are typically imposed:

$$\begin{aligned} \min_{u(\cdot)}\quad & \mathbb E\bigl[C(P_G(\xi))\bigr] \ \text{s.t.}\quad & f(x(\xi),u(\xi))=0 \quad \text{a.s.\ in }\xi \ & \mathbb P\bigl[ h(x(\xi),u(\xi)) \le 0 \bigr] \ge 1-\epsilon \end{aligned}$$

where $h(\cdot)$ encodes voltage, line-flow, and generation limits, and $\epsilon$ is the risk tolerance (Wang et al., 2024).

2. Surrogate Models and Spectral Methods

Surrogate modeling is central to making stochastic AC-OPF computationally tractable. Adaptive stochastic spectral embedding (ASSE) constructs piecewise polynomial expansions of the parametric map $\xi \mapsto y(\xi)$, partitioning the uncertainty domain $\Omega$ into subregions $D_i$ and fitting local orthonormal polynomial bases $\{\psi_\alpha\}_{\alpha\in A_i}$:

$$y(\xi) \approx \hat y(\xi) = \sum_{\alpha\in A_i} c_\alpha^{(i)} \psi_\alpha(\xi),\quad \xi\in D_i$$

Coefficients are estimated by projection:

$$c_\alpha^{(i)} = \frac{ \mathbb E \bigl[ y(\xi)\, \psi_\alpha(\xi)\, \mathbf 1_{D_i}(\xi) \bigr] }{ \mathbb E \bigl[ \psi_\alpha^2(\xi)\, \mathbf 1_{D_i}(\xi) \bigr] }$$

The ASSE method further adaptively partitions domains along directions of largest Sobol’ index, efficiently targeting regions of high local complexity. This surrogate enables fast probabilistic assessment—orders of magnitude faster than Monte Carlo—while sharply controlling global error (Wang et al., 2024).

3. Chance–Constraint Enforcement and Solution Methodologies

Several approaches exist for enforcing chance constraints in stochastic AC-OPF:

• Sampling-based methods: Random scenario generation, as in the scenario with certificates (SwC) methodology, enforces feasibility over a finite set of sampled realizations. Feasibility for $N$ samples yields a guaranteed risk upper bound (with high confidence) dependent only on $N$ and the number of design variables, not on the underlying dimension of uncertainty. The SwC approach introduces per-sample “certificate” variables for non-affine recourse and leverages semidefinite relaxations for tractability, requiring as few as $N=O(\epsilon^{-1}\log(1/\beta))$ samples for violation probability $\epsilon$ and confidence $1-\beta$ (Chamanbaz et al., 2017).
• Polynomial chaos expansions (PCE/gPCE): For arbitrary finite-variance uncertainties, the general polynomial chaos expansion (gPCE) represents stochastic variables and their nonlinear interactions in a basis of orthogonal polynomials tailored to the input distribution. Closed-form expressions for all moments are available, enabling deterministic reformulation of chance constraints (e.g., via two-moment Chebyshev inequalities):

$$\underline x \leq \mathbb E[\mathsf x] \pm \lambda(\epsilon) \sqrt{ \mathrm{Var}[\mathsf x] } \leq \overline x$$

where $\lambda(\epsilon)$ is the distribution-dependent risk multiplier (Mohy-ud-din et al., 26 Sep 2025).

• Data-driven surrogates: Gaussian process (GP) regression models, trained on data from full AC power flow samples, provide a nonparametric surrogate of the PF map. Predictive uncertainty is propagated analytically (e.g., via Taylor expansions or exact moment matching), and deterministic reformulations of the chance constraints are applied. Hybrid surrogates combine DC power flow with a GP on residuals, enhancing both accuracy and robustness (Mitrovic et al., 2022, Mitrovic, 2024).

4. Scalability and Algorithmic Aspects

Modern stochastic AC-OPF solvers leverage parallelization and decomposition:

• Two-stage decomposition: The problem can be decomposed into a first-stage (e.g., transmission network) master and multiple parallelizable second-stage (subnetwork or scenario) subproblems, with communication between levels consisting only of boundary variables and derivatives. Smoothing (barrier) techniques ensure differentiability for efficient gradient-based optimization at scale, with numerical validation up to tens of millions of buses (Tu et al., 2020).
• Adaptive surrogates: Surrogate models such as ASSE or sparse GP are refined adaptively, focusing computational effort on regions of high modeling error or strong local nonlinear behavior, thus mitigating curse of dimensionality (Wang et al., 2024).

5. Model Selection: Distributional and Policy Choices

Key modeling choices include:

• Uncertainty representation: gPCE accommodates non-Gaussian and correlated uncertainties; scenario-based approaches make minimal distributional assumptions but scale poorly in high input dimensions.
• Corrective policies: Piecewise-affine policies for voltage and power recourse are widely used, balancing fidelity and tractability. The SwC approach relaxes parametric restrictions, assigning independent recourse variables per scenario (Venzke et al., 2017, Chamanbaz et al., 2017).
• Risk measures: Most formulations are based on individual or joint chance constraints; some methods use Conditional Value-at-Risk (CVaR) for risk-averse formulations (see distributionally robust optimization approaches).
• Security constraints: Security-constrained stochastic AC-OPF extends the base framework to model contingency outcomes (e.g., $N-1$ security), propagating uncertainty through each possible post-contingency network state (Mohy-ud-din et al., 26 Sep 2025).

6. Numerical Performance and Validation

Empirical evidence demonstrates the computational and statistical efficiency of state-of-the-art methods:

Method	Violation (target/achieved)	Cost Rel. Error	Typical CPU Time	Reference
ASSE surrogate	$0.05\%$ (target $0.1\%$)	$<0.1\%$	$55$ s ($10^4$ eval)	(Wang et al., 2024)
SwC (39-bus)	$0.0014$ ($\epsilon=0.02$)	not reported	$49$ min (5105 sc.)	(Chamanbaz et al., 2017)
GP hybrid CC-OPF	$2.48\%$ ($2.5\%$)	negligible	$1.23$ s	(Mitrovic, 2024)
gPCE-CC-SCOPF	$<0.1$ pu all vars	$<$1\%	2–100× deterministic	(Mohy-ud-din et al., 26 Sep 2025)

For large-scale AC-OPF instances, two-stage decomposition with smoothing achieves practical scalability; stochastic runs with up to $1,280$ scenarios ($\sim$11 million buses) are feasible within 2.3 hours using 40 cores (Tu et al., 2020).

7. Extensions, Limitations, and Practical Implications

• Scalability: Adaptive surrogate constructions and network decompositions offer substantial scalability, circumventing the exponential sample growth of naive scenario methods (Wang et al., 2024, Tu et al., 2020).
• Model fidelity: Methods based on gPCE and data-driven surrogates retain the full nonlinearity of AC physics, crucial for response accuracy under high renewable penetration (Mohy-ud-din et al., 26 Sep 2025, Mitrovic, 2024).
• Distributional robustness: Distributionally robust optimization (DRO) techniques using Wasserstein ambiguity sets offer explicit out-of-sample risk guarantees, a critical property under data scarcity or nonstationarity (Guo et al., 2017).
• Practical adoption: Surrogates (e.g., ASSE, hybrid GPs) enable near-real-time risk assessment, suitable for integration into operational decision-making frameworks.

A plausible implication is that the ongoing development of stochastic AC-OPF formulations—especially those leveraging adaptive surrogate models and data-driven approaches—will be essential as power systems progress towards deeper renewable integration, demanding tractable yet rigorously risk-aware control solutions.