Stochastic Programming

The core mechanism of stochastic programming is the representation of uncertainty through a discrete set of scenarios, each with an associated probability weight. A scenario tree captures the sequential unfolding of uncertainty over time. For grid applications, a scenario might represent a specific combination of wind generation (e.g., 150 MW), solar output (e.g., 80 MW), and load (e.g., 1.2 GW) at a future hour. The optimization problem is then formulated to minimize the expected cost across all scenarios rather than optimizing for a single deterministic forecast. This transforms a brittle point estimate into a robust decision that hedges against adverse outcomes.

The most common architecture is the two-stage recourse model, which mirrors real-world grid operations:

• First-stage (here-and-now) decisions: Actions taken before uncertainty is resolved, such as day-ahead unit commitment, generator dispatch schedules, or reserve procurement. These decisions must be feasible regardless of which scenario materializes.
• Second-stage (wait-and-see) decisions: Corrective actions taken after the random variable is observed, such as deploying reserves, curtailing renewables, or activating demand response. These recourse actions incur scenario-dependent costs that penalize first-stage decisions that leave the system exposed.

The objective minimizes first-stage costs plus the expected value of recourse costs across all scenarios, ensuring the initial plan is hedged against expensive corrective measures.

Stochastic programming requires explicit specification of the probability density function for each uncertain parameter. In renewable energy contexts, this often involves:

• Wind power: Weibull distributions fitted to historical wind speed data, transformed through turbine power curves.
• Solar irradiance: Beta distributions capturing the bounded nature of cloud cover and clear-sky indices.
• Load forecasting errors: Truncated normal distributions derived from historical forecast residuals.

Scenario generation techniques such as Monte Carlo sampling, Latin Hypercube Sampling, or moment matching are used to discretize these continuous distributions into a manageable scenario set. Scenario reduction algorithms like fast forward selection then prune the set to preserve statistical moments while maintaining computational tractability.

Standard stochastic programming optimizes expected value, which may be insufficient for grid reliability where extreme tail events cause cascading failures. Risk-averse formulations incorporate coherent risk measures into the objective or constraints:

• Conditional Value at Risk (CVaR): Minimizes the expected cost of the worst α-percentile of scenarios, explicitly penalizing high-impact low-probability events like simultaneous generator outages during peak load.
• Robust optimization hybrids: Combine stochastic scenarios with uncertainty sets to guarantee feasibility within a defined uncertainty budget.
• Chance constraints: Enforce probabilistic guarantees (e.g., "the probability of voltage violation must be less than 0.1%") using the scenario distribution.

These extensions are critical for N-1 security-constrained dispatch where the cost of a single contingency scenario can dominate the objective.

The scenario-based formulation causes the problem size to grow multiplicatively with the number of scenarios, quickly exceeding the capacity of commercial solvers. Specialized decomposition algorithms exploit the problem's block-angular structure:

• Benders decomposition: Separates the problem into a master problem (first-stage decisions) and subproblems (one per scenario). Optimality and feasibility cuts are iteratively generated from the dual solutions of subproblems and added to the master problem, converging to the optimal solution without explicitly forming the full extensive form.
• Progressive Hedging: A scenario-based decomposition that solves each scenario independently and penalizes deviations from a consensus first-stage decision, iteratively converging to non-anticipativity.
• Stochastic Dual Dynamic Programming (SDDP): Extends Benders decomposition to multi-stage problems by approximating the cost-to-go function with piecewise linear cuts, enabling horizon lengths of weeks or months for hydro-thermal coordination.

A fundamental requirement in stochastic programming is non-anticipativity: decisions made at any stage cannot depend on future outcomes that have not yet been revealed. In a two-stage model, this means the first-stage decision vector must be identical across all scenarios. Mathematically, this is enforced by constraints that equate the first-stage variables across scenario subproblems. In multi-stage formulations, non-anticipativity is enforced at each node of the scenario tree, ensuring decisions at a given node depend only on the history of observations up to that point. This constraint is what distinguishes stochastic programming from deterministic scenario analysis and is the source of the problem's computational complexity.

The foundational architecture for grid commitment problems. First-stage (here-and-now) decisions—such as unit commitment and day-ahead market bids—must be made before uncertainty is realized. Second-stage (wait-and-see) decisions—like real-time dispatch adjustments and reserve deployment—are corrective actions taken after observing the actual wind or load outcome.

• Mechanism: Minimizes first-stage costs plus the expected value of second-stage recourse costs
• Grid Application: Determining optimal day-ahead generator schedules while accounting for the probability-weighted cost of real-time balancing actions
• Key Distinction: Recourse decisions are scenario-dependent, creating a decision tree rather than a single deterministic path

The process of constructing a computationally tractable representation of uncertainty from historical data or probabilistic forecasts. Raw continuous distributions are discretized into a finite scenario tree that preserves statistical moments while remaining solvable.

• Techniques: Monte Carlo sampling, Latin Hypercube Sampling, and moment-matching methods
• Reduction Methods: Fast-forward selection and backward reduction algorithms prune similar scenarios to maintain tractability while preserving distribution tails
• Grid Context: A wind farm's 24-hour forecast might generate 1,000 scenarios, reduced to 20 representative trajectories that capture the 5th and 95th percentile extremes

An alternative formulation where constraints are expressed as probabilistic guarantees rather than hard requirements. A constraint such as 'line flow must not exceed thermal limit' becomes 'the probability of exceeding the thermal limit must be less than 5%'.

• Mathematical Form: P(g(x,ξ) ≤ 0) ≥ 1 − ε, where ε is the acceptable violation probability
• Grid Application: Ensuring voltage limits are satisfied with 95% confidence despite stochastic solar injection, allowing controlled risk in exchange for higher renewable utilization
• Solution Approaches: Individual chance constraints can be reformulated as deterministic equivalents using the inverse cumulative distribution function when the underlying distribution is known

Extends the two-stage framework to sequential decision-making where uncertainty is revealed gradually over multiple time periods. Decisions at each stage are non-anticipative—they can only depend on information available up to that point.

• Structure: A scenario tree where nodes represent decision points and branches represent uncertainty realizations
• Grid Application: Weekly hydro-thermal coordination where reservoir release decisions are made each day as updated inflow forecasts arrive
• Computational Challenge: Scenario tree size grows exponentially with stages, requiring decomposition methods like Stochastic Dual Dynamic Programming (SDDP) for tractability

Standard stochastic programming minimizes expected cost, which treats extreme losses and minor deviations equally. Conditional Value at Risk (CVaR) augments the objective to penalize the tail of the loss distribution, protecting against catastrophic outcomes.

• Definition: CVaR at confidence level α is the expected loss given that the loss exceeds the Value at Risk (VaR) threshold
• Grid Application: A distribution system operator minimizes expected operational cost while adding a weighted CVaR term to avoid scenarios where voltage collapse occurs in the worst 5% of wind drop-off events
• Tuning: The risk-aversion weight λ balances expected performance against tail-risk protection, creating an efficient frontier for decision-makers

Large-scale stochastic programs with millions of scenarios cannot be solved monolithically. Decomposition techniques exploit the problem's block-diagonal structure to achieve computational tractability.

• Benders Decomposition: Separates the problem into a master problem (first-stage decisions) and subproblems (scenario-specific recourse), iterating via cutting planes
• Progressive Hedging: A scenario-based decomposition that penalizes deviations from a consensus solution, gradually enforcing non-anticipativity constraints
• Grid Relevance: Enables solving the stochastic unit commitment problem for ISO-scale systems with thousands of buses and hundreds of wind scenarios within operational timeframes