⚡ electrical engineering

Power Reserve Procurement Considering Dependent Random Variables with PCE

This paper proposes a tractable chance-constrained optimization framework for probabilistic power reserve procurement that utilizes Gaussian copulas and generalized polynomial chaos to effectively model and solve problems involving dependent random variables.

Original authors: Nicola Ramseyer, Matthieu Jacobs, Mario Paolone

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nicola Ramseyer, Matthieu Jacobs, Mario Paolone

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the manager of a massive, high-stakes power grid. Your job is to keep the lights on, even when things go wrong. To do this, you buy "insurance" in the form of Ancillary Services—extra power reserves that kick in when the grid gets shaky.

In the old days, you bought this insurance from big, reliable power plants. But now, you're buying from a mix of Distributed Energy Resources (DERs): thousands of solar panels on roofs, electric vehicles charging in driveways, and heat pumps. These are great, but they are unpredictable. A solar panel might not produce power if a cloud passes over, and an EV might not be plugged in when you need it.

The Problem: The "Weather" of Uncertainty

Traditionally, when you buy insurance from these scattered sources, you assume they act independently. You think, "Well, if the sun hides in Zone A, maybe it's still shining in Zone B. They won't all fail at the same time."

But that's a dangerous gamble. In reality, these resources are dependent. If a massive storm hits a whole region, all the solar panels and wind turbines in that area might fail simultaneously. If you assume they are independent, you might buy too little insurance, thinking you're safe, only to find out you're exposed when a real storm hits.

The Solution: A New Way to Predict the Storm

This paper introduces a clever mathematical toolkit to handle these connected uncertainties. Think of it as upgrading from a simple weather forecast to a super-accurate, 3D simulation of how the weather affects the whole region at once.

Here is how they do it, broken down into three simple steps:

1. The "Gaussian Copula": The Master Connector

Imagine you have a bag of marbles. Some are red (solar), some are blue (wind). You know how likely a red marble is to be picked (its individual probability) and how likely a blue one is. But you don't know how they relate. Do they always get picked together?

The authors use a mathematical tool called a Gaussian Copula. Think of this as a special glue that sticks the individual probabilities together based on how they actually behave in the real world. It allows the model to say, "Ah, when the temperature drops, the solar output drops AND the EV charging behavior changes in a specific, linked way." It captures the hidden dance between different variables.

2. "Polynomial Chaos": The Crystal Ball

Now that we have our "glued" probabilities, we need to predict the outcome. The paper uses Generalized Polynomial Chaos (gPC).

Imagine you are trying to guess the final score of a chaotic soccer match where the wind, the referee, and the players' moods are all random. Instead of running the game a million times (which takes forever), you build a mathematical crystal ball. This crystal ball is a fancy equation (a polynomial) that learns the shape of the chaos. Once you build it, you can instantly predict the likely outcome without running thousands of simulations.

The paper's big breakthrough is figuring out how to build this crystal ball even when the variables are glued together (dependent). Previous methods either broke when things got complicated or took too long to compute.

3. The "Smart Shortcut": Dimensionality Reduction

Calculating these connections for hundreds of variables is like trying to solve a puzzle with a million pieces. It's too slow.

The authors found a smart shortcut. They realized that while the whole system is complex, the interactions that matter most are often local. It's like realizing that in a crowded room, you only really need to worry about the people standing next to you, not everyone in the building. By focusing only on the relevant connections and ignoring the "noise," they made the math fast enough to be useful in real-time.

The Real-World Test: The Cross-Border Power Trade

To prove this works, the authors simulated a scenario where two neighboring countries (let's call them Zone X and Zone Y) need to buy power reserves together.

The Old Way (Independent): They assumed the solar panels in Zone X and Zone Y were unrelated. They bought a certain amount of power, thinking they were safe.
The New Way (Dependent): They used their new "glue" (Copula) to realize that a storm in Zone X would likely hit Zone Y too.

The Result:
When they ran the simulation, the "Old Way" was dangerously optimistic. It thought they had enough power 99% of the time, but in reality, they were failing more often because they didn't account for the simultaneous failures. The "New Way" correctly identified the risk, adjusted the purchase, and ensured the grid stayed safe without wasting money on unnecessary extra insurance.

Why This Matters

This isn't just about math; it's about keeping the lights on.

For Grid Operators: It means they can buy cheaper, more efficient insurance because they aren't over-paying for "safety" they don't need, nor under-paying for risks they didn't see.
For the Planet: It makes it safer to rely on green energy (solar and wind) because we finally have a way to mathematically handle their unpredictability when they act together.

In short, this paper gives grid operators a super-powerful telescope to see the hidden connections between random events, allowing them to make smarter, safer, and cheaper decisions in an increasingly chaotic energy world.

1. Problem Statement

Distributed Energy Resources (DERs) such as heat pumps, electric vehicles, and photovoltaic (PV) systems are increasingly participating in ancillary services (AS) markets. However, their power availability is stochastic and uncertain.

The Challenge: Existing probabilistic reserve procurement methods often assume that stochastic inputs (bids) are independent. In reality, these resources exhibit dependencies (e.g., PV output in neighboring regions depends on correlated weather patterns).
The Risk: Ignoring these dependencies can lead to an underestimation of the probability of violating critical thresholds (e.g., failing to meet reserve requirements), potentially compromising grid stability.
The Gap: While Generalized Polynomial Chaos (gPC) is effective for uncertainty propagation, standard approaches struggle with dependent variables. Existing methods for handling dependencies (e.g., Gram-Schmidt orthogonalization or Rosenblatt transformations) often suffer from numerical ill-conditioning, high computational costs, or require complex inverse transforms for every sample.

2. Methodology

The authors propose a tractable framework combining Generalized Polynomial Chaos (gPC) with Gaussian Copulas to model arbitrary dependencies between random variables.

A. Modeling Dependencies via Copulas

Sklar's Theorem: The joint probability density function (PDF) is separated into marginal distributions and a dependence structure (Copula).
Gaussian Copula: The authors select the Gaussian Copula because it allows for efficient scaling to high dimensions via a correlation matrix ( $\Sigma$ ). It links marginal Cumulative Distribution Functions (CDFs) to the joint CDF.
Transformation: The joint PDF is formulated as the product of the Copula density and the marginal PDFs.

B. Tractable Computation Scheme

To avoid the numerical issues of Gram-Schmidt procedures and the sampling errors of Monte Carlo methods, the paper introduces a Quadrature Integration approach with Dimensionality Reduction:

Variable Transformation: The integration domain is mapped from the physical space $\xi$ to a unit cube $[0,1]^d$ using marginal CDFs, and then to a Gaussian latent space $z$ using the inverse normal CDF.
Gauss-Hermite Quadrature: The integral is approximated using tensor-grid Gauss-Hermite quadrature on the latent space.
Cholesky Factorization: To handle the correlation matrix $\Sigma$ , a Cholesky decomposition ( $\Sigma = LL^T$ ) is used to "color" the uncorrelated quadrature nodes, effectively transforming independent nodes into correlated ones.
Dimensionality Reduction Strategy:
- Instead of integrating over the full $d$ -dimensional space (which scales exponentially), the method exploits the fact that inner products of polynomials can be expanded into sums of monomial expectations.
- Expectations of monomials only depend on the specific dimensions involved in that monomial.
- This allows pre-computing lower-dimensional joint PDFs (marginals of the Copula) and limits the integral dimensionality to $2\nu$ (where $\nu$ is the polynomial degree), decoupling computational cost from the total number of random variables.

C. Chance-Constraint Reformulation

The method reformulates probabilistic (chance) constraints into convex constraints:

The weighted sum of gPC expansions results in a new gPC expansion.
The mean ( $\mu$ ) and variance ( $\sigma^2$ ) of the output are derived directly from the expansion coefficients.
Using a quantile approximation (e.g., $\mu + \lambda \sigma$ ), the chance constraints are converted into deterministic convex constraints suitable for standard optimization solvers.

3. Key Contributions

Arbitrary Dependency Modeling: A method to model arbitrary dependencies between random variables using Gaussian Copulas within the gPC framework.
Tractable Computation: A novel integration scheme using quadrature and dimensionality reduction that avoids the numerical instability of Gram-Schmidt orthogonalization and the high cost of Monte Carlo sampling.
Convex Reformulation: A method to transform probabilistic ancillary service procurement problems with dependent variables into convex optimization problems.
Parallelization: The pre-computation of expectation values allows for parallel processing, significantly reducing computation time.

4. Results

The method was validated on a cross-border ancillary service procurement problem involving two zones (Austria, Czechia, Germany context) with 8 random variables (representing 4 locations with temperature and irradiation).

Case 1: Normal Marginals & Linear Functions:
- The gPC expansion achieved a maximum numerical error of $\approx 10^{-7}\%$ .
- The optimization tightly satisfied the 99% chance constraints only when dependencies were modeled. Assuming independence led to constraint violations.
Case 2: Arbitrary Marginals (Beta) & Non-linear Functions:
- Used Beta distributions (common for solar irradiance) and polynomial bid functions.
- A 21-term polynomial basis was computed in ~2 minutes (non-parallelized) using 15 quadrature points per dimension.
- The expansion error remained negligible ( $\approx 10^{-6}\%$ ).
- The optimization satisfied constraints effectively, though the non-normality of distributions meant the Gaussian quantile factors ( $\lambda$ ) were slightly conservative.

5. Significance and Limitations

Significance:

Grid Stability: Provides Transmission System Operators (TSOs) with a rigorous tool to procure reserves while accounting for real-world correlations, reducing the risk of under-delivery.
Efficiency: Offers a computationally efficient alternative to distributionally robust optimization or Monte Carlo-based chance-constrained programming, enabling real-time or near-real-time decision-making.
Scalability: The dimensionality reduction strategy makes the approach viable for problems with many stochastic inputs.

Limitations & Future Work:

Quantile Factors: The exact reformulation of chance constraints relies on knowing the quantile factor $\lambda$ . While exact for normal distributions, it is unknown for arbitrary distributions, requiring conservative bounds.
Dimensionality Reduction: The efficiency gain depends on the dimensionality of the specific random functions; if functions depend on all variables, the reduction is less effective.
Future Directions: Extending the framework to time-coupled problems (accounting for temporal correlations) and investigating joint chance constraints.