⚡ electrical engineering

Distributionally Robust Model Predictive Control for Virtual Power Plants

This paper proposes a distributionally robust model predictive control (DR-MPC) framework for Virtual Power Plants that utilizes time-varying Wasserstein ambiguity sets to optimize operations under electricity price uncertainty, demonstrating improved economic performance over standard methods when the ambiguity radius is appropriately calibrated.

Original authors: Nikolas Recke, Mathias Hudoba de Badyn

Published 2026-05-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nikolas Recke, Mathias Hudoba de Badyn

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 a neighborhood in Norway that acts like a single, giant, smart energy company. This is called a Virtual Power Plant (VPP). Instead of one big factory, it's a team made up of:

Suns and Winds: Solar panels and wind turbines that generate electricity when the weather is nice.
Batteries: Giant storage tanks that hold extra power for later.
Houses: Homes that need heating and electricity, but can also save energy when it's cheap.

The goal of this "team" is simple: Make money. They do this by buying electricity when it's cheap, storing it, and selling it back to the grid when prices are high. They also have to keep the houses warm and comfortable.

The Problem: The Weather and Prices are Unpredictable

The tricky part is that the future is a foggy crystal ball.

The Weather: Will the sun shine? Will the wind blow?
The Prices: Will electricity cost 10 cents or 50 cents tomorrow?

If the VPP guesses wrong, they might buy expensive power or miss a chance to sell cheap power. Traditional computer programs (called "Model Predictive Control" or MPC) usually just guess the most likely future and plan based on that. But if the guess is slightly off, the plan can fail, costing money.

The Solution: A "Safety Net" Strategy

The authors of this paper propose a smarter way to plan, called Distributionally Robust Model Predictive Control (DR-MPC).

Think of it like packing for a trip:

Standard Planning (MPC): You check the forecast, it says "sunny," so you pack only a t-shirt. If it rains, you get soaked.
Overly Cautious Planning (Robust MPC): You assume it might rain, snow, or hail, so you pack a heavy winter coat, an umbrella, and a raincoat. You stay dry, but you are carrying too much weight and can't move fast.
The New "DR-MPC" Strategy: You look at the forecast data and say, "It's likely sunny, but there's a small chance of a sudden shower." You pack a light rain jacket. You aren't assuming the worst-case disaster, but you aren't ignoring the risk either. You are preparing for a range of likely possibilities.

How It Works in the Paper

Learning from Data: The system uses a smart AI (called TiREx) to look at past weather and price data. Instead of giving just one number (e.g., "Price will be $50"), it gives a range of possibilities (e.g., "Price will likely be between $40 and $60, but could be $30 or $70").
The "Ambiguity Set": The computer creates a "safety bubble" around these predictions. It asks: "What is the worst thing that could happen inside this bubble of likely outcomes?"
The Decision: The VPP plans its actions (charging batteries, heating homes) to survive that "worst-case" scenario within the bubble, without being paralyzed by extreme, impossible scenarios.

What They Found (The Results)

The researchers tested this on real data from Norway during two seasons: Spring (sunny, mild) and Autumn (cloudy, cold).

The Sweet Spot: They found that if the "safety bubble" is small and just right, the VPP makes more money (about 0.8% more) than the standard guessing method. It's like finding an extra $80 in your pocket over a month just by being slightly more careful with your planning.
Too Much Caution: If the "safety bubble" is too big, the system gets scared. It stops taking risks and misses out on good deals. In this case, they actually lost money compared to the standard method.
The Main Driver: The biggest risk to their profits was the electricity price, not the weather. By focusing their "safety net" specifically on price uncertainty, they got the best results.

The Bottom Line

This paper shows that for a Virtual Power Plant, being smartly cautious pays off. By using data to understand the shape of uncertainty (not just a single guess), the system can make better financial decisions. However, you have to be careful not to be too scared of the unknown, or you'll miss out on the profits.

The authors note that while this works well now, future work could make the "safety bubble" even smarter by looking at how prices change over time (e.g., if it's expensive now, will it be expensive tomorrow?), rather than just looking at each hour in isolation.

Technical Summary: Distributionally Robust Model Predictive Control for Virtual Power Plants

Problem Statement
Virtual Power Plants (VPPs) aggregate dispatchable and non-dispatchable sources, residential buildings, and battery storage to participate in electricity markets and provide grid support. However, optimal VPP operation is complicated by inherent uncertainties in renewable generation, thermal demand, and electricity market prices. Traditional Model Predictive Control (MPC) offers some robustness through its receding-horizon structure but can suffer under large, unmodeled disturbances. Robust MPC (RMPC) ensures constraint satisfaction for all admissible disturbances but is often overly conservative, while Stochastic MPC (SMPC) reduces conservatism but relies on precise probabilistic knowledge that is rarely available in practice. Forecast errors in these variables can propagate through the control system, leading to suboptimal scheduling, increased costs, or constraint violations.

Methodology
The paper proposes a Distributionally Robust Model Predictive Control (DR-MPC) framework designed to handle electricity price uncertainty while optimizing economic performance and thermal comfort.

Unified VPP Model: The authors formulate a unified model aggregating residential buildings (modeled via a 3R2C thermal dynamics approach), battery storage, and renewable generation (solar and wind). The system operates under a discrete-time power balance constraint linking renewable production, building demand, battery operation, and grid exchange. The objective function minimizes energy costs, thermal discomfort, and battery degradation costs.
Data-Driven Forecasting: The framework utilizes the Time-series Representation Extractor (TiREx), a pre-trained zero-shot model based on an xLSTM architecture. TiREx generates causal probabilistic forecasts for electricity prices, ambient temperature, solar irradiance, and wind speed. Specifically, it outputs conditional quantiles over a 24-hour horizon.
Uncertainty Quantification: Instead of using scenario-based stochastic programming or assuming precise distributions, the method constructs stage-wise empirical predictive distributions from the forecasted quantiles. These distributions serve as the centers for time-varying Wasserstein ambiguity sets.
DR-MPC Formulation: The control problem is reformulated as a distributionally robust optimization problem. Crucially, the stage cost is affine with respect to the uncertain electricity price. This affine structure allows for a tractable closed-form reformulation of the worst-case expected cost. The resulting formulation includes a correction term proportional to the price sensitivity (net market exposure) and the radius of the Wasserstein ball ( $\epsilon$ ). The radius controls the trade-off between robustness against distributional shifts and expected profit.
Implementation: The framework operates in a receding-horizon manner. At each time step, the controller solves a finite-horizon optimization problem (24 hours) using the current state and the constructed ambiguity sets, applies the first control input, and advances the horizon.

Key Contributions

Formulation of a DR-Optimal Control Problem: The paper formulates a real-time VPP control problem that jointly optimizes economic performance and thermal comfort while explicitly robustifying against electricity price uncertainty.
Time-Varying Ambiguity Sets: It constructs time-varying Wasserstein ambiguity sets derived from empirical quantile forecasts of market prices, allowing the uncertainty model to adapt to forecast dispersion.
Integration with Causal Forecasts: The approach integrates these sets within a receding-horizon DR-MPC framework using causal probabilistic forecasts from the TiREx forecaster.
Empirical Validation: The method is validated on a realistic Nordic case study using real weather and market data, demonstrating consistent revenue improvements over standard forecast-based MPC.

Results
The framework was evaluated using real data from Oslo, Norway, across two seasonal scenarios: a spring period (April–May) with moderate temperatures and significant solar generation, and an autumn period (October) with higher heating demand and reduced irradiance. Three strategies were compared:

PF-MPC: Assumes perfect future knowledge (upper bound).
FC-MPC: Uses TiREx point forecasts without robustness margins (baseline).
DR-MPC: Augments FC-MPC with the distributionally robust price term.

Findings:

Small Radii Improve Performance: When the ambiguity radius ( $\epsilon$ $ϵ$ ) is chosen appropriately (small values), DR-MPC consistently outperforms the standard forecast-based MPC.
- In the spring scenario, $\epsilon = 0.25$ yielded a revenue gain of +0.82% under full uncertainty and +0.78% in the price-only case.
- In the autumn scenario, $\epsilon = 0.5$ performed best, achieving gains of +0.76% (full uncertainty) and +0.57% (price-only).
Conservatism Penalty: Larger radii (e.g., $\epsilon = 2$ ) became overly conservative, significantly reducing revenue (drops of up to -7.77% in the autumn scenario). This underscores the importance of proper radius selection to avoid restricting profitable arbitrage.
Price Uncertainty Dominance: Experiments isolating price uncertainty (with perfect weather knowledge) confirmed that price uncertainty is the primary driver of economic risk in this context, though removing weather uncertainty increased absolute revenue levels.

Significance and Claims
The paper claims that the proposed DR-MPC framework offers a practical and effective enhancement to standard MPC for uncertainty-aware VPP operation. By integrating predictive distribution information directly into real-time decision-making via time-varying Wasserstein sets, the method achieves consistent economic benefits (up to 0.8% revenue improvement) without requiring precise probabilistic models. The authors emphasize that the approach bridges the gap between the conservatism of RMPC and the data dependency of SMPC.

Limitations and Future Work
The authors note a limitation regarding the stage-wise construction of the empirical price distributions. Because the ambiguity sets are constructed independently at each prediction step from marginal quantile forecasts, the temporal dependencies in the price process are not explicitly captured. This may lead to overly conservative robustification of intertemporal arbitrage decisions (e.g., battery charge/discharge scheduling) where the joint distribution of prices matters.

Future work is proposed to address this by constructing joint Wasserstein ambiguity sets over price trajectories. Additionally, the authors suggest extending the VPP model to heterogeneous portfolios, incorporating uncertainty in renewable generation and thermal demand via sample average approximation or scenario-based methods, and developing data-driven mechanisms for dynamically selecting the Wasserstein radius based on season or time.