Feature-Driven Strategies for Trading Wind Power and Hydrogen

TL;DR

This work addresses trading for a hybrid wind–electrolyzer plant by introducing a feature-driven prescriptive framework that maps contextual forecasts to day-ahead electricity trades and hydrogen production via linear policies: $p^{\rm DA}_t=\mathbf{q}^{\rm{DA}}\mathbf{X}_t^\top$ and $p^{\rm H}_t=\mathbf{q}^{\rm{H}}\mathbf{X}_t^\top$, with price-domain partitioning into $M$ domains to form stepwise bids. It optimizes policy parameters through a MILP trained on historical data, while enforcing constraints on wind balance, electrolyzer capacity, and a daily hydrogen quota. The real-time component introduces a rule-based adjustment to respond to realized wind and balancing prices, ensuring quota feasibility. Numerical results show that the HA+PD configuration with a sliding-window retraining and a rule-based real-time adjustment achieves profits very close to a perfect-information benchmark (oracle) and outperforms a deterministic baseline. The approach offers a practical, data-driven alternative to stochastic optimization for hybrid plants and suggests directions for online adaptation and expanded asset integration.

Abstract

This paper develops a feature-driven model for hybrid power plants, enabling them to exploit available contextual information such as historical forecasts of wind power, and make optimal wind power and hydrogen trading decisions in the day-ahead stage. For that, we develop different variations of feature-driven linear policies, including a variation where policies depend on price domains, resulting in a price-quantity bidding curve. In addition, we propose a real-time adjustment strategy for hydrogen production. Our numerical results show that the final profit obtained from our proposed feature-driven trading mechanism in the day-ahead stage together with the real-time adjustment strategy is very close to that in an ideal benchmark with perfect information.

Figures (5)

• Figure 1: The discretization of $p^{\rm DA}_t=$$\textbf{q}^{{\rm{DA}}} \textbf{X}^\top_t$ to create piece-wise price-quantity bids $\{\lambda_b,p_b\}_{b=1, ..., B}$.
• Figure 2: Illustration of sliding window approach to determine training data in each retraining interval.
• Figure 3: Distribution of day-ahead market price (right) and wind power generation (left) forecast errors.
• Figure 4: Ex-post profit of the hybrid power plant with (a) four policy architectures (GA, HA, GA+PD, HA+PD); (b) the HA+PD model for various training dataset lengths and available feature vectors (AF, RF, FM), compared to the Deterministic and Hindsight models; and (c) the HA+PD model before and after the real-time adjustment, compared to the deterministic (Det.) and Hindsight model, and the Optimal Adjustment strategy.
• Figure 5: Example of resulting price-quantity bidding curves in the day-ahead market, when the hybrid power plant buys (right) and sells (left) electricity. Recall $\rho^{\rm H}\lambda^{\rm H}$ is the hydrogen price, and $\lambda^{\rm 90\%}$ is the $90$% percentile of realized day-ahead prices in the training data.