Introduction
Large electricity consumers and retailers face a significant financial challenge: timing their purchases in forward markets. Instead of buying power on the day of delivery, they hedge against price volatility by purchasing "Calendar" (CAL) contracts up to three years in advance. Deciding whether to buy today or wait for a price dip is a complex sequential decision-making problem. Currently, many organizations rely on human heuristics or opaque stochastic models that lack transparency and are difficult to trust.
This paper introduces the Uniformity-based Procurement of Electricity (UPE) algorithm. By combining deep learning-based trend forecasting with a "uniformity" safety net, UPE achieves a 1.65% cost reduction over standard benchmarks while maintaining the explainability required for industrial deployment.
Methodology: The UPE Algorithm
The UPE algorithm moves away from pure "black-box" prediction by employing a two-pillar strategy that balances AI-driven insights with robust risk management.
1. Trend Forecasting
The first pillar uses a Deep Neural Network (DNN) to predict the dominant market direction (upward or downward). To train this model effectively, the authors use a "lag-free low-pass filter" on historical prices. This filter strips away short-term noise to reveal the "true" underlying trend, which serves as the ground truth for the neural network's training phase.
2. Procurement Uniformity Level ($u_t$)
The second pillar is the "uniformity" indicator, which tracks how far the current procurement progress deviates from a perfectly uniform strategy (buying exactly $1/T$ of the total needs every day).
The logic is intuitive: if the AI predicts a price drop, the algorithm allows procurement to "lag" behind the uniform schedule. Conversely, if a price hike is predicted, it "leads" the schedule. Crucially, this deviation is bounded by tunable "trigger" values ($u^+$ and $u^-$). These act as a safety net, ensuring that even if the forecast is incorrect, the strategy never drifts too far from a safe, diversified buying path.
Data and Evaluation
The study utilizes data from the Belgian forward market (ICE Endex), focusing on yearly base load "Calendar" (CAL) products. The dataset spans CAL 2012 to CAL 2019, with each product consisting of approximately 750 daily prices. To prevent data leakage and ensure realistic testing, the model for a specific year (e.g., CAL 2018) is trained using data from three years prior (CAL 2015).
Key Results and Insights
The UPE-DL (Deep Learning) approach significantly outperformed standard benchmarks, including naive balanced strategies and standard moving average trend-followers:
- Quantifiable Savings: UPE-DL achieved an average 1.65% reduction in costs compared to the mean electricity price. For a major consumer with 100 GWh annual usage, this translates to roughly €75,100 in yearly savings.
- Stability and Consistency: While traditional trading strategies (like EPMA) failed significantly in certain volatile years like 2019, the UPE-DL model remained stable and profitable across diverse market conditions.
- Explainability as a Requirement: The uniformity indicator provides a clear rationale for decisions. A human supervisor can easily interpret why the AI is waiting: "We are currently 10% ahead of schedule and a downward trend is predicted, so we are waiting for a better entry point."
Conclusion and Future Work
The UPE algorithm demonstrates that AI can be both powerful and transparent in the energy sector. By anchoring deep learning forecasts to a uniform procurement schedule, the authors mitigate the risks associated with pure trend-following strategies.
Future research may involve incorporating exogenous data—such as gas prices, carbon credits, or macroeconomic indicators—to further sharpen forecasts. Additionally, the authors suggest that Deep Reinforcement Learning (DRL) could be a natural next step to solve this sequential decision problem, potentially discovering even more sophisticated buying patterns.
Summary updated