TL;DR
This paper introduces the Hybrid Renewable Energy Forecasting and Trading Competition (HEFTcom), a novel challenge that required participants to forecast and trade electricity from a hybrid wind and solar portfolio in real-time. The competition revealed that while gradient boosted trees remain strong for forecasting, successful trading hinges on a combination of forecast accuracy and sophisticated bidding strategies, highlighting the crucial link between prediction and market action.
What Problem Does This Paper Address?
Accurately forecasting renewable energy generation (wind and solar) and effectively trading this variable power in electricity markets are critical for a stable, low-carbon energy system. However, comparing different forecasting and trading strategies is challenging due to varying datasets, evaluation metrics, and the potential for data leakage. This paper addresses this by presenting HEFTcom, a competition designed to mimic real-world operational practice, requiring daily forecasts and market bids for a hybrid renewable energy portfolio. It aims to establish best practices, evaluate new forecasting advances, understand the value of forecast accuracy in decision-making, and assess the complexity of trading strategies.
Methods
HEFTcom featured two main tracks:
- Forecasting Track: Participants submitted probabilistic forecasts for the combined output of a 3.6 GW wind and solar portfolio. The primary evaluation metric was the Pinball Score, a standard for probabilistic forecasting, which penalizes inaccurate quantile predictions.
- Trading Track: Participants submitted day-ahead market bids to sell the forecasted energy. Revenue was calculated based on day-ahead market prices and imbalance settlement, with a model accounting for the participant's impact on the imbalance price.
The competition utilized a live, operational setup with daily submissions, mirroring real-world energy trading. Participants were encouraged to use various modeling techniques, and their performance was tracked using leaderboards.
Datasets
The competition utilized:
- Historical and operational data: Including wind and solar power production data for Great Britain.
- Weather forecasts: Provided from two models (DWD's ICON-EU and NCEP's GFS) with hourly resolution.
- Electricity market data: Day-ahead auction clearing prices and imbalance prices for Great Britain.
The paper notes that the competition data, including participant submissions, is shared for further analysis and benchmarking.
Key Insights and Ideas
- Gradient Boosted Trees Still Reign Supreme (Mostly): For day-ahead wind and solar forecasting, popular gradient boosted tree algorithms continued to show strong performance, but implementation details were crucial for success.
- Forecast Skill Isn't Everything in Trading: While forecast accuracy (measured by Pinball Score) is important, it's not the sole determinant of trading success. Sophisticated trading strategies that effectively leverage forecast information and market dynamics are equally, if not more, critical.
- Strategic Bidding Pays Off (with Risk): Teams that employed strategic bidding, actively managing their bids based on price spread forecasts and market conditions, achieved significantly higher revenues. However, this often came with increased revenue volatility.
- Hybrid Portfolios Present Unique Challenges: Forecasting the combined output of wind and solar farms, especially with operational constraints like the Hornsea 1 cable fault, introduced complexities not always present in single-technology forecasting competitions.
- Real-World Operational Challenges Matter: The competition highlighted the importance of robustness to unexpected events (like the cable fault), missing data, and technical issues, which are often simplified or ignored in purely academic settings.
Key Results
- Forecasting: The top-performing forecasting teams consistently used gradient boosting methods, often combining multiple models and employing advanced feature engineering and hyperparameter tuning. The winning approach (SVK) used stacked CatBoost models with a meta-model for quantile regression.
- Trading: The trading track was more volatile than forecasting. SVK also led the trading track, demonstrating a strong combination of forecasting and strategic bidding. Teams that successfully engaged in strategic bidding added over £500,000 to their revenue compared to a naive strategy, equivalent to a significant forecast improvement.
- Correlation: While there was a correlation between forecasting and trading performance, it wasn't perfect, indicating that trading strategy played a significant role.
- Participant Demographics: Over 170 teams registered, with 66 participating. Most participants had backgrounds in energy forecasting, with a strong representation from Europe.
Limitations and Future Work
- Limited External Expertise: The competition primarily attracted participants from the energy sector, missing an opportunity to draw in novel ideas from other fields. Future competitions might benefit from broader outreach or platforms that attract diverse expertise.
- Competition Duration and Structure: While the three-month duration provided meaningful results, some teams struggled with the daily submission burden and adapting to early challenges. Future competitions might consider multi-round formats or different submission frequencies to balance realism and participation.
- Focus on Day-Ahead: The competition focused on day-ahead markets. Future work could explore intraday and medium-term forecasting and trading horizons, which are gaining importance.
- Explainability and Human-Computer Interaction: As AI models become more prevalent, future competitions should prioritize emerging needs such as explainability, human-computer interaction, and the delivery of more prescriptive outputs.
- Data and Software: Leveraging advancements in open and synthetic data, and open software, will be essential for maximizing the impact of future competitions.
Summary updated