TLDR: A new research paper introduces an AFE–DFL framework for optimizing Battery Energy Storage System (BESS) operations, particularly in data-scarce, real-world scenarios. It addresses the limitations of traditional Predict-Then-Optimise (PTO) methods by integrating prediction and optimization, using Automated Feature Engineering (AFE) to extract richer data representations. Validated on a real-world UK property dataset, the framework demonstrates that DFL, especially when enhanced by AFE, significantly reduces operating costs compared to PTO, proving DFL’s practical viability and offering substantial economic benefits for energy management.
Managing energy effectively, especially with Battery Energy Storage Systems (BESS), is a complex task because future electricity prices and demand are always uncertain. Traditional methods, often called Predict-Then-Optimise (PTO), first predict these unknowns and then use those predictions to make decisions about when to charge or discharge batteries. The problem with PTO is that any errors in the prediction stage can lead to poor decisions later on, because the models are designed to minimize forecasting errors, not to optimize the final outcome, like reducing electricity costs.
A newer approach, Decision-Focused Learning (DFL), aims to overcome this limitation by directly integrating the prediction and optimization processes. Instead of just trying to make accurate forecasts, DFL methods train models to make predictions that lead to the best possible decisions, considering the actual costs or benefits of those decisions. This is done by using a ‘task-aware’ loss function, such as regret, which measures the difference between the optimal outcome and the outcome achieved with the model’s predictions.
While DFL shows great promise, it’s a relatively new field and has mostly been tested on simplified, synthetic datasets or small-scale problems. Applying DFL to real-world scenarios, like managing a household BESS, introduces additional challenges such as greater variability in data and, crucially, data scarcity. Real-world data collection can be limited, making it difficult for complex models to learn effectively.
To address these real-world challenges, a new framework has been proposed that enhances DFL with Automated Feature Engineering (AFE). AFE is a technique that automatically extracts richer, more informative features from existing data, such as timestamps and historical energy consumption. This process reduces the need for human experts to manually create features, which is often time-consuming and requires specialized knowledge. By leveraging AFE, the framework can make the most out of limited data, providing DFL models with better representations to improve their decision-making capabilities.
The proposed AFE–DFL framework is designed for small datasets and focuses on forecasting electricity prices and demand while simultaneously optimizing BESS operations to minimize costs. Its effectiveness was rigorously tested using a novel, real-world dataset collected from a UK property over 55 days. This dataset size is significantly smaller than those typically used in other studies, making it a strong test for the framework’s practical viability under data-scarce conditions.
The study compared the performance of DFL methods (specifically SPO+ and DBB) against the traditional PTO approach, both with and without the AFE enhancement. The results were compelling: DFL methods, on average, led to lower operating costs than PTO. More importantly, the addition of AFE significantly boosted the performance of DFL methods, improving them by 22.9% to 56.5% compared to the same models without AFE. The SPO+ method, when combined with AFE, consistently achieved the lowest regret (a measure of decision error), demonstrating its superior ability to make cost-effective battery scheduling decisions.
These findings provide strong empirical evidence for the practical viability of DFL in real-world settings, particularly when enhanced by AFE. The framework offers significant economic benefits for energy management systems by enabling more effective charge and discharge schedules for BESS, even with limited data and without extensive domain expertise. This is crucial for the accelerating integration of renewable energy sources and the optimization of energy storage.
Also Read:
- Simplifying Decision-Making Under Uncertainty: A Minimalist Bayesian Approach
- Optimizing Industrial Scheduling: A Novel DRL Environment for Flexible Job-Shops
However, the research also highlighted some limitations. The relatively small dataset size might affect the generalizability of the findings across different energy markets. Additionally, computational constraints limited the scope of hyperparameter optimization, suggesting that even greater performance gains might be possible with more extensive tuning. Future work aims to expand the framework to include renewable generation sources like solar panels, explore other DFL approaches, extend forecasting horizons beyond one day, and incorporate multi-objective optimization to balance economic benefits with environmental considerations like carbon emissions. For more details, you can read the full research paper here.
