Artificial Intelligence for Robust Black-Box Decision Support Systems: a Case of Study on Unit Commitment

The growing use of Smart Energy Systems (SES) supports data-driven strate- gies to boost efficiency, reduce costs, and improve reliability. In industry, the Unit Commitment Problem (UCP) helps optimise machine configura- tions but often struggles with long-term uncertainties when solved determin- istically. This work introduces a stochastic UCP formulation, using machine learning to predict uncertain parameters and a simulation environment to apply adaptive recourse actions. It compares two learning paradigms: Prediction- Focused Learning (PFL), which trains models independently, and Decision- Focused Learning (DFL), which incorporates decision outcomes. A Score Function Gradient Estimation (SFGE) method is tested as a DFL approach against a PFL baseline using both industrial and surrogate planners. Results show SFGE achieves more robust, stable performance, laying groundwork for real-world decision support systems in smart energy settings.