Decision-focused learning for dynamic aggregation optimization of flexible resource clusters in active distribution networks
Active distribution networks require flexible resources to respond at the right location and time while respecting device and feeder constraints. Forecast accuracy alone does not measure the operating consequences of flexibility-estimation errors. This study formulates decision-focused dynamic aggregation by connecting a graph-temporal representation of resource capability to cluster formation and network-constrained dispatch. The model distinguishes dispatch instructions from contracted delivery, links device controls to nodal power balances, and trains continuous dispatch decisions through a conditional convex optimization layer. Cluster assignments and discrete device modes are handled outside that layer. In the reported numerical comparison, the normalized regret index is 0.56 against a deterministic reference of 1.00; the corresponding voltage and under-delivery penalty indices are 0.18 and 0.15. The highest photovoltaic-penetration case reports curtailment of 8.7%, compared with 16.4% for deterministic aggregation and 13.5% for forecast-then-optimize scheduling. The numerical experiments demonstrate improved dispatch quality and renewable utilization in the studied operating cases. The framework aligns flexibility estimation with the device and feeder constraints that determine its operational value.
Authors
- Lei An (ORCID: https://orcid.org/0000-0002-9778-2787)
- Jianan Liu (ORCID: https://orcid.org/0000-0001-5200-8125)
- Jing Xiang
- Yan Lu
- Xuan Liu
Institutions
- State Grid Corporation of China (China) (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-12
- DOI
- https://doi.org/10.1038/s41598-026-77850-6
- Primary Topic
- Optimal Power Flow Distribution
- Type
- article
- Field-Weighted Citation Impact
- 0.00