Multi-Market Bidding of Virtual Power Plants with Machine Learning-Assisted Asynchronous Distributed Optimization
Résumé
Virtual power plants aggregate distributed energy resources, such as renewables, flexible loads, and storage units, for coordinated participation in multiple electricity markets. This requires the repeated optimization of complex bidding strategies within short time windows. Distributed algorithms address this by decomposing the bidding problem; their convergence is however impeded by synchronous schemes where every update requires outputs from all agents. Alternatively, asynchronous distributed algorithms can use auto-regressive methods to predict outputs, but they remain vulnerable to errors and degrade under prolonged delays. We propose to consider the impact of predictions on the objective costs by directly minimizing a decision regret. We design 2 decision-focused learning methods, and integrate them to a distributed bidding strategy based on the alternating direction method of multipliers. Simulations under failures and delays show that our method converges faster than existing asynchronous distributed alternatives, while approaching the optimal cost of centralized optimization.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |