A modified modeling approach of virtual power plant via improved federated learning

Abstract

The Virtual Power Plant (VPP) is a promising solution to power systems' challenges. However, its energy management system (EMS) faces challenges due to centralized Big-Data analysis, i.e., complexity and high computational cost. Decentralized strategies, e.g., federated learning (FL), have been proposed to mitigate these challenges. Although conventional FL can be implemented in VPPs, its effectiveness is limited by its typical structure, which relies on single-pattern modeling. Indeed, this particular structure leads to delays and inaccuracies in optimizing the heterogeneous generator dispatch patterns, particularly renewables. Hence, addressing this critical hindrance is essential for FL's efficient and effective deployment in VPPs. To overcome this drawback, this paper proposes novel VPP modeling utilizing multi-task FL modernized by multi-pattern modeling to decentralize the EMS processing. The proposed FL defines electricity generators' agents as clients collaborating to train the decentralized local data coordinated by a central server. The innovative idea is to create a novel multi-pattern-multi-task FL for electricity generation prediction so that the generator's agents learn to provide an optimized dispatch. Thus, the conventional multi-task FL is improved by proposing an optimized neural-network clustering technique. Simulation results demonstrate that multi-pattern-multi-task FL is 39%-41% faster than the best-reported methods, i.e., teaching–learning-based and honey-bee-mating optimization algorithms.© 2017 Elsevier Inc. All rights reserved.