Source-Load Scenario Generation Based on Weakly Supervised Adversarial Learning and Its Data-Driven Application in Energy Storage Capacity Sizing

Abstract

The historical measured data of renewable energy sources and loads can be processed in various ways to generate scenarios for energy storage planning. With the development of advanced forecast technology, the valuable reference of massive forecast data accumulated by the prediction platform in sce-nario generation is ignored. To this end, we propose a new par-adigm based on scenario generation for energy storage planning considering source-load uncertainties. First, a novel generative adversarial network (GAN) is constructed under weakly super-vised learning. The network can extract the prediction errors between predicted data and measured data to generate source-load profiles. The loss function of the network is redesigned to adapt to the data generation task. To match the planning pro-cess, scenario generation and reduction algorithms are embed-ded in the GAN. Then, an energy storage sizing model consider-ing battery health constraints is established. Case studies show that the proposed method can accurately portray the local fluc-tuation characteristics of source-load power and properly re-duce conservatism compared with the model-driven method