Attention-Based Graph Generative Adversarial Networks for Scenario Generation of Wind Power

Day-ahead scenario generation of renewable energy sources (RES) is a powerful tool of addressing uncertainty in short-term microgrid scheduling. Most of studies only focus on the overall distribution features of the historical data, while neglect adequately the day-ahead pattern properties of RES. In this paper, an attention-based graph generative adversarial network (AG-GAN) model is proposed to generate day-ahead scenarios by utilizing RES output data and meteorological features under the form of the graph structures. In the proposed model, an attention-based spatiotemporal block is developed to process the different dimensional spatiotemporal features, which makes it possible to generate the desired day-ahead scenarios with various patterns. Numerical experiments are conducted to show the outstanding performance of AG-GAN in inferring correct day-ahead patterns and capturing inherent spatiotemporal dependence compared with two state-of-the-art methods.