Sequence Generative Adversarial Networks for Wind Power Scenario Generation

With the rapid increase in distributed wind generation, considerable efforts have been devoted to the microgrid day-ahead scheduling. The effectiveness of those methods will highly depend on the selection of the uncertainty sets. We propose a distribution-free approach for wind power scenario generation, using sequence generative adversarial networks. To capture the temporal correlation, the model adopts the long short-term memory architecture and uses generative adversarial networks coupled with reinforcement learning, which, in contrast to the existing methods, avoids manual labeling and captures the complex dynamics of the weather. We conduct case studies based on the data from the Bonneville Power Administration and the National Renewable Energy Laboratory, and show that the generated scenarios can better characterize the variability of wind power and reduce the risk of uncertainties, compared with those produced by Gaussian distribution, vanilla long short-term memory, and multivariate kernel density estimation. Moreover, the proposed method achieves better performance when applied to the day-ahead scheduling of microgrids.