2022 IEEE 5th International Electrical and Energy Conference (CIEEC 2022)

The uncertainty of renewable energy brings adverse effects to renewable energy consumption, and therefore, how to accurately describe the uncertainty of renewable energy becomes more and more important. Though great progress has been made in this field, these existing methods cannot consider the variation characteristic of wind power well. To tackle this problem, this paper decomposes the historical data of wind farms into state components and variations, where state components of wind power output are used to train WGAN-GP. Through the game training of WGAN-GP, the generative model can establish the mapping between noise distribution and wind power state component set. Then, variations are sampled from the corresponding t location-scale distribution and later added to the state component to generate scenarios of wind power. The simulation results show that the generated data by the proposed model closest imitates the probability distribution of historical data.

Scenario generation; Wind power and its variations; Generative Adversarial Networks with Wasserstein distance;