Predicting Ultra-Short-Term Wind Power Combinations Under Extreme Weather Conditions

Abstract

The prediction of wind power generation is an important basis for the rational scheduling of new energy sources in wind power. However, the severe fluctuations in wind power output under extreme weather conditions pose a serious challenge for ultra-short-term wind power output prediction. A combination forecasting method for ultra-short-term wind power that comprehensively considers extreme weather and normal weather is proposed to address the above issues. To explore the differences in power time series characteristics under different scenarios in terms of the relationships between wind power output prediction errors and different weather types, time series generative adversarial networks are used to expand small sample datasets under extreme weather conditions, and a bidirectional long short-term memory (BiLSTM) deterministic prediction model based on a time attention mechanism is established. On this basis, a kernel density estimation probability prediction model is proposed for different extreme scenarios. A case study was conducted to verify the data of a wind farm in China. At the 90% confidence level, the proposed method improved the PICP and PINAW of various extreme weather conditions by averages of 10.04% and 20.47%, respectively. The results showed that the proposed method has good adaptability for wind power prediction under extreme weather conditions.

Keywords

• Deep learning
• extreme weather
• Kernel density estimation
• time GAN
• wind power prediction