Short-term prediction of wind power and its ramp events based on semi-supervised generative adversarial network

Short-term predictions of wind power and its ramp events play a critical role in economic operation and risk management of smart grid. This paper proposes a hybrid forecasting model based on semi-supervised generative adversarial network (GAN) to solve the short-term wind power outputs and ramp event forecasting problems. In the proposed model, the original time series of wind energy data can be decomposed into several sub-series characterized by intrinsic mode functions (IMFs) with different frequencies, and the semi-supervised regression with label learning is employed for data augmentation to extract non-linear and dynamic behaviors from each IMF. Then, the GAN generative model is used to obtain unlabeled virtual samples for capturing data distribution characteristics of wind power outputs, while the discriminative model is redesigned with a semi-supervised regression layer to perform the point prediction of wind power. These two GAN models form a min-max game so as to improve the sample generation quality and reduce forecasting errors. Moreover, a self-tuning forecasting strategy with multi-label classifier is proposed to facilitate the forecasting of wind power ramp events. Finally, the real data of a wind farm from Belgium is collected in the case study to demonstrate the superior performance of the proposed approach compared with other forecasting algorithms.

Automated summary

This paper introduces a hybrid forecasting model based on semi-supervised generative adversarial network for short-term wind power and ramp event prediction. By decomposing wind energy data time series and employing semi-supervised regression, non-linear and dynamic behaviors are extracted to enhance forecasting accuracy, along with a self-tuning forecasting strategy for improved performance.