A novel prediction method for low wind output processes under very few samples based on improved W-DCGAN

Abstract

The threat of long-term low wind output processes (LWOP) on the supply ability of the power system is escalating with the increasing integration of wind power. Accurate prediction of LWOP is crucial for maintaining the stable operation of the power system. However, the occurrence probability of LWOP is low and the available samples are lacking, limiting the high-accuracy predictive modeling of LWOP. Therefore, a novel prediction method for LWOP under very few samples based on improved Wasserstein deep convolutional generative adversarial networks (W-DCGAN) is proposed here. Firstly, a multi-dimensional identification method is proposed to accurately identify historical LWOP. Then, an LWOP sample generation model based on improved W-DCGAN is established. The model integrates a long short-term memory layer into the deconvolutional layer of the generator to enhance the temporal characteristics of generated samples. Finally, three prediction algorithms are used to construct LWOP prediction models based on both generated and actual samples, respectively. The wind power operation data from a province in China is taken as an example to verify the effectiveness of the proposed method. The results show that the prediction accuracy of LWOP can be improved by 14.36%–55.85%.