Ultra-short-term Prediction of Small-sample Photovoltaic Power Based on WGAN-GP and BiLSTM-NGO

In response to the issue of low photovoltaic power prediction accuracy caused by the lack of original data and poor measurement data quality in photovoltaic power stations, this paper proposes a photovoltaic power ultra-short-term prediction model based on the Wasserstein generative adversarial network with gradient penalty (WGAN-GP) and bidirectional long short-term memory (BiLSTM). Firstly, three environmental variables highly correlated with photovoltaic power generation, namely wind speed, temperature, and radiation, are selected as input features to increase the feature dimension and improve prediction accuracy. Secondly, the WGAN-GP is utilized to learn the distribution pattern of the original photovoltaic data, and it generates high-quality new samples similar to the original data. The northern goshawk optimization (NGO) algorithm is employed to optimize the network parameters of the BiLSTM network. The BiLSTM-NGO composite model is constructed to predict photovoltaic power. Finally, simulation modeling is conducted using data from a specific photovoltaic power station. The experimental results indicate that training the prediction model with augmented samples effectively enhances the prediction accuracy. Moreover, as the proportion of augmented data to the original training set increases, the model exhibits higher accuracy in photovoltaic power prediction.