Small Sample Data Augmentation Method for Photovoltaic Power Generation Based on Improved Variational Auto-encoder

This study proposes a combined model of attention-enhanced Long Short-Term Memory (LSTM) networks and Variational Auto-Encoder (VAE) to address the challenge of augmenting small sample data, aiming to tackle the issues of limited training data and imbalanced supervised learning in photovoltaic power prediction for newly established solar power plants. Firstly, weather clustering using K-means is applied to the data. Subsequently, VAE extract reliable local features within short time windows, encoding information. Concurrently, LSTM networks integrate with attention mechanisms to infer long-term correlations within the time series. Finally, an isolated forest method is employed to analyze the quality of generated data, further assessing its quality level. Validation experiments conducted on data from a new photovoltaic station compared the proposed approach with an existing VAE data generation model, confirming the feasibility and effectiveness of the proposed method. This study offers a robust solution for addressing data scarcity and imbalance issues in photovoltaic power prediction.