Short-term PV prediction using multiperiod similar days and TimeGAN-inception
Xiaomei Wu, Ranran Wu, S. David Wu, Wei Li, Haijie Chen, Ning Tong
Abstract
Accurate photovoltaic (PV) power prediction is critical for new power systems due to the increasing grid-connected PV stations. However, short-term forecasting is hindered by meteorological uncertainty, questionable small-sample data quality, and inefficient feature extraction to balance accuracy with training efficiency. To address these, a short-term PV power prediction method using multiple similar-day periods, Time-Series Generative Adversarial Networks (TimeGANs), and Multi-Layer Parallel Inception Networks (MLPIs) is proposed. First, historical data utilization is enhanced through local similarity matching, resolving traditional approaches’ suboptimal local performance and mitigating meteorological uncertainty. Second, training sample quality is improved by leveraging TimeGANs’ capability to synthesize realistic, diverse time-series data. Third, MLPIs deeply extract features from historical power and meteorological data using a multi-scale 1D convolutional structure. Results demonstrate that the proposed model effectively expands the pool of high-quality PV samples and extracts time-series characteristics from PV data. Reductions in root mean square error (RMSE) of 11.8%, 20.6%, and 26.0% are achieved under clear, cloudy, and rainy conditions, respectively.