Stacking algorithm based framework with strong generalization performance for ultra-short-term photovoltaic power forecasting

Standard forecasting methods are commonly applicable to specific types of photovoltaic (PV) systems, and the prediction accuracy drops significantly for PV systems in different modules or locations. This paper proposes an ultra-short-term PV power forecasting framework based on the Stacking ensemble algorithm (StAB), which achieves high precision prediction and good generalization performance through multi-model integration and data mining. Firstly, StAB utilizes Spearman’s rank correlation coefficient to select appropriate features to reduce computational complexity. In addition, a correlation guided fast Fourier transform (CGFFT) is proposed to choose reasonably the optimal decomposition frequency, and decompose PV power into high-frequency and low-frequency components to reduce the prediction difficulty. Importantly, StAB builds an ensemble prediction model based on the Stacking algorithm, by solving optimization problems in the base models and meta-model selection process, enabling the ensemble model to maintain desired performance for different distributed PV systems. The framework is tested on 12 PV power datasets, and the results demonstrate that the proposed framework outperforms 12 advanced artificial intelligence models and another two Stacking algorithm based models, in terms of prediction accuracy and generalization performance.