Photovoltaic power uncertainty quantification system based on comprehensive model screening and multi-stage optimization tasks

Accurately predicting photovoltaic output power is crucial for grid management and energy dispatch. However, the objectivity of benchmark model determination in combination strategy, the stability of deterministic prediction results, the rationality of parameter setting in error distribution fitting and the validity of upper and lower bounds of prediction interval have become major challenges for current research on interval prediction. To address these issues, this paper integrates a comprehensive model evaluation mechanism with fluctuation quantification theory, and proposes a multi-stage optimized photovoltaic power interval prediction system. The system first reduces computational complexity caused by redundancy using mutual information techniques. Accordingly, the model selection module calculates comprehensive proximity to adaptively determine benchmark models. Finally, three types of parameter optimization tasks are designed to improve the reliability and resolution of the prediction intervals. The system is trained and validated using historical power data from various locations in Hebei Province, China. Results show that in four prediction scenarios, the proposed model's average comprehensive rankings are 1, 1.5, 2, and 1.5, outperforming other comparative models in predictive performance. This indicates that the method not only provides an effective solution to the current research challenges, but also offers a new vehicle for grid operators in energy dispatch and management.