Hybrid photovoltaic power prediction method based on subsystem staged optimization and sample-level fine classification

Accurately forecasting short-term photovoltaic power generation is crucial for ensuring power system stability. However, traditional methods, due to the limitations of "fixed-stage sub-model training" and "coarse weather classification," struggle to cope with the high randomness of photovoltaic power generation, resulting in insufficient prediction accuracy and practicality. This paper proposes a hybrid photovoltaic power forecasting system that incorporates fuzzy time series preprocessing, subsystem selection, and model optimization modules. Its key breakthroughs lie in two key strategies: First, a "full-stage performance testing - optimal stage selection" mechanism is designed to screen five seed models from linear, machine learning, and pure attention models. The optimal output stage for each sub-model is determined through iterative verification, addressing the performance loss associated with fixed training. Second, a sample-level fine classification system is constructed, eliminating coarse classification and incorporating features such as irradiance and temperature to achieve one-to-one "model-to-sample" adaptation, mitigating the influence of individual differences. The system utilizes fuzzy preprocessing to improve data quality, and employs an enhanced whale optimization algorithm (EWOA) combined with simulated annealing (SA) to avoid local optima. Experiments on the Safi-Morocco three-dataset system show that the system significantly improves prediction accuracy and stability through the synergy of subsystem optimization and fine classification, providing a new path for high-fluctuation energy power prediction.