Zihan Ni ¹ Nuoheng Sha ^1* Chen Su ²

• ¹ Hunan Agricultural University, Changsha, China
• ² Hebei University, Baoding, Hebei Province, China

The accurate forecasting of photovoltaic (PV) power generation is pivotal for the ultra-short-term scheduling of PV power plants and the operational management of power generation strategies. To address the stochastic and volatile nature of PV power generation, a novel ultra-short-term forecasting framework is proposed. It innovatively combines the quadratic decomposition strategy and a combinatorial prediction model. The quadratic decomposition strategy consists of the Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN), Permutation Entropy (PE), and Symplectic Geometric Mode Decomposition (SGMD), which reduce the complexity and non-stationarity of the PV series. The combined prediction model consists of the Multivariate Linear Regression (MLR) and fusion neural network, improving the accuracy of the prediction of different frequency components. Firstly, ICEEMDAN was employed for the primary decomposition. The decomposed components were aggregated into low, medium, and high-frequency components by PE and the high-frequency components were further subjected to quadratic decomposition utilizing SGMD. Secondly, a composite prediction model was utilized to estimate components at various frequencies. Predictions for the high- and mid-frequency components were conducted using a fusion network model encompassing the Whale Optimization Algorithm (WOA), Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Self-Attention Mechanism (SA). Simultaneously, the low-frequency component was forecasted using the MLR model. Thirdly, the predicted values of the various components were reconstructed to yield the final forecasted outcomes. Corroborated by empirical instances, the Root Mean Square Error (RMSE) of the proposed approach ranged from 0.2778 to 0.4115 under varying meteorological conditions, representing an approximate 75% reduction compared to other models. The results demonstrate that the proposed approach enhances the granularity of photovoltaic power decomposition, fully capitalizing on the strengths of different forecasting models, and significantly improving both prediction accuracy and stability compared to existing models.