关键词: 光伏功率预测, 模态分解, 卷积神经网络, BiLSTM神经网络, 注意力机制

Abstract:

Due to the intermittency and instability of solar radiation， photovoltaic（PV） power generation shows high randomness and fluctuation， posing challenges to the stable operation of power grids. To improve prediction accuracy， the complete ensemble empirical mode decomposition with adaptive noise（CEEMDAN） was applied to decompose PV power data into intrinsic mode functions（IMFs） with different frequencies. These IMFs were clustered using K-means based on sample entropy， categorizing into high-， medium-， and low-frequency components. The high-frequency components were further decomposed using variational mode decomposition（VMD） for refined analysis. A hybrid deep learning prediction model was established by integrating convolutional neural network（CNN）， bidirectional long short-term memory （BiLSTM） network， and attention mechanism. The osprey optimization algorithm（OOA） was employed to optimize the model's hyperparameters. The experimental results showed that the proposed hybrid prediction model based on modal secondary decomposition and OOA-CNN-BiLSTM-Attention achieved a root mean square error of 4.11 kW， a mean absolute error of 2.88 kW， a mean absolute percentage error of 3.08%， and a coefficient of determination of 98.89%， outperforming other models. It is demonstrated that the proposed method effectively captures the multi-scale features of PV power generation with strong generalization ability and application potential.

Key words: photovoltaic power prediction, modal decomposition, convolutional neural network, BiLSTM neural network, attention mechanism