关键词: 光伏发电, 功率预测, 神经网络, 自适应稀疏自注意力机制, 模态分解

Abstract: Aiming at the problem that the volatility and randomness of meteorological data in PV power data lead to poor prediction accuracy of a single model in PV power prediction, it is proposed to construct a short-term PV output prediction framework integrating empirical modal decomposition and hybrid neural network architecture. Firstly, an improved fully integrated empirical modal decomposition method is used to process the PV historical data; secondly, a hybrid PV power prediction model based on the adaptive coefficient attention mechanism is constructed, in which the LSTM model extracts the local time-series correlation features, and the ASTransformer model captures the cross-cycle dependencies. Finally, the component prediction results are superimposed to obtain the final PV power prediction results. In this study, real data from a PV site in Jiangsu Province, China, are used to validate the model performance, and the results show that the prediction accuracy and stability of the proposed method are significantly better than that of the control model under different weather conditions, and it has good prediction results.

Key words: Photovoltaic power generation, Power prediction, Neural network, Adaptive sparse self-attention mechanism, Modal decomposition