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:

Due to the strong volatility and randomness of photovoltaic power data caused by meteorological factors， existing single prediction models generally suffer from insufficient accuracy and poor generalization performance ability. To address this issue， a short-term photovoltaic power output prediction framework integrating improved complete ensemble empirical mode decomposition with adaptive noise（ICEEMDAN） and a hybrid neural network architecture consisting of long short- term memory（LSTM） and adaptive sparse Transformer（ASTransformer） was proposed. According to the Pearson correlation coefficients between various meteorological factors （irradiance， diffuse radiation， temperature， relative humidity， and wind speed） and photovoltaic power， the core input features with the highest correlations to photovoltaic power were selected. The ICEEMDAN method was adopted to denoise and reduce the dimensionality of the non stationary historical photovoltaic time series data， decomposing the original complex sequence into multiple intrinsic mode functions （IMFs） and residual components with more stable characteristics. A hybrid prediction model for photovoltaic power generation with an adaptive sparse attention mechanism（LSTM-ASTransformer） was constructed， in which the LSTM accurately extracted local temporal correlation features， while the ASTransformer efficiently captured cross cycle dependency relationships in the sequence. Through a dual branch adaptive fused sparse attention mechanism， irrelevant redundant features were filtered out， and an optimized feature refinement feed forward network was employed to further extract effective information. The prediction results of all subcomponents were inversely superimposed to obtain the final photovoltaic power prediction value. The model performance was verified using real data from a photovoltaic power station in Jiangsu Province， China， during 2019 —2020. The results showed that under different weather conditions， the prediction accuracy and stability of the ICEEMDAN-LSTM-ASTransformer model were significantly better than those of the comparison models. Specifically， under sunny weather conditions， compared with the Crossformer model， the mean absolute error and root mean square error were reduced by 22.76% and 7.09%， respectively. Under rainy weather conditions， compared with the TimesNet model， they were reduced by 22.15% and 31.92%， respectively， and under cloudy weather conditions， compared with the Crossformer model， they were reduced by 10.03% and 8.96%， respectively， demonstrating good prediction performance. The ICEEMDAN-LSTM-ASTransformer integrated the advantages of multiple models， which can improve prediction accuracy from multiple perspectives and optimize prediction efficiency.

Key words: photovoltaic power generation, power prediction, neural network, adaptive sparse self-attention mechanism, modal decomposition