Ultra-short-term photovoltaic power prediction based on modal reconstruction and BiLSTM-CNN-Attention model

Accurate ultra-short-term photovoltaic (PV) power prediction is crucial for ensuring the power grid’s stable operation and economic dispatch. This study proposes a PV power prediction model based on modal reconstruction and bidirectional long and short-term memory network stacked convolutional neural network with embedded attentional mechanism (BiLSTM-CNN-Attention). Given that meteorological conditions impact PV power generation, the maximum information coefficient (MIC) is initially utilized to screen the vital meteorological features as input units of the prediction network. Second, the raw power is decomposed into subsequences of multiple frequencies employing the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) strategy. The sample entropy (SE) is calculated for each subsequence, and the modal reconstruction is carried out based on the similarity of the subsequences to obtain the high-frequency component, the periodic component, and the low-frequency component. Finally, the BiLSTM-CNN-Attention network is constructed for each reconstructed component, extracting deep high-level features and predicting power based on the spatio-temporal correlation. The experimental results show that under different weather conditions, the combined prediction models proposed in this study can achieve high accuracy, and the R² can reach more than 96.5%.