Short-term photovoltaic power prediction based on a VMD-SE-CNN-LSTM combined model

Abstract Solar power’s weather-driven intermittency complicates secure and flexible grid operation. This study’s key innovation is a principled coupling of entropy-guided sequence reconstruction with multiscale variational mode decomposition (VMD) and a hybrid CNN–LSTM, forming a VMD-SE-CNN-LSTM framework tailored to non-stationary photovoltaic(PV) signals. The pipeline (1) uses VMD to alleviate mode mixing; (2) performs sample entropy (SE)–based component reconstruction to suppress noise and reduce dimensionality; and (3) integrates CNN’s localized feature extraction with LSTM’s temporal modeling for prediction. Using operational data from a PV plant in Ningxia, the proposed model outperforms strong baselines (LSTM, CNN, CNN-LSTM, and VMD-CNN-LSTM), reducing MAE and RMSE by 14.9% and 14.6%, respectively, and achieving R 2 = 0.986. These results show that entropy-guided reconstruction, when fused with multiscale decomposition and hybrid deep learning, yields robust, real-plant accuracy gains, offering tangible value for PV regulation and grid dispatch.