一种用于短期光伏发电功率预测的三阶段混合模型

Abstract Accurate prediction of Photovoltaic (PV) power generation is critical for renewable energy dispatch. To improve prediction accuracy under fluctuating meteorological conditions, this paper proposes a three-stage hybrid model for short-term PV power prediction, integrating similar day optimization, multi-level signal processing and hybrid prediction. Firstly, historical data are segmented into time periods. The Principal Component Analysis-Newton-Raphson-Based Optimizer-K-means++ (PCA-NRBO-K-means++) clustering algorithm is used to distinguish different weather categories within each period. Similar time periods to the target day are selected within the same weather category, and a similar day dataset is reconstructed. Secondly, through the combined method of Variational Mode Decomposition-Fuzzy Entropy (VMD-FE) and Complete Ensemble Empirical Mode Decomposition with Adaptive Noise-Fuzzy Entropy (CEEMDAN-FE), the similar day PV power is decomposed into high-frequency, low-frequency, trend and residual terms. Finally, a hybrid model is constructed for different reconstructed sequences. After separate prediction and fusion, the result is optimized by introducing error compensation. Experiments based on data from a PV power plant in Ningxia, China, demonstrate that the proposed model reduces the average root mean square error (RMSE) by 72.4 % (from 1.2925 MW to 0.3572 MW) and the average absolute error (MAE) by 73.3 % (from 1.0472 MW to 0.2791 MW), compared to the baseline model. The model also performs well under complex weather conditions. Under overcast and rainy conditions, RMSE is decreased by 71.0 % and 65.7 %, while MAE is decreased by 72.7 % and 66.1 %. The results indicate that the model maintains high adaptability under diverse meteorological conditions.