A Hybrid CNN-GRU-SE Forecasting Method for Short-Term Photovoltaic Power Considers AFD and Data Aggregation

To enhance the accuracy and robustness of short-term photovoltaic (PV) power forecasting, this paper proposes a novel forecasting method that integrates data aggregation, adaptive frequency decomposition (AFD), modified improved beluga whale optimization (MIBWO), and a CNN-GRU-SE hybrid model. First, the Pearson correlation coefficient and the entropy weight method are combined to screen meteorological features that are strongly correlated with PV power output. Considering the geographical distance, a spatial data aggregation strategy is proposed to exploit the spatial correlation among neighboring PV stations and suppress the output volatility of individual stations. Then, the AFD is adopted to adaptively decompose the PV power series into trend and seasonal components, and the MIBWO algorithm is utilized to optimize the cutoff frequency of AFD and key hyperparameters of the CNN-GRU-SE forecasting model simultaneously. Finally, the SHAP method is employed for model interpretability analysis to quantify the contribution of each feature to the prediction results. Simulation results verified the power forecasting accuracy and robustness of the proposed method. Compared with CNN-GRU and BWO-CNN-GRU-SE, the proposed method reduces MAE by 96.23% and 95.03%, respectively. The method maintains stable performance with sunny and cloudy conditions.