Research on ultra-short-term photovoltaic power forecasting using multimodal data and ensemble learning

To enhance the ultra-short-term prediction capability of photovoltaic power generation, this study proposes a forecasting method integrating ensemble learning with multimodal data. After systematically comparing the predictive performance of six independent machine learning models (RF, XGBoost, CatBoost, LightGBM, LSTM, and GRU), a fused model was developed using the stacking ensemble strategy. The ensemble model achieved the highest coefficient of determination (R 2 =0.9698) along with the lowest normalized mean square error (NMSE=0.0020) and normalized root mean square error (NRMSE=0.0451). Compared to meteorological data models and ground-based cloud image models, the proposed multimodal ensemble learning model improved R 2 by 20.8% and 17.9%, reduced NMSE by 84.8% and 83.3%, and decreased NRMSE by 60.9% and 58.8%, respectively. Contribution analysis revealed that power from the previous moment, light of ground-based cloud images, and station irradiance were critical factors influencing model predictions. • The ensemble model integrates meteorological data, ground-based cloud imagery, and historical power records, achieving significantly improved prediction accuracy compared to single-modality approaches. • By combining diverse machine learning models via a stacking strategy, the proposed framework outperforms standalone and hybrid models in ultra-short-term photovoltaic power forecasting. • GAN-based occlusion removal and MLP imputation address missing values and sky image distortions, while K-means clustering ensures outlier handling, significantly improving data quality. • Power from the prior moment, cloud image Light intensity, and station irradiance are key contributors, validated through feature importance analysis.