Research on Ultra-Short-Term Photovoltaic Power Forecasting Method Based on Decision-Level Multimodal Fusion

Abstract

: To address the problems of limited prediction accuracy of photovoltaic (PV) power forecasting under complex meteorological conditions when using a single data source or a single model, an ultra-short-term PV power forecasting method based on decision-level multimodal fusion is proposed. This method constructs two parallel prediction paths. Path 1 adopts a two-stage architecture of “first predicting cloud images, then regressing to power” to establish a cloud-image-to-power prediction model. Path 2 takes multi-source meteorological data as input and directly predicts PV power using the CatBoost model. At the decision level, the outputs of the two paths are combined through weighted fusion with fixed weights. Validated on measured data, the fusion model achieves a normalized root mean squared error (NRMSE) of 0.0608, a normalized mean absolute error (NMAE) of 0.0382, and a coefficient of determination (R 2 ) of 0.9550. All accuracy metrics are significantly better than those of any single-modality prediction model.