RTI-Net: Physics-informed deep learning for photovoltaic power forecasting

This paper introduces a novel deep learning approach for photovoltaic (PV) power generation forecasting that leverages sky images by incorporating the physical principles of solar radiation transmission. Unlike conventional methods that treat images as black-box inputs, the proposed model calculates a Radiation Transmission Index (RTI) between multi-scale pooled images and original images to directly model cloud attenuation effects on solar radiation. This physically-grounded approach quantifies cloud layer spatial distribution into meaningful features, creating a bridge between deep learning and actual PV physical mechanisms. The network architecture extracts detailed features from sky images using convolutional neural networks, computes the RTI through multi-scale pooling operations, and dynamically fuses different scale transmission features using trainable attention mechanisms. These features are then combined with historical power data for prediction through temporal modeling. Experimental validation demonstrates significant performance improvements compared to traditional methods, with a 16.4% reduction in Mean Absolute Error (MAE) and a 17.8% reduction in Root Mean Square Error (RMSE). The model shows particularly strong improvements in accurately predicting PV output during rapidly changing cloud cover conditions, which traditionally pose the greatest forecasting challenge.