Hybrid Physics-Informed and Data-Driven Zero-Shot Forecasting for Shaded Photovoltaic Systems

This paper investigates zero-shot photovoltaic (PV) power forecasting for shaded systems, examining a hybrid approach which combines data-driven models with physics-based shading information derived from a digital surface model (DSM). Five Forecasting models are trained using data from a reference system and can then be transferred to other systems without the need for further training. Site-specific shading effects are accounted for using a parallel PV simulation based on ray tracing to compute module-level I-V characteristics and system-level shading ratios. These are then applied as a correction to the forecasts. Experiments on the two ground-mounted photovoltaic (PV) systems demonstrate consistent reduction in forecast error under all shading conditions, with the greatest reduction observed in strongly shaded systems featuring multiple structures. An analysis of periodic retraining reveals that its effectiveness depends primarily on the availability of local data and the model architecture. On the moderately shaded system, retraining can slightly outperform DSM-corrected forecasts for TSMixer, while other models remain inferior. On the strongly shaded system, however, retraining improves the baseline performance, but does not match the performance of the DSM-corrected forecasts. ... mehrThese results demonstrate that incorporating geometric shading information enables effective zero-shot adaptation to shaded PV systems. A hybrid strategy is therefore recommended. DSM-based correction provides a robust solution in the absence of local data, while retraining only becomes competitive after sufficient observations have been collected.