From persistence to foundation models: A critical look at zero-shot solar forecasting

Abstract

The high-frequency variability of solar irradiance, driven by complex cloud–aerosol–radiation interactions, remains a formidable challenge for solar grid integration. Despite extensive research, eliminating the phase lag between forecasts and observations—often called the “persistence hurdle”—remains difficult. Accordingly, I investigate a paradigm shift in the field: the transition from traditional regression-based forecasting to zero-shot in-context learning via time series foundation models (TSFMs). Using a rigorous experimental framework across seven geographically diverse sites, I evaluate state-of-the-art TSFMs (released in 2024–2025) against established supervised baselines. Intra-hour forecasting results reveal a pronounced “efficiency gap”: most TSFMs fail to overcome the persistence-induced phase lag, and several yield negative forecast skill relative to the reference. In contrast, TabPFN—a tabular prior-data-fitted network—demonstrates superior utility, achieving a consistent 5% forecast skill for both global and beam irradiance. These findings suggest that, for stochastic irradiance transients, the structural priors of regression-based models are more effective than the general-purpose TSFMs. This work provides a new benchmark for univariate solar forecasting and raises important considerations regarding the zero-shot applicability of large language models in forecasting in a physical setting. Data and code are released for reproducibility.