Evaluating TabPFN for regression tasks in solar energy meteorology

Regression using tabular data is a cornerstone technique in solar energy meteorology for prediction, bias correction, and trend analysis. For over two decades, gradient-boosted trees have dominated this task. Recently, however, a generative transformer-based foundation model—the Tabular Prior-data Fitted Network (TabPFN)—has been proposed as a high-performance alternative. To evaluate this claim, we apply TabPFN to three well-established problems in solar energy meteorology: bias correction of satellite-derived irradiance, decomposition of global horizontal irradiance into diffuse and beam components, and solar power forecasting with numerical weather prediction. Our results demonstrate that TabPFN consistently outperforms gradient-boosted decision trees, with the most notable gains observed in nonlinear and high-dimensional tasks. More broadly, these findings suggest that TabPFN has the potential to advance predictive analysis using tabular data across atmospheric sciences.