Transferable Solar Power Forecasting Using Hectometric NWP and Foundation models

Accurate forecasting of solar power generation is critical for grid stability and energy planning, particularly as photovoltaic (PV) systems expand across Europe. However, the inherently location-dependent nature of PV production, coupled with limited availability of site-specific data, presents a major challenge for generating reliable forecasts across spatial and temporal scales. This study presents a scalable and transferable machine learning framework that combines synthetic data and real-world observations to enhance solar PV forecasting in data-scarce regions 

We generate synthetic PV production time series across randomly selected locations in Europe using high-resolution hectometric numerical weather prediction (NWP) simulations. To increase realism and robustness, we integrate several additional data sources: ERA5 reanalysis for climatological consistency and gap filling, CAMS satellite-based radiation products for improved irradiance realism, and the high-resolution New European Wind Atlas (NEWA) for supplementary wind and solar surface fields. PV output is modelled using PVLib, using realistic metadata (e.g., panel tilt, azimuth, location) to simulate realistic production patterns.  

In addition to synthetic sites, a set of real PV locations is used to anchor the dataset and validate model behaviour. These real cases can also be perturbed or scaled to test robustness and generalization. A hybrid machine learning setup is then trained on this combined dataset, leveraging both foundation models and classical ML techniques. The training pipeline includes standardized preprocessing and feature engineering to ensure consistent input preparation across all sites and conditions. 

The trained models are evaluated on unseen PV sites and extreme weather cases to assess their generalization capacity and transferability. Our results show that synthetic data, especially when enhanced with multi-source auxiliary datasets, significantly improves forecast accuracy in previously unobserved or data-scarce areas. This approach lays the foundation for a transferable, pan-European PV forecasting system.