Accurate measurements of global solar radiation are essential for applications ranging from photovoltaic (PV) yield assessment to grid operation and agricultural management. However, direct measurements with pyranometers are sparse and costly, motivating alternative approaches based on widely available PV power production data. This thesis investigates how accurately global radiation can be estimated using machine learning models trained on PV power output data, supplemented with weather reanalysis data.

Two tree-based models, Random Forest and Gradient Boosting, were applied to PV data from two sites in the Netherlands combined with globally available weather reanalysis data, using on-site pyranometer measurements as the ground truth for training and validation. Both single-location models, trained and validated on the same site, and cross-location models, trained on one site and applied to another, were evaluated. For comparison, a linear regression baseline model was also tested.

The single-location models achieved strong performance, with R2 ≈ 0.97, mean absolute errors (MAE) of 10–15 [W/m²], and near-zero bias at a 15-minute resolution, substantially outperforming established reanalysis products such as ERA5. Cross-location models retained reasonable accuracy (R² ≈ 0.94, MAE 15–30 [W/m²]), though with increased bias. Feature importance analysis highlighted the dominant influence of PV power output and the clear-sky index for photovoltaics (K_PV), while reanalysis variables contributed little. The top-of-atmosphere radiation proved to be a consistently useful predictor.

These results demonstrate that PV power data can be transformed into accurate, high-frequency radiation estimates, offering a cost-effective complement to pyranometer measurements. While site-specific calibration currently limits full generalisation, the findings point toward the feasibility of scalable models based on transferable features such as KP V . Expanding datasets across more climates and system configurations, and exploring advanced model architectures, are promising next steps towards generalised PV-based radiation models that could be applied without requiring on-site pyranometer measurements for training.