Reconstruction and Probabilistic Forecasting of National PV Generation

About this project

For my MSc Applied Physics at TU Delft, I developed a reproducible framework to reconstruct national solar generation and produce probabilistic forecasts 15 to 345 minutes ahead. The research ran from November 2025 to September 2026. The framework combines observations from individual PV systems, irradiance-based physical models and installed-capacity data. Forecasting uses LightGBM and national-level calibration. A central limitation is the absence of complete, independent national measurements at 15-minute resolution. Conclusions about model performance therefore depend on the reconstructed reference and a limited, predominantly residential sample. The report explicitly discusses these assumptions and limitations.