PV Growth Forecasting

This project used multivariate time series models from the Darts library to forecast electricity generated by photovoltaic systems in Germany. The model incorporated eight historical metrics and one predicted covariate, with architecture search focusing on N-BEATS. Multiple models were evaluated and documented, and the forecasting pipeline was used in a paper on the implications of national PV growth.

The main finding was that model selection matters less than feature engineering and proper validation when working with noisy energy data. Raw forecast accuracy can be misleading without accounting for seasonal patterns, policy shocks, and the structural breaks that characterize renewable energy adoption curves.