Seasonal Forecasting of Photovoltaic Power in Germany Using Ensemble Climate Predictions

This thesis presents a machine learning framework for seasonal photovoltaic power forecasting in Germany, where the growing share of solar generation has increased the need for reliable forecasts for grid planning, maintenance scheduling and energy trading. The main aim is to investigate whether machine learning models can estimate photovoltaic power output several months in advance and to develop a consistent approach for evaluating such forecasts.
Physically meaningful features are derived from meteorological inputs, including plane of array irradiance, cell temperature, and time based variables, and are used to predict normalized photovoltaic power at daily, monthly and three-month mean resolution. Linear Regression, LightGBM and SARIMAX are trained separately for each forecast lead month, and every seasonal window is built from a single model initialization so that all months of a season share the same atmospheric starting state. The performance of the models is compared against a climatological baseline across lead months one to six.
Results show that LightGBM outperforms the other models during seasonal forecasting, providing the most stable predictions once the meteorological signal weakens at longer horizons. Model accuracy is quantified using root mean square error, mean absolute error and mean bias, together with the skill score derived from root mean square error, which together provide a reliable basis for judging predictive accuracy across leads, resolutions and seasons.
Overall, this work provides a structured methodological foundation for evaluating seasonal photovoltaic forecasts and improving the reliability of long term photovoltaic power forecasts, contributing to more informed renewable energy planning and grid integration.